Close
About
FAQ
Home
Collections
Login
USC Login
Register
0
Selected
Invert selection
Deselect all
Deselect all
Click here to refresh results
Click here to refresh results
USC
/
Digital Library
/
University of Southern California Dissertations and Theses
/
00001.tif
(USC Thesis Other)
00001.tif
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
INFORMATION TO USERS This manuscript has been reproduced from the microfilm master. UMI films the text directly from the original or copy submitted. Thus, same thesis and dissertation copies are in typewriter face, while others may be from any type of computer printer. The quality of this reproduction is dependent upon the quality of the copy submitted. Broken or indistinct print, colored or poor quality illustrations and photographs, print bleedthrough, substandard margins, and improper alignment can adversely affect reproduction. In the unlikely event that the author did not send UMI a complete manuscript and there are missing pages, these will be noted. Also, if unauthorized copyright material had to be removed, a note will indicate the deletion. Oversize materials (e.g., maps, drawings, charts) are reproduced by sectioning the original, beginning at the upper left-hand comer and continuing from left to right in equal sections with small overlaps. Each original is also photographed in one exposure and is included in reduced form at the back of the book. Photographs included in the original manuscript have been reproduced xerographically in this copy. Higher quality 6” x 9” black and white photographic prints are available for any photographs or illustraiioTis appearing in this copy for an additional charge. Contact UMI directly to order. UMI A Bell & Howell Information Company 300 North Zeeb Road, Ann Arbor MI 48106-1346 USA 313/761-4700 800/521-0600 A S tu dy o f R esource R eservation P rotocol Scaling and D ynam ics in Integrated Service Packet N etw orks by Danny James Mitzel A Dissertation Presented to the FACULTY OF TH E GRADUATE SCHOOL UNIVERSITY OF SOUTHERN CALIFORNIA In Partial Fulfillment of the Requirements for the Degree D O CTO R OF PHILOSOPHY (Computer Science) December 1995 Copyright 1995 Danny James Mitzel UMI Number: 9617123 UMI Microform 9617123 Copyright 1996, by UMI Company. All rights reserved. This microform edition is protected against unauthorized copying under Title 17, United States Code. UMI 300 North Zeeb Road Ann Arbor, MI 48103 UNIVERSITY OF SOUTHERN CALIFORNIA T H E G RA D U A TE SCHOO L U NIVERSITY PARK LOS A N G ELES, CA LIFO R N IA 90007 This dissertation, written by under the direction of h.iS. Dissertation Committee, and approved by all its members, has been presented to and accepted by The Graduate School, in partial fulfillment of re quirements for the degree of DOCTOR OF PHILOSOPHY Dean of Graduate Studies D a te.... DISSERTATION COMMITTEE Chairperson A ck n o w led g em en ts I would like to begin by thanking my thesis supervisor, Dr. Deborah Estrin, for her guidance, support, and patience. It was during attendance at my very first, course taught by Professor Estrin that my interest was kindled in pursuing research in the field of computer networking protocols. During the past six years of working together, Deborah has put forth substantial time and effort in assisting me in crossing all the hurdles inherent in completing the Ph.D. program, and in honing my research skills. In addition, I am greatly indebted to Dr. Scott Shenker of Xerox PARC. The insights and encouragement Scott provided during my thesis work were invaluable. I would also like to acknowledge the contributions of the late Dr. Kim Korner and Dr. Peter Danzig at USC, and Dr. Lixia Zhang at Xerox PARC. In addition to Deborah, Professor Korner was a great influence during my initial studies of computer networking, distributed systems, and operating systems principles; he was also a great friend outside of the classroom. Professor Danzig helped me develop many of the underlying skills required to perform my research in computer system performance analysis and simulation; he also gave me my first opportunity to co author a published research paper, and his high level of enthusiasm was a constant inspiration. Much of Lixia’s early work on the R.SVP reservation protocol sparked my interest in pursuing research in this area, and she contributed a considerable amount of time and insight during my investigations. Other members of the computer science department faculty at USC who con tributed their time on my qualifying exam committee include Dr. Shahram Ghan- deharizadeli and Dr. Rafael Saavedra. I would especially like to thank Dr. John Silvester, of the electrical engineering department, who participated on both the qualifying exam and dissertation committees. Their contributions helped me focus my research plan and have enhanced the quality of this research. I am grateful for the financial support provided by the Hughes Aircraft Com pany Masters and Doctoral fellowship programs. Hughes provided not only financial support, but also an interesting work environment and the ability to tailor my work schedule to enable me to complete my coursework and research. None of this would have been possible without the support of my management, and I would like to thank them all for their support, including: Richard Loftus, Gladys and Gerry Hoshijo, Ron Gillet, Larry Hunt, Walt Cohn, Geoffrey Luke and Ed Lesnansky. Over the course of the many projects I have worked on at Hughes I have made many friends who have made the work more enjoyable and provided friendship outside of Hughes also; I would like to especially thank: Beckie Arakawa, Roger Barker, Richard Ching, Ken Cockerill, Theresa Distaso, Vickie Kubo, Cheri Lee, Keith Massey, Garrick Oka, Mary Parker, and Ron Watanabe. Working in the Network and Distributed Systems Laboratory at USC has pro vided a stimulating research environment and a great community of friends. I would like to thank all of the lab members, both old and new, whom I’ve shared experi ences with both in and outside the lab. There have been far too many people to completely list, but I would like to especially thank: Lee Breslau, Ron Cocchi, Doug Fang, Shai Herzog, Steve Hotz, Sugih Jamin, Abhijit Khale, Kraig Meyer, Louie Ramos, Gene Tsudik, and Daniel Zappala. Ron Cocchi and I shared a number of interests and experiences over the years, including the Hughes fellowship program, the same thesis advisor, our Ford Mustangs, our Sun SparcStations, downhill skiing, and the fine cuisine at a large number of greasy drive thrus; he’s been a great friend. Finally, I would like to recognize the first friend I met in the lab, Gary Frankel, who is dearly missed. Outside of work and school I have had a large support group of friends that have helped me enjoy my love of the beach, volleyball, the outdoors, and dive bars. There have been far too many people, too many road trips, and too many incidents to list, so I will just say a big thanks to all of the beach crowd. I would like to specifically thank my long list of roommates who have been responsible for maintaining my day-to-day sanity, especially: Doug Hehn, Dave Rowley, Brian (Clayton) Gardner, and Bob Koller. Finally I wish to thank my family for their continuing love and support. My parents, Herman and Karen, my brother, Dwayne, and my sister and brother-in- law, Marie and Tim Moore have been a source of encouragement and much needed camping and skiing trips over the years. C o n ten ts Acknowledgem ents ii List Of Tables viii List Of Figures ix A bstract xi 1 Introduction 1 1.1 Integrated Services Packet Network (ISPN) A r c h ite c tu r e ........................ 2 1.2 Resource R eserv atio n ............................................................................................. 4 1.3 Thesis O u tlin e........................................................................................................... 5 2 R elated W ork 8 2.1 Integrated Services Packet Networks .............................................................. 8 2.1.1 Flow Specification ................................................................................... 9 2.1.2 Reservation Establishment and S ig n a lin g ........................................ 11 2.1.3 Admission C o n t r o l ................................................................................... 14 2.1.3.1 Predictive Service..................................................................... 15 2.1.3.2 Advanced R eservations.......................................................... 15 2.1.4 Packet Service............................................................................................. 16 2.1.5 R o u t i n g ........................................................................................................ 18 2.1.5.1 Multicast R o u t i n g ................................................................. 18 2.1.5.2 QoS Sensitive R o u tin g .......................................................... 20 2.2 Resource Allocation Thrashing and D eadlock............................................... 21 2.2.1 Deadlock Detection and R eco v ery ...................................................... 21 2.2.2 Thrashing in Communication N etw orks........................................... 22 2.2.3 Thrashing in Database S y s t e m s .......................................................... 22 3 An Introduction to Novel Solutions to Resource Reservation 24 3.1 Protocol Overview ................................................................................................. 25 3.1.1 ST-II P rotocol............................................................................................. 26 3.1.2 RSVP P r o to c o l .......................................................................................... 28 v 3.2 Static Analysis ....................................................................................................... 3.2.1 Supporting Self-Limiting Applications .......................................... 3.2.2 Supporting Heterogeneous G ro u p s..................................................... 3.2.3 Supporting Channel S e le c tio n ............................................................ 3.3 Dynamic A nalysis................................................................................................... 3.3.1 Network D y n a m ic s ................................................................................. 3.3.2 Group-Membership D y n a m ic s ............................................................ 3.4 Chapter Summary ................................................................................................ 4 A sym ptotic Resource C onsum ption in M ulticast R eservation Styles 4.1 Network M o d e l ....................................................................................................... 4.2 Self-Limiting A p p lic a tio n s ................................................................................. 4.3 Channel Selection................................................................................................... 4.3.1 Assured Channel Selection A lternatives.......................................... 4.3.2 Dynamic Filter vs. Non-assured Selection O v e rh e a d .................. 4.3.2.1 Chosen Source Worst Case (C S worst) .............................. 4.3.2.2 Chosen Source Average Case (C S avg) .............................. 4.3.2.3 Chosen Source Best Case (CS/,est ) .................................... 4.4 Chapter Summary ................................................................................................ 5 Resource Reservation D ynam ics and Thrashing 5.1 T h r a s h in g ................................................................................................................. 5.2 Network M o d e l ....................................................................................................... 5.3 Point-to-point Reservations .............................................................................. 5.3.1 Point-to-Point Reservations with Blocked Reservation Retry . 5.4 Multipoint Session M o d e ls ................................................................................. 5.4.1 Bi-directional Session M o d e l............................................................... 5.4.2 Scaling Session S iz e ................................................................................. 5.4.2.1 Retry-all-receivers Session Retry P o licy......................... 5.4.2.2 Retry-blocked-receivers Session Retry P o lic y ................ 5.5 Reservation Retry B a c k o f f ................................................................................. 5.5.1 Unidirectional Point-to-Point Reservations with Exponential Retry B a c k o ff............................................................................................ 5.5.2 Session Reservations with Exponential Retry B ackofT .............. 5.6 Chapter Summary ................................................................................................ 6 Sum m ary and F uture W ork 6.1 Summary of C o n trib u tio n s................................................................................. 6.1.1 Reservation Protocol S caling ................................................................ 6.1.1.1 Service M o d e l s ....................................................................... 6.1.1.2 Dynamic R e s p o n s e ................................................................. 6.1.2 Reservation T h r a s h i n g .......................................................................... 30 30 33 35 37 38 39 42 43 46 49 51 54 55 55 56 58 58 60 61 62 64 66 70 71 72 73 75 76 77 78 80 81 SI 81 82 83 83 vi 6.2 Significance of Work ............................................................................................. 84 6.3 Future W o rk ............................................................................................................... 85 6.3.1 Reservation Protocol S caling ................................................................... 85 6.3.1.1 Detailed Q u e s tio n s ................................................................. 86 6.3.1.2 Fundamental Q u e s t i o n s ....................................................... 87 6.3.2 Reservation D y n a m ic s ............................................................................ 87 6.3.2.1 Detailed Q u e s tio n s ................................................................. 88 6.3.2.2 Fundamental Q u e s t i o n s ....................................................... 89 Reference List 90 vii L ist O f T ables 3.1 Network-wide resource requirements for ST-II and RSVP protocols in support of a 40 receiver audio lecture, with various numbers of low quality receivers......................................................................................................... 35 3.2 Network-wide resource allocation required by each of the channel se lection mechanisms in support of N-way conference.................................... 37 4.1 Summary of RSVP reservation styles................................................................. 45 4.2 Topological properties summary........................................................................... 48 4.3 Summary of resource allocations for self-limiting applications with N ■ — 1 r i 0 snnsrc — 1................................................................................................................. 4.4 Summary of resource allocations for assured channel selection with Nsini-clian — 1.............................................................................................................. 54 4.5 Summary of resource allocations for non-assured channel selection with — 1..................................................................................................... 56 5.1 Summary of reservation success and failures on the square topology for uni-directional point-to-point reservations with a one second reser vation request retry interval and 60 second teardown delay..................... 68 5.2 Summary of reservation success and failures on the binary tree topol ogy for uni-directional point-to-point reservations with a one second reservation request retry interval and 60 second teardown delay. . . . 69 5.3 Blocking probability for various N-way session sizes on the binary tree topology for no blocked reservation retry and immediate blocked reservation teardown................................................................................................ 73 5.4 Session blocking probability at thrashing onset point for various tree depths and N-way session sizes with a one second reservation request retry interval, one second blocked reservation teardown delay and retry-all-receivers session retry policy............................................................... 74 L ist O f F igu res 1.1 Components of Integrated Services Packet Network (ISPN) Architecture 3 3.1 Scaling of resource requirements in support of N-way audio conference using RSVP Shared Tree and ST-II Independent Streams reservation styles.............................................................................................................................. 31 3.2 Link reservations installed by RSVP in support of a heterogeneous audio lecture using Independent Trees reservations (40 receivers total, 20 “low quality” )...................................................................................................... 34 3.3 Network-wide protocol overhead for ST-II and RSVP supporting ho mogeneous receivers independently joining a multicast group................. 41 4.1 Network topologies considered in resource consumption analysis. . . . 47 4.2 Ratio of chosen source average and worst case for selected topologies. 57 5.1 Number of successful reservation requests on the square topology for uni-directional point-to-point reservations...........................................................65 5.2 Number of successful reservation requests on the binary tree topology for uni-directional point-to-point reservations....................................................66 5.3 Number of successful reservation requests on the square topology for uni-directional point-to-point reservations with a one second reserva tion request retry interval..................................................................................... 67 5.4 Number of successful reservation requests on the binary tree topol ogy for uni-directional point-to-point reservations with a one second reservation request retry interval........................................................................ 69 5.5 Number of successful session reservation requests on the binary tree topology for 2-way sessions with a one second reservation request retry interval and retry-all-receivers session retry policy...................................... 71 5.6 Number of successful session reservation requests on the binary tree topology for 4-way sessions with a one second reservation request retry interval and retry-all-receivers session retry policy...................................... 74 5.7 Number of successful session reservation requests on the binary tree topology for 2-way sessions with a one second reservation request retry interval and retry-blocked-receivers session retry policy............................. 76 ix 5.8 Number of successful reservation requests on the square topology with uni-clirectional point-to-point reservations with exponential retry backoff........................................................................................................................... 77 5.9 Number of successful session reservation requests on the binary tree topology for 2-way sessions with exponential retry backoff and retry- all-receivers session retry p o lic y ........................................................................ 78 5.10 Number of successful session reservation requests on the binary tree topology for 2-way sessions with exponential retry backoff and retry- blocked-receivers session retry policy................................................................. 79 x A b stra ct The Integrated Services Packet Network (ISPN) architecture is proposed to com bine the multiplexing, multipoint communication and robustness of packet switched networks with the service guarantees of the circuit switched model. The ISPN ar chitecture incorporates a resource reservation phase for those connections requiring Quality of Service (QoS) guarantees. The resource reservation protocol is a critical component of the ISPN architecture, requesting allocation and release of network resources along the data distribution path to ensure QoS requirements are met. The resulting network utilization and efficiency depends to a great extent on the reservation protocol’s service model and dynamic response. Studies of resource reservation protocol design and performance have been pri marily limited to the circuit switched network model. Novel reservation protocol mechanisms are required to support the large dynamic groups and heterogeneous receiver requirements introduced in the ISPN. This work presents results from an initial study of network resource reservation within the context of the ISPN archi tecture. We restrict our focus to two primary areas of investigation, protocol scaling and reservation dynamics. Our interest in protocol scaling is based upon the observation that scaling is the critical factor in support of services across a global internet. We discuss a number of novel resource reservation mechanisms proposed to support the ISPN and con trast their flexibility, network performance, and scaling characteristics to those of traditional protocols. We continue with a more rigorous analysis of the reservation styles mechanism, which defines rules to control aggregation of reservations at inter mediate switches. We show, using both simulations and analytical models, that the novel mechanisms explored can result in significant savings in support of a number of anticipated application classes. The incorporation of resource reservation and the diversity of possible intercon nections can lead to complex interdependencies. These interdependencies may lead to reduced system throughput and thrashing. In our study, we observe several neces sary conditions to induce thrashing. We use simulations to establish that thrashing can occur even in simplistic scenarios, and then evaluate the effects of support ing more complex multipoint conferences. We conclude by proposing a cooperative method for end-users or applications to improve system stability and throughput; our simulations demonstrate the effectiveness of the proposal. C h a p ter 1 In tro d u ctio n There is a growing desire in the communications field to provide integrated services networks utilizing a single cable-plant to carry multiple classes of traffic, each having its own set of traffic characteristics and performance requirements. This trend is being driven by requirements to integrate voice, video, and data communications, to support new applications, and by the promise of high bandwidth broadband networks in the near future, which will make it feasible to carry much larger volumes of data over a wide-area. Traditional data and telecommunications networks have taken very different ap proaches to providing their respective services. Data communication networks, such as those based on the T C P /IP protocol suite [96, 97], exhibit several distinctive characteristics. Datagram networks offer best-elfort service, which means that no guarantees are made as to when and whether packets will be delivered. Opera tionally, best-effort service is typically implemented using simple FIFO service at the network switches with no admission control or resource reservation. Recent work on multicast routing [27, 29] has shown that multicast data distribution can also be accommodated to increase efficiency in support of multipoint communica tions. Thus, the datagram network model can maximize network utilization by multiplexing multiple bursty data streams, can provide multipoint communication, and can provide robustness by adapting to network dynamics. However, datagram networks provide only the best-effort delivery service. In current circuit switched telecommunication and Integrated Services Digital Network (ISDN) [18] networks, where traffic is dominated by voice calls with tight delay-jitter bounds, Quality of Service (QoS) guarantees are provided. To meet the strict QoS constraints, current voice networks use a circuit model, allocating the maximum resources required for each connection during a setup phase; this ensures no other network activity can affect the capacity of the circuit. However, the circuit model leads to inefficient use of network resources when sending bursty data, does not adapt to link and router failures, and lacks support for multipoint communication. There are a number of application requirements that neither the traditional data gram nor circuit switched network model handle efficiently. For instance, applica tions such as Variable Bit Rate (VBR) voice and video, interactive conferencing, real-time sensing, and virtual reality often have stringent real-time requirements on the delivery of data. In addition to the QoS constraints, many of these applications can be much more efficiently accommodated in a network that directly supports bursty data streams and multipoint communication. Efficient integration of these diverse applications and traffic classes requires much more than just high bandwidth links. The goal of the Integrated Services Packet Network (ISPN) is to merge the two network paradigms; combining the multiplexing, multipoint communication and robustness of packet switched networks, with the service guarantees of the circuit switched model. 1.1 In teg ra ted S erv ices P a ck et N etw o rk (IS P N ) A rch itectu re There has been significant recent work [22, 42, 53, 59, 71, 92, 115] showing that by carefully scheduling packets, and utilizing admission control to prevent network overload, one can achieve real-time data delivery delay bounds in a packet-switched environment. It is widely believed that providing delay bounds only to those flows that need them is much more efficient than continuing to use FIFO scheduling and merely adding enough bandwidth to provide low delay service to all flows. This work has led to the definition of an ISPN architecture within the Internet research community [12], We assume an ISPN architecture similar to that proposed in [12] as the context for our work. The ISPN model has several distinct components, including: (1) a flow QoS Routing A voids low bandw idth, high delay < link that cannot satisfy Q oS requested Multicast Forwarding R eplicates each packet only once on each distribution link H2 H3 Reservation Setup < S 2 -S 3 . F lo w S p e o J^A c-cpt R cquestS < S 2 -H 3 , F l o w S p e o R cqucsi'v^R cJccl Call Admission A v a ila b le R e s o u rc e s S 2 -S 3 : 2 5 6 K b /s A v a ila b le R e s o u rc e s S 2 -H 3 : 3 2 K b /s 56K b/s Ltnk, 3(K)ms D elay 1.5M b/s Link. I Oms Delay Flow Specification M ax R a te - 6 4 K b /s M ax D e la y - 2 0 0 m s M ax L o s s R a te - 1% Packet Service D - D [ ] ^ ] 0 - * m O Q [ 3 [ ] D atagram T raffic Q R eal-T im e D ala Figure 1.1: Components of Integrated Services Packet Network (ISPN) Architecture specification defining the source traffic stream and receiver service requirements; (2) a routing protocol supporting QoS and multicast data distribution; (3) a reservation protocol to create and maintain resource reservations; (4) an admission control algo rithm to maintain network load at a proper level; and (5) a packet service algorithm to schedule packet transmissions in an order that maintains service guarantees for individual data streams. Figure 1.1 presents a simple illustration of how the ISPN components inter operate. The figure shows five host systems interconnected via three internetwork switches. In this example, an application on host HI is sending to the other group members. The data stream generated by the application at HI has specific character istics, such as average and peak transmission rate, delay constraints, and sensitivity to data loss; the flow specification is used to express these characteristics. The reservation protocol must request allocation of network resources along the data distribution path from HI to each of the other hosts. The figure shows the inter action between the reservation protocol and call admission at switch S2, where the reservation protocol must request resource allocation along links S2-S3 and S2-H3. Call admission must translate the resource request, presented as a flow specifica tion, into network level resource requirements (link bandwidth, router buffer space, etc.) and confirm that acceptance of the request will not violate current resource guarantees. In the example shown, the request on link S2-S3 is accepted, while the request for link S2-H3 is rejected due to insufficient resources currently available. Most existing routing protocols implement a shortest path algorithm. Note in the figure that the QoS-sensitive routing has avoided the H1-S3 link, even though it rep resents the shortest path from HI to hosts H4 and H5; this is because the H1-S3 link cannot satisfy the requested QoS. Multicast routing can improve network utilization by eliminating the sending of duplicate packets along a link when delivering data to multiple destinations; note at switch S3 in the figure only one copy of each data packet is received, each packet is then transmitted only once on each next hop link. Satisfying real-time constraints while maintaining high network utilization requires a sophisticated priority-based packet service algorithm. Note the example at switch S3, while real-time data and best-effort datagram traffic arrive intermixed, the real time data gets serviced first; this behavior is quite different than current First In First Out (FIFO) packet service. 1.2 R eso u rce R eserv a tio n The resource reservation protocol is a critical component of the ISPN architecture, it is responsible for requesting allocation and release of network resources along the data distribution path to ensure QoS requirements are met. The resulting network utilization and efficiency depends to a great extent on the reservation protocol’s ser vice model and dynamic response. Service models can be characterized by the set of communication styles (point-to-point, multipoint) and reservation styles (to control aggregation of reservations at intermediate switches) supported, and by the ability to support heterogeneous group members. The dynamic response of the reservation protocol can be characterized by the support for dynamic group memberships, and the response to link and router failures. Previous studies on resource reservation protocol design and performance have been primarily limited to the circuit switched network model, such as current tele phone networks. As mentioned earlier, current packet switched networks do not in corporate resource reservation. The traditional source-initiated point-to-point reser vation model between a static set of peer entities has proven adequate for the circuit switched network because the current use is limited to a single application, tele phony. However, the wider range of potential Internet applications require novel reservation protocol mechanisms to support the ISPN architecture efficiently. The source-initiated reservation model may introduce a processing or state bottleneck at the source, and may be inefficient in supporting the large, dynamic groups, and heterogeneous receiver requirements introduced in the ISPN. Note also that under the ISPN architecture resource reservation introduces a new form of resource contention for the shared network resources not addressed by previous studies of packet switched networks. Other domains that incorporate resource allocation (e.g., database systems) have uncovered phenomena - usually called thrashing - having great detrimental effects on overall system performance and throughput. 1.3 T h esis O u tlin e This dissertation represents an initial effort to study network resource reservation in support of application QoS requirements within the context of an Integrated Services Packet Network architecture. We restrict our focus to two primary areas of investigation, protocol scaling and reservation dynamics. A study of reservation protocol scaling has a large number of dimensions that can be explored, including network size and dynamics, and the number, size, and application group dynamics that can be accommodated. We quantify the effects that various reservation protocol design alternatives have on these scaling proper ties. We believe that scaling should be a primary consideration whenever evaluating reservation protocol design tradeoffs. This interest in scaling is based upon the ob servation that scaling is the critical constraint faced today in support of services across a global internet.1 Our interest in resource reservation dynamics derives from the realization that the ISPN architecture represents a new and previously unexplored domain. The incorporation of resource reservation into the packet switched network, and the di versity of possible interconnections introduced by multipoint applications lead to complex interdependencies among the individual reservations. These interdepen dencies may lead to reduced system throughput and thrashing. We believe it is critical to develop an understanding of these dynamics within the context of the ISPN architecture. The set of necessary conditions leading to thrashing onset must be identified, the effects of thrashing on system performance need to be quantified, and methods to increase system stability need to be identified and evaluated. The remainder of this dissertation is organized as follows. Chapter 2 surveys related work in the design and analysis of each of the individual ISPN architecture components. In addition, we also survey several studies on performance analysis and thrashing in other domains employing resource reservation and resource contention. Chapter 3 begins our analysis of reservation protocol scaling. We begin by intro ducing several of the novel features of a resource reservation protocol under design for use in the ISPN environment. We contrast the flexibility, network performance, and scaling characteristics of the novel mechanisms to those of a more traditional reser vation protocol. This work establishes the importance of the resource reservation protocol design on overall system performance, and demonstrates the effectiveness of the novel mechanisms in improving scalability within the ISPN architecture. In Chapter 4 we continue our study of reservation protocol scaling by more closely focusing on a single novel mechanism that shows great promise in improving reservation efficiency. The mechanism we analyze is reservation styles, which define rules to control aggregation of reservations at intermediate switches. Because this 1 As an exam ple of the rapid growth in network size driving this requirem ent for scalable so lutions, consider the T C P /IP Internet. In [110] it was reported that when originally established in 1988, 400 networks were reachable through the U.S. NSFNet, backbone. W ith network growth resulting in a doubling o f this num ber every nine m onths, more than 20,000 networks were inter connected by 1994. Current projections are for over 100,000 interconnected networks by 1996, and greater than one m illion by the year 2000. 6 is a novel mechanism and has not been previously analyzed its performance is not well understood. We undertake a rigorous evaluation of reservation styles under a number of network topologies and several application models and quantify the effects of the different styles on network scaling and efficiency. Chapter 5 presents the results of our investigation into resource reservation dy namics and thrashing. Our investigation is quite novel in the fact that we are assuming the ISPN architecture; including long delays and multipoint-to-multipoint applications. We observe several necessary conditions to induce thrashing. We use simulations to establish that thrashing can occur even in simplistic scenarios, and then evaluate the effects of supporting more complex multipoint conferences. We conclude by proposing a cooperative method for end-users or applications to improve system stability and throughput; our simulations demonstrate the effectiveness of the proposal. Finally, in Chapter 6 we summarize the contributions of our work. We also present directions for future work under the topics of reservation protocol scaling analysis, and resource reservation dynamics and thrashing. C h a p ter 2 R e la te d W ork In this chapter we explore technical issues related to the development of the Inte grated Services Packet Network architecture. We survey related work in the design and analysis of each of the individual ISPN architecture components; flow specifi cation (Section 2.1.1), resource reservation (Section 2.1.2), admission control (Sec tion 2.1.3), packet service algorithms (Section 2.1.4), and multicast and QoS routing (Section 2.1.5). Reservation protocols introduced in Section 2.1.2 form the basis for our investigations on protocol scaling. We also survey several studies related to the analysis of resource contention and thrashing (Section 2.2); these motivate the development of our thrashing model. 2.1 In teg ra ted S erv ices P a ck et N etw o rk s Previously, much of the research in communications focused on designing networks for a specific technology. This has led to the development of separate hardware and protocols for telephone networks, data communication networks, and cable TV distribution. This development of distinct technologies has made integration of services more difficult. For example, the current telephone network is adequate fol low bandwidth fixed rate traffic. However, it is highly inefficient in supporting bursty data traffic or high rate video streams. The trend in integrating voice and data communications is being driven by sev eral factors. Integrated service networks can be cheaper and more efficient to design than to maintain distinct networks for each service [114, 119]. A second factor is im provements in packet switching technology to support a wider range of applications 8 [54, 100]. Finally, the arrival of high bandwidth broadband networks in the near future will make it feasible to carry much larger volumes of data over a wide-area [4, 107, 119]. Early work on Integrated Services Digital Network (ISDN) protocols [18, 77] led to systems providing relatively narrowband voice and data services. The ISDN protocols developed were still inefficient in supporting the QoS, high rate bursty data streams, and multipoint communication recpiirements of many applications. Additional research on integrating support for packetized voice, video, and data [20, 22, 42, 72, 88] has led to the development of an Integrated Services Packet Network architecture within the Internet research community [12]. The Internet ISPN architecture is based upon the current packet switched tech nology and multicast data delivery, with additions to the service model to support both best-effort and real-time services. The service model defines the externally visible behavior of the system, and includes three levels of service: best-effort, guar anteed, and predictive. Best-effort data deliver is the current service model support ing non-real-time applications. Guaranteed service offers a perfectly reliable upper bound on delay. Guaranteed service is appropriate for intolerant playback applica tions; these are applications that must use a fixed offset delay to avoid distortion in the playback. Predictive service supplies a fairly reliable, but not perfectly reliable, delay bound. Predictive service is appropriate for adaptive playback applications; these are applications that can tolerate some late packets and may adapt their offset delay to reduce latency. The realization of the Internet ISPN service model has five components: (1) flow specification, (2) resource reservation, (3) admission control, (4) packet service, and (5) multicast and QoS routing. In subsequent sections we discuss technical issues related to the design and analysis of the components of the ISPN architecture. 2.1.1 Flow Specification Before an application can characterize its data stream or request a specific service level from the network, a grammar must be defined to express these quantities. Current circuit switched networks providing QoS guarantees do not require this specification, the entire network is constructed to provide a single homogeneous 9 service level to every circuit. A goal of the ISPN is to provide the flexibility to support the broad range of possible data rates and service levels that may be required by any future application. The flow specification (or “flow spec” ) is required to express this service request. Much of the early work on providing QoS guarantees in packet switched networks focused on defining the packet service discipline. In these studies the data streams and service requirements were typically expressed in terms directly implemented by the packet service algorithm. The traffic specification in Virtual Clock [124], Hierarchical Round Robin (HRR) [71] and Stop-and-Go [53] are essentially the same: a transmission rate (AR) averaged over an interval (.47). The Delay-EDD and Jitter- EDD [42] service disciplines require three traffic parameters: X mi„ is the minimum packet inter-arrival time, X arg is an average packet inter-arrival time, and I is the interval over which X avg is computed. The DASH resource model [2, 3] uses a Linear Bounded Arrival Process (LBAP) abstraction based on maximum message size {Smtir bytes), maximum message rate (RmaT messages per second), and maximum burst size (Bmax messages). Finally, many leaky bucket based service disciplines [20, 22, 98, 114] characterize service requests in terms of a bucket size (b) and service rate (r). There are several problems with the schemes coupling the flow specification di rectly to the packet service algorithm parameters. The parameters understood by the queueing discipline might not be the most intuitive for applications to express their service requirements. Also, the diversity in parameters across the multiple service disciplines leads to a plethora of application characterizations that must be mastered when porting an application across service models. Finally, the most crit ical deficiency is that a single packet service discipline must be deployed across all routers in the ISPN. During work on development of the Internet STream (ST) protocol [47, 113] it was recognized that application service requirements should be abstracted, using a common data structure for applications to request services and then requiring each packet service mechanism to perform the mapping into its service parameters. A problem was that knowledge of the flow spec format was tied into the protocol, and no mechanism for extension was included. Partridge [94] defines an initial flow spec data structure for consideration in the Internet. More recent work in the Internet 10 Engineering Task Force (IETF) Integrated Services work group resulted in the defi nition of a service request d ata structure that is independent of any reservation setup protocol or packet service mechanism, and provides extensibility for experimenta tion and evolution [121]. It has also been determined that it is useful to distinguish between the flow specs describing the aggregate source traffic (TSpec) and the re ceiver service requirements (RSpec); this leads to added flexibility in defining service definitions [103, 104, 105, 106]. 2.1.2 R eservation E stab lish m en t and Signaling The resource reservation protocol is responsible for requesting allocation and release of network resources along the data distribution path to ensure QoS requirements are met. In this section we briefly survey previous work in reservation protocol design. Several of the protocols introduced are covered in greater detail in Section 3.1, where they act as the basis for our comparative investigation of reservation mechanism alternatives. Resource reservation procedures have traditionally been associated with circuit setup in telecommunication networks, where the process is referred to as signaling. Although the point-to-point fixed-bandwidth service model of Plain Old Telephone Service (POTS) and data communication over Public Data Networks (PDN) (such as X.25 [17]) has remained fixed, the underlying signaling protocols have evolved. Early signaling protocols required the end-user to physically generate a request signal (ring down) and verbally communicate addressing information to an operator; this evolved to automated setup using stimulus signaling; finally evolving to functional signaling based on packet switching technology. Functional signaling is currently used primarily for “common channel” inter-office switching (such as the CCITT Common Channel Signaling System No. 7 (SS7) protocol [122]), making possible additional features such as improved 800 services and 900 numbers. Narrowband ISDN (NISDN) [18] uses the Q.931 Digital Subscriber Signaling System No. 1 (DSS 1) [64] to extend functional signaling to the user; however, the NISDN service model continues to be limited to point-to-point fixed-bandwidth circuits. Broadband ISDN (BISDN) proposals based on the Asynchronous Transfer Mode (ATM) [61] may pro vide additional service models. Q.931 can not be used in the BISDN environment 11 as it does not support the establishment of ATM virtual path or virtual channels. ATM-based BISDN relies on the Digital Subscriber Signaling System No. 2 (DSS 2) in the Recommendation Q.2931 [62] for basic connection setup. Additional rec ommendations have been proposed to add extensions, such as multicast [63] and Quality of Service parameter negotiation [65, 66, 67]. The lack of accepted standards in early work on the BISDN signaling led to development of vendor proprietary signaling protocols [45, 46]. The ATM Forum consortium has attem pted to increase interoperability by developing a number of implementors agreements. This has resulted in the ATM User Network Interface Specification, Version 3.0 (UNI 3.0) [48] based on the Q.93B proposal [60], and the UNI 3.1 [49] updated for the newer Q.2931 recommendation. In addition, the lack of multiparty support in the original Q.931 ISDN signaling protocol led to proposals to extend the service model. The Connection Management Access Protocol (CMAP) [15. 33] is concerned primarily with adding multipoint communication. Connection setup begins with the establishment of one controlling end-point (the “owner” ), other participants are added at the recjuest of the owner, client, or a third party; in addition, receiver and transm itter attributes can be assigned to each participant. The EXPANSE signaling protocol [84] provides not only multiparty communication, but also control over application mapping and presentation (e.g., conversion between multiple media type within the network), and the ability to define synchronization constraints among logical connections. Early work in supporting guaranteed service and multipoint communication in the Internet community resulted in development of the Internet STream protocol (ST) [47]. ST supports unicast and N-way duplex (omniplex) connections; however, it requires a centralized Access Controller to coordinate among the participants. Knowledge gained from experience with ST was incorporated into the follow on ST- II protocol [113], the primary difference in service model was the simplification of the stream model to a source-rooted point-to-multipoint simplex data flow. Stream setup is initiated by the source sending a Connect message to all initial receivers; the source must then wait for a Reply from each receiver before transmitting data. Each Reply indicates the actual resources allocated to the receiver, the source must then treat the stream as a homogeneous distribution path, forcing it to conform to the least demanding or least capable path/receiver. Group membership dynamics are 12 accommodated by allowing the source to add new receivers via additional Connect messages; receivers may be deleted by either the receiver sending a Refuse message or by the source sending a Disconnect message. Recent work on the ST-II protocol has attem pted to formalize the protocol specification to eliminate ambiguities uncovered during several prototype implementation efforts [35, 58, 93], and to incorporate new features such as receiver initiated reservations [36]. This has resulted in the ST2+ proposal [34]. We discuss the details of the ST-II protocol in much greater depth in Section 3.1.1. Anderson et al. [2] propose the Session Reservation Protocol (SRP) for perform ing reservation setup in an IP network. Setup requires an end-to-end data exchange with the source sending a Request specifying the flow specification, and the receiver returning a Reply indicating its service requirements. This setup mechanism allows the receiver to express its local service requirements; however, the unidirectional point-to-point service model, and tight integration with the DASH resource model [3] place severe limitations on its flexibility. The real-time channel establishment protocol proposed by Ferrari and incorporated into the Tenet protocol suite [10, 42] similarly use an end-to-end messaging protocol. A recent reservation setup protocol proposed in support of the the ISPN archi tecture is the Resource ReSerVation Protocol (RSVP) [13, 123]. RSVP is primarily a vehicle for establishing and maintaining state in routers along the data distri bution path. RSVP state include Path state and Reservation state. Path state includes information regarding senders using the router and is used to propagate control messages toward specific senders. Reservation state includes information on network resources reserved and which sender(s) can use the resources. A soft state1 mechanism is employed where Path and Reservation state are refreshed periodically by active senders and receivers; this feature automatically adapts to changes in net work and group membership. A provision to “merge” control messages is included 'C lark [21] characterizes the concept o f soft state in support o f type o f service as follows; “It, would be necessary for the gatew ays to have flow state in order to remem ber the nature o f the flows which are passing through them , but the state inform ation would not be critical in m aintaining the desired type o f service associated w ith the flow. Instead, th at type of service would be enforced by the end points, which would periodically send m essages to ensure that the proper type o f service was being associated with the flow. In this way, the sta te inform ation associated with the flow could be lost in a crash w ithout perm anent disruption o f the service features being used.” to reduce control traffic overhead to one message on each link during each refresh period. Reservations are modeled as a simplex flow of data from the sender to the receivers. Reservations are receiver-initiated; receivers choose the level of resources required and are responsible for initiating and maintain their reservations. Reser vation requests are only propagated until it “splices” into the distribution tree at a point where sufficient resources are allocated to meet the requested QoS. Application- level service requirements are captured within RSVP by offering several reservation styles, these dictate how reservations may be aggregated within the network. Shared Tree reservations2 allow any source to use the reserved resources; Independent Trees reservations setup resources to a fixed set of sources and allow aggregation of re sources across multiple receivers for a common source; Dynamic Filter reservations allow a receiver to dynamically switch its resources among different sources. We discuss the details of the RSVP protocol in much greater depth in Section 3.1.2. 2.1.3 A dm ission C ontrol Admission control implements the decision algorithm that a router or host uses to determine whether a new flow request can be granted the requested QoS without impacting earlier guarantees. This is done by converting the requested flowspec into an equivalent buffer space, bandwidth and scheduling slots, and confirming that none of the resources is over-booked. The guaranteed class of service offers a perfectly reliable upper bound on delay, it is appropriate for those applications requiring hard bounds on packet loss and delay characteristics. For services offering these deterministic guarantees the allocated bandwidth must not exceed the bandwidth of the link. Schedulability to provide delay guarantees must also be considered. For those service disciplines with a natural regeneration cycle [53, 71] the schedulability constraint is automatically met provided bandwidth in each frame time is not over-booked. In schemes without a natural regeneration point [42] the test must construct the worst case packet arrival pattern from all channels and examine if all packets can meet their deadlines. “T he term inology o f the reservation styles in R SV P is in flux, so here vve adopt a som ew hat independent term inology to avoid direct inconsistencies. T he Shared Tree reservation style is currently called W ildcard Filter in [13]. 2.1.3.1 P redictive Service A problem with systems offering only the guaranteed service is that it may lead to low network utilization. The hard bound in guaranteed service reflects the worst- case behavior of network traffic, this results in over-reservation for bursty streams. In addition, many applications can tolerate some level of intermittent loss, thus the hard guarantees are not essential. These observations have led to the definition of a predictive service class [22], which attem pts to minimize the post facto delay bound but makes no absolute guarantees. The advantage of offering the predictive service is that one can more fully utilize the network. The disadvantage is that admission control becomes much more complex, as the decision is no longer based on the simple worst-case characterization of traffic. Jamin et al. [69, 70] define an admission control scheme for the CSZ service model [22]; including guaranteed, datagram, and multiple levels of predictive service. When making an admission control decision the proposed scheme uses the a priori source characterization only for the incoming flows; it uses measurements to characterize those flows that have been in place for a reasonable duration. This measurement- based admission control scheme assumes that time-averaged measurement of the actual traffic is a valid predictor of future behavior, and is made possible by the loose performance bounds and application adaptability assumed for the predictive service classes. Simulation results presented in [70] clearly show that the combination of predictive service and the proposed admission control scheme can result in much greater network utilization than the guaranteed service alone, especially with bursty data streams. Predictive delay bound violations are also shown to be minimal. 2.1.3.2 Advanced R eservations Admission control decisions and resource allocation are typically discussed within the context of being performed immediately at the time of application startup. An interesting alternative is to consider the effects of advance reservations; allowing resources to be “booked” for use at a specified time in the future without actually allocating any resources at the initial request time. Ferrari et al. [43] propose a scheme for advance reservations within the Tenet protocol suite [10] by augmenting the setup request with: (i) the starting time, and 15 (ii) the duration. Network resources are divided into an immediate partition and an advance partition, however fragmentation is minimized using a movable boundary algorithm. Future reservation state is distributed among the network switching elements and maintained in an efficient representation called an interval table; no provision for state recovery after a crash is currently defined. The resulting scheme is highly efficient and flexible in terms of granularity of start times and durations. Reinhardt [99] and Delgrossi et al. [37] discuss advanced reservations in the context of the ST-II [113] protocol; their signaling is similar except for the use of fixed-size time intervals. Degermark et al. [32] propose extensions to the measurement-based admission control of [69, 70] to introduce advance reservation of predictive service flows. The reservation request interface is extended to specify start and end times. Admission decisions for predictive service are now based on measured rates for active flows and biased by the requested rate for any overlapping advance reservations. State aggregation methods are also proposed to reduce the scaling of state information required. 2.1.4 Packet Service Traditional data networks incorporate a First Come First Served (FCFS) packet service discipline, this introduces many problems when attem pting to provide ser vice guarantees. Based on packet arrival ordering and network load level, individual data packets may exhibit large variations in end-to-end delay and packet loss rates; providing service guarantees requires some form of priority queueing. Three levels of service guarantees can be defined: deterministic, statistical, and best-effort. A deterministic service is an absolute guarantee required to meet hard real-time con straints; a statistical guarantee provides a probabilistic bound on failure to meet the guarantee; while best-effort only promises that the system will try its best to satisfy the requested service. Several queueing disciplines have been investigated that queue packets according to static priorities and service the packets within a priority class FCFS [23, 74, 76]. These policies are attractive due to their ease of implementation, however high priority traffic can monopolize the output channel for long periods. This leads to 16 degraded service for lower priority traffic; no maximum delay guarantees can be enforced. Schemes to dynamically increment priorities as waiting time increases have been explored to increase fairness [6, 74, 75]. The Fair Queueing [38, 89] service discipline demonstrated that non-FC'FS service could be incorporated into a packet switched network to better allocate a uniform share of resources among all data streams. Fair Queueing conceptually allocates a separate queue to each conversation traversing the router; it only provides increased fairness, no real-time service guarantees are made. Zhang’s VirtualClock service pro tocol [124, 125] attem pts to emulate Time Division Multiplexing (TDM); it is similar to Fair Queueing, however it allows each flow to specify an average arrival interval, providing for non-uniform service allocations. VirtualClock can enforce long-term average resource guarantees, but can not provide hard real-time guarantees. Much of the research on providing bandwidth enforcement for ATM networks has focused on the use of the Leaky Bucket service discipline [20, 98, 114]. The leaky bucket service requires each cell to obtain a token before transmission; the service can be characterized by two parameters, the average data rate (r) and a token bucket size (6). Tokens are generated at rate r, if the bucket fills tokens are discarded; this method enforces the average input rate while allowing for a certain degree of burstiness (defined by the bucket size). Leaky bucket is more precisely a rate control algorithm. Parekh defines the Generalized Processor Sharing (GPS) [92] service discipline that, when combined with leaky bucket rate control, provides worst-case service guarantees. GPS is functionally equivalent to the Fair Queueing service discipline mentioned earlier, with the exception that each data stream may request a non-uniform resource allocation (weighted fair queueing). Ferrari and Verma [42] define a mechanism to establish a “channel” across a wide-area network providing either deterministic or statistical bounds on maximum end-to-end delays. Packet service is deadline based using a modified Earliest Due Date (EDD) policy [57], which gives highest priority to deterministic traffic. Clark, Shenker, and Zhang [22] define an architecture supporting both guaranteed and predictive service classes. Guaranteed service is provided by Parekh’s GPS service discipline described earlier; while additional service classes use a modified First Come First Served algorithm called F1FO+. The GPS discipline provides isolation between the guaranteed service classes and predictive classes, while FIFO + provides 17 sharing among flows within a predictive class. This architecture attem pts to provide a small number of flows supporting hard real-time guarantees, while supporting more adaptive applications via the predictive service to increase network utilization. [44, 68] defines a link-sharing architecture. Packets are classified based on a hierarchy of user classes, the packet service discipline attem pts to allocate bandwidth fairly among the data flows in each class. This scheme provides a long-term “fair share” of bandwidth among the classes, while allowing classes to dynamically “borrow” bandwidth from other classes to accommodate short-term burstiness. 2.1.5 R outing The ISPN architecture places two critical demands on the network routing infras tructure; the need for multipoint data distribution and QoS sensitive routes. 2.1.5.1 M ulticast R outing Multicast routing is an enabling technology in supporting multipoint communication. Using the traditional point-to-point service model, a separate packet is sent to each receiver (we call this simultaneous unicasts); multiple copies of packets are sent over the overlapping portions of the routes to individual receivers. Multicast routing solves this problem by having the network, whenever the routes diverge, send a single copy along each path. Early work in the Internet research community on supporting multipoint commu nication resulted in the development of the Internet STream (ST) [47] and follow on ST-II [113] pi'otocols. Data distribution is based upon a source-rooted uni-directional distribution tree using multicast forwarding. Explicit end-to-end messaging is re quired to control group membership, this can lead to inefficiencies and processing bottlenecks. Deering [27, 29] proposes extensions to distance-vector and link-state routing algorithms to build truncated-broadcast and multicast routing tables. The routing tables calculated produce a shortest-path tree to each receiver. Novel concepts introduced by Deering include the notion of a host group and receiver-initiated group join. Rather than identifying individual receivers, the source addresses each packet to a single destination group address; thus, the source has no knowledge of the 18 set of receiver hosts. A host membership protocol allows any receiver to request membership in a group. A variation of Deering’s truncated-broadcast distance- vector algorithm, documented in [116], is in use today in the experimental Internet Multicast backbone (MBONE), the group membership protocol in use is documented in [26]. Despite its success to date in the experimental MBONE, there are several well known scaling constraints inherent in the current truncated-broadcast and pruning protocols. The Internet Open Shortest Path First (OSPF) routing protocol [87] is a link- state routing protocol. Work on incorporating multicast extensions into the protocol has been completed (MOSPF) [86], and in fact several end sites participating in MBONE experiments have been running it internally to their sites. However, the OSPF (and M OSPF extensions) were explicitly designed to operate in the capacity of an Intra-Domain routing protocol and suffer from several well known scaling limitations of link-state routing algorithms. Current work on defining a scalable multicast routing protocol has been placed under the auspice of the IETF Inter-Domain Multicast Routing (IDMR) working group. At present two competing protocol designs are under consideration. The Core Based Trees (CBT) protocol [7, 8, 9] adopts Deering’s host group abstrac tion and receiver-initiated group join, but seeks to reduce overhead in tree setup and dependencies between underlying routing protocols. Under this architecture a core-router is defined towards which data packets are transmitted; multicast dis tribution then branches out from the core node. This routing architecture reduces network-wide overhead at the cost of possibly longer delays. The Protocol Indepen dent Multicast (PIM) proposal [28, 30, 31] also utilizes Deering’s abstractions, and achieves independence from underlying unicast routing protocols. PIM includes a sparse mode, intended for use in wide-area multicast routing. PIM sparse mode uses a Rendezvous Point (RP) to form a shared tree but supports optional switching to a shortest path tree if traffic warrants it; a soft state protocol is employed to maintain state information. There have been several studies conducted analyzing the tradeoffs and dynamics of the IDMR multicast distribution alternatives [117, 118], however no definitive conclusions or selection have been made at this time. 19 2.1.5.2 QoS Sensitive R outing Most current packet switched routing protocols typically base next-hop selection on a shortest path algorithm [56, 80, 87, 91, 116]. Given the heterogeneous and widely varying capabilities of links within a large internetwork and the widely varying ap plication resource requirements, this can lead to selection of next hop links which may be incapable of accommodating the requested resource level. Quality of service sensitive routing attem pts to increase the probability of successful end-to-end reser vation establishment by biasing route selection toward a path predicted capable of accommodating the requested resource level. Incorporating QoS support into the routing architecture requires additional in formation in routing updates beyond simple connectivity. The Internet Protocol [96] defines a Type Of Service (TOS) field, and several other routing protocols such as the Internet OSPF [87] and OSI IS-IS [91] can provide separate next-hops for each TOS. However, this mechanism has seen little use so far, and it is not known how well it would work in practice. Although not designed specifically for QoS routing, protocols incorporating policy constraints [111] may be able to provide limited QoS routing support. In [14] a source routed protocol is analyzed. An advantage to the source routed technique is that multiple routes can be computed to a single des tination to reduce traffic concentration along the route, further increasing network throughput and utilization. [109] presents another QoS routing protocol for packet switched networks, which is based closely on concepts developed for alternate path routing in telephone networks [1, 5, 81]. Requests blocked on their primary path are routed onto alternate paths. Simulation results show this scheme results in im proved performance, given simple Poisson assumptions and only local knowledge at each switch. The QoS support provided by the protocols described above are all based on static information. Dynamically adapting to measured network loading can lead to oscillations, as found in the early ARPANET experiments [73]. [50, 82] have begun preliminary investigations of adaptation to dynamic QoS information. A strength of packet switched network routing is its ability to adapt to network dynamics. However, once a satisfactory route is obtained providing a requested end-to-end QoS level, it may not be desirable to always adapt to changes in the 20 routing state unless absolutely necessary. A change to a routing metric does not always indicate failure of a route, it may have been caused by a new shortest path becoming available. The proposal defined in [40] represents a first attem pt at defining the routing support and interface to the resource reservation protocol required to support the ISPN environment. Routing support functions are defined to provide notification of routing state changes, to pin a route in place once established, and to request additional next hops after a setup request has been rejected. These features are meant to enable the resource reservation mechanism to maintain established QoS guarantees once established, while allowing adaptation to network dynamics to occur in a controlled manner. 2.2 R eso u rce A llo c a tio n T h ra sh in g and D ea d lo ck Addition of resource reservation into the Integrated Services Packet Network to support QoS guarantees introduces a new form of resource contention for shared network resources. W ith reservations, admission control will deny access if there arc not sufficient unreserved resources available. Previous work in other domains in corporating resource allocation (e.g., database systems) has uncovered phenomena - usually called thrashing - having great detrimental effects on overall system per formance and throughput. Introduction of resource dependency circularity among independent reservations also raises the possibility of system deadlock. In this section we survey a number of previous studies in deadlock detection, thrashing, and distributed database resource locking. Several of the concepts pre sented motivate the establishment of our thrashing model in Section 5.1. 2.2.1 D eadlock D etectio n and R ecovery There is a large body of research in deadlock detection and recovery in centralized systems, Coffman, Elphick, and Shoshani [24] survey many such algorithms. Per forming deadlock detection in a distributed system introduces additional difficulties due to the lack of shared memory, and non-negligible delays in message delivery. One approach to distributed deadlock detection is to perform a straight forward extension 21 of a centralized algorithm, constructing and maintaining a resource graph at each node. Menasce and Muntz [83] propose one such scheme; however, Gligor and Shat- tuck [51] note several inconsistencies that can be introduced due to arbitrary delays and out-of-order messages. Goldman [52] proposes a distributed algorithm based on the Ordered Blocked Process List (OBPL), this protocol reduces constraints on synchronized state across the nodes; however, it incurs the expense of creating the OBPL every time deadlock detection is invoked, and has a restriction preventing multiple resource requests. 2.2.2 T hrashing in C om m unication N etw orks Studies of resource allocation in communications networks typically employ analyt ical modeling techniques. Dziong and Mason [39] model a system providing call admission, routing, and Grade of Service (GOS) using the policy iteration algorithm from the theory of Markov decision processes. Bernabei et al. [11] model the perfor mance of an ATM switch implementing the Circuit Emulation Mode (GEM). The fully distributed “flooding” set-up technique is analyzed using an M /G /l/K queue ing model. Chesnais et al. [19] propose a method for studying shared data access and distributed deadlock detection via simulation. Much research in telecommunication networks has studied the effects of alter nate path, routing [1, 5, 41, 81], The work of Akinpelu [1] uses simulation to show a form of thrashing in nonhierarchical networks using a simple alternate path algo rithm. The analysis showed instabilities in trunk-group blocking probability when considered as a function of offered load. Two realizable states were found: one state showed low network blocking with almost all calls using their shorter first-choice path; the alternate, congested state, showed a large proportion of calls using their longer alternate paths and many calls being blocked. Additional scenarios investi gated included engineered networks and trunk reservation, these did not exhibit the instabilities under overload conditions. 2.2.3 T hrashing in D atab ase S ystem s The definition of the concept of a transaction [90] enabled consistency and deadlock to be studied in database systems. A transaction is an abstraction for the processing 99 required to take a system from one consistent state to the next consistent state. Each transaction requires the request and allocation of a sequence of locks. Parallels between resource allocation for transactions and end-to-end reservations in ISPN suggests that similar phenomena may be observed. A number of studies have been conducted to analyze performance of static and dynamic locking in distributed database [101, 108, 112, 120]. Typical distributed database deadlock detection algorithms are based on creating a Transaction Wait-For Graph (TW FG), which represents the dependencies between transactions waiting for allocation of resources. Obermarck [90] defines a TW FG-based algorithm where the graph is calculated on each invocation of deadlock detection, while Maekawa [78] describes a continuous deadlock-detection algorithm. The work of Tay [112] demonstrates a form of thrashing in database systems. Tay showed that with a fixed transaction length and total resources, scaling re source demand results in an initial increase in throughput to a point and then a steady decrease due to contention. This work demonstrates that significant system degradation is possible even without deadlock. Two contention resolution scenarios were investigated: in the No Waiting case a transaction is immediately restarted; in the Waiting case a transaction blocks and waits for the next resource to become available, rather than restarting the entire transaction. Both contention resolution scenarios exhibited the thrashing phenomena, however the Waiting case can also enter the deadlock state if a transaction abort timeout is not included. C h ap ter 3 A n In tro d u ctio n to N o v e l S o lu tio n s to R eso u rce R eserv a tio n The reservation protocol is responsible for requesting allocation and release of net work resources along the data distribution path to ensure QoS requirements, are met. The resulting network utilization and efficiency depends to a great extent on the reservation protocol’s service model and dynamic response. Service models can be characterized by the set of communication styles (point-to-point, multipoint) and reservation styles (to control aggregation of reservations at intermediate switches) supported, and by the ability to support heterogeneous group members. The dy namic response of the reservation protocol can be characterized by the support for dynamic group memberships and the response to link and router failures. Initial work in supporting multicast end-to-end guaranteed service within the Internet protocol suite resulted in the development of the Internet STream (ST) pro tocol [47], and the later development of a second version of the protocol, ST-II [113], which was specified as an experimental protocol within the Internet community. The reservation establishment portion of the ST-II protocol, as will be described in more detail in Section 3.1.1, incorporates a number of features from the traditional circuit switched signaling protocols. These features include source initiated reservations, and end-to-end reliable messaging. A more recent proposal targeted at supporting the resource reservation requirements of the ISPN architecture is the RSVP protocol [13, 123]. The RSVP protocol is currently in the design phase, an IETF working group has been formed to evolve the protocol along the standard track. It incor porates several novel features designed to improve reservation protocol scaling and flexibility in support of the ISPN. This chapter acts as an introduction to several of the novel resource reservation protocol mechanisms proposed to support the ISPN architecture. We outline the novel components of the proposed reservation protocol and introduce our m ethod ology for investigating their scaling properties. Next, we define several application scenarios we expect to be typical of an ISPN and analyze the scaling behavior of the novel reservation protocol components, and contrast their performance to that of the more traditional signaling protocol mechanisms. In the remainder of this chapter we investigate the operation of the ST-II and RSVP protocols in support of applications that we expect will be typical of an ISPN. In Section 3.1 we present an overview of the two protocols. We divide our comparison of the protocols into two distinct topics: static resource requirements and dynamic behavior. The static resource investigation looks at network resource requirements to support a fixed set of communicating applications over a range of communication styles; these results are presented in Section 3.2. In Section 3.3 we describe the protocol mechanisms for supporting network dynamics, and look at the protocol overhead associated with accommodating group membership dynamics. Section 3.4 concludes with a summary of our findings. 3.1 P r o to c o l O v erview A simplistic reservation service could be implemented on top of a point-to-point service model by establishing a separate reservation between each pair of commu nicating applications. This service model might be sufficient if the only goal of the ISPN was to extend the current IP point-to-point service model with QoS support; however, the goal of the ISPN architecture is to provide efficient support for ap plications requiring QoS support and multipoint communications. As we shall see the simplistic point-to-point reservation mechanism is very inefficient in terms of network resource allocation required to support multipoint communications. An enabling technology for supporting multipoint communication incorporated into both ST-II and RSVP is multicast routing. Deering [27, 29] describes how mul ticast distribution can be incorporated into a datagram network to improve network resource utilization; however, ST-II and RSVP make different assumptions about the level of multicast support provided by the network. ST-II builds a multicast distribution tree based upon unicast routing tables, and performs the replication and forwarding of data packets. RSVP is decoupled from the multicast routing and data forwarding functions; it assumes they are provided by the underlying network. This difference in assumptions about the level of multicast support provided by the network is largely historical. At the time ST-II was developed there was no inter network multicast routing. While incorporating multicast forwarding into the ST-II protocol adds some processing overhead it does not affect the resource allocation or protocol messaging overhead and thus does not affect our comparative analysis. The multipoint communication capabilities of ST-II and RSVP provide improved network resource utilization when compared to the simplistic point-to-point reserva tion model. Additional gains in terms of improved resource utilization are possible by incorporating application-level communication requirements into the reservation service model. In the following subsections an overview of the ST-II and RSVP reservation protocol and service model are presented; these act as the basis for our comparisons throughout the remainder of the chapter. It should be emphasized that these descriptions provide only a summary of the protocol functions relevant to the discussion. For a complete protocol description the appropriate protocol documents should be consulted. 3.1.1 ST -II P rotocol ST-II [113] models a resource reservation as a simplex data stream rooted at the source and extending to all receivers via a multicast distribution tree. Stream setup is initiated when a source ST agent generates a Connect message listing the flow specification and initial set of participants. Connect processing at each intermediate ST agent involves determining the set of next hop subnets required to reach all downstream receivers, installing multicast forwarding state, and reserving network level resources along each next hop subnet. If the actual resource allocation obtained along a subnet is less than the amount requested then this is noted in the Connect packet by updating the flow specification. Upon receiving a Connect indication a receiver must determine whether it wishes to join the group, and return either an Accept or a Refuse message to the stream source. In the case of an Accept the receiver may further reduce the resource request by updating the returned flow specification. During connection setup the stream source must wait for an Accept/Refuse reply from each initial receiver before beginning data transmission. ST-II treats the entire stream as a homogeneous distribution path. Whenever the source receives an Accept with a reduced flow specification it must either adapt to the lower QoS for the entire stream or reject group participation for the specific receiver by sending it a Disconnect message. Group membership dynamics are accommodated by allowing stream receivers to be added or deleted after initial stream setup. Each addition of a receiver requires an interaction with the stream source to trigger the sending of a Connect message. This interaction is not defined by the protocol specification but is instead performed out-of-band using IP. As in the initial setup the stream source must examine the flow specification in a returned Accept and either reduce its QoS or reject the new receiver if the resources allocated are less than those currently allocated for the stream. Deletion of receivers may be done asynchronously by a receiver sending a Refuse message or the source sending a Disconnect message; the Disconnect message can either list individual receivers to remove or set the Global-Disconnect flag to tear down the entire stream. Reliability and robustness are incorporated into the ST-II protocol via two sepa rate mechanisms. First, all control messages used to create and manage a stream are transmitted reliably using hop-by-hop acknowledgments with retransmission. Sec ond, a Hello protocol is used to query the status of neighboring ST agents sharing active streams. When a change in reachability between neighboring ST agents is detected automatic stream recovery may be attem pted. The only service model directly supported by ST-II is that of a homogeneous reservation over a point-to-multipoint simplex distribution tree. We call this the In dependent Streams reservation style; a separate and independent resource reservation is allocated for each distribution tree. The ST-II protocol specification defines the concept of a group of streams, which may be useful in defining more sophist icated reservation styles. Groups can be used to express relationships among individual streams or for performing operations on the group as a whole. However, the group mechanism is an experimental feature and no stream relations have been defined at this time. We do not consider the group mechanism in any of our analysis. 3.1.2 R SV P P rotocol RSVP [13] is similar to ST-II in that a data stream is modeled as a simplex dis tribution tree rooted at the source and extending to all receivers. However, the mechanisms for group sources and receivers to establish resource reservations and the reservation styles supported differ substantially from the ST-II model. Under RSVP a source application begins participation in a group by sending a Path message containing a flow specification to the destination multicast, address. The Path message serves two purposes: to distribute the flow specification to the receivers, and to establish Path state in intermediate RSVP agents to be used in propagating reservation requests toward specific sources. RSVP does not restrict a source from transmitting data even when no receiver has installed a reservation to it; however, data service guarantees are not enforced. Before establishing a reservation each receiver must first join the associated mul ticast group to begin receiving Path messages. This multicast group join operation is a function of the multicast routing protocol and is outside the scope of RSVP. Each receiver may use information from Path messages and any local knowledge (computing resources available, application requirements, cost constraints) to deter mine its QoS requirements; it is then responsible for initiating its own Reservation request message. Intermediate RSVP agents reserve network resources along the subnet leading toward the receiver then use the established Path state to propagate the Reservation request toward the group sender(s). Reservation message propa gation ends as soon as the reservation “splices” into an existing distribution tree with sufficient resources allocated to meet the requested QoS requirements. This receiver-initiated1 reservation style enables RSVP to accommodate heterogeneous receiver requirements. ’T he receiver-initiated approach was inspired by D eering’s work on m ulticast routing [29] in which the receiver is responsible for initiatin g group m em bership requests. 28 RSVP incorporates a datagram messaging protocol with periodic refreshes to maintain soft state in intermediate switches to provide reliability and robustness. Path refreshes automatically adapt to changes in the multicast distribution tree and install Path state in any new branches of the tree. Reservation refreshes maintain es tablished reservations and incorporate new receiver reservations. This refresh based mechanism allows orphaned reservations and state to be automatically timed out and recovered. RSVP models a reservation as two distinct components, a resource allocation and a packet filter. The resource allocation specifies what amount of resources is reserved while the packet filter selects which packets can use the resources. This distinction between the resource reservation and packet filter, and an ability to change the packet filter without changing the resource allocation enables RSVP to offer several different reservation styles. A reservation style captures application-level commu nications requirements: these dictate how reservation requests from individual re ceivers should be aggregated inside the network. At the moment RSVP has defined 3 reservation styles, these are Shared Tree. Independent Teees. and Dynamic Filter. other styles may be identified as new multicast applications with different needs are developed. A Shared Tree reservation indicates that a source specific reservation is not required and that any packets destined for the associated multicast group may use the reserved resources. This allows a single resource allocation to be made across all distribution paths for the group. When a source specific reservation is required a receiver may indicate whether it desires to receive a fixed set of sources or the abil ity to dynamically switch its reservation among the sources. An Independent Trees reservation cannot be changed during its lifetime without re-invoking setup and ad mission control: this allows the reservation to be shared among multiple requests for the same source.2 The Dynamic Filter reservation allows a receiver to modify its packet filter over time. This requires that sufficient resources be allocated to handle the worst case when all receivers take input from different sources. 2.\o te that w hile the ST-II Independent Stream s and the R SV P Independent Trees reservation styles can result in equivalent reservations, we use distinct nam es to distinguish the m echanistic differences. 29 3.2 S ta tic A n a ly sis The protocol descriptions in Section 3.1 noted that ST-II and RSVP are similar in that they model a data stream as a simplex point-to-multipoint distribution tree. However, the RSVP protocol incorporates heterogeneous receiver requests and mul tiple reservation styles, providing additional opportunities to improve network-wide resource utilization. In this section we look at several applications expected to be typical of an ISPN, map the service model of the two protocols to the application communication requirements and compare the network-wide resource requirements for supporting the application. 3.2.1 Supporting Self-L im iting A pp lications A number of multipoint-to-multipoint applications have application-level constraints that prohibit all data sources from transmitting simultaneously; one example is an audio conference. In an audio conference there is typically only one person speaking at a time because when more than a few speakers are simultaneously active the result is usually unintelligible. Therefore, instead of reserving sufficient resources for every potential speaker to transmit simultaneously it may be adequate to reserve only enough resources to handle a few simultaneous audio channels. RSVP is able to capture these application communication requirements exactly using the Shared Tree reservation and requesting resources for the maximum number of simultaneously active sources. ST-II requires that an Independent Stream reservation be established for each audio source. In this section we compare the total network-wide resource allocation to sup port an N-way audio conference under the two reservation protocols. We model a hierarchical network containing 60 routers interconnected via 82 links and vary the number of conference participants from 2 up to 65. Each audio source was randomly distributed among the 60 nodes3 and is modeled as a request for a 64Kb/s PCM audio stream. Figure 3.1 presents the total network-wide resources allocated under the two reservation protocols to support audio conferences of various sizes. For the 3T he random placem ent function used throughout the sim ulations selects a random order for adding participants to unique nodes, this precludes m ultiple participants at a single node until all nodes have at least one participant. 30 C/3 23 & c _ o £ < u C /3 < D & < D T 3 O £ ( L ) z ■ 4 — » O H 250000 200000 150000 100000 50000 RSVP Shared Tree (1 resv) RSVP Shared Tree (2 resv) RSVP Shared Tree (3 resv) ST-II Independent Streams • * „ Q .............Q ...... —"S -V —+—......... +..... 10 20 30 40 50 60 Number of Group Participants 70 Figure 3.1: Scaling of resource requirements in support of N-way audio conference using RSVP Shared Tree and ST-II Independent Streams reservation styles. RSVP Shared Tree reservation style we show the resource requirements when each participant requests a reservation for 1, 2, or 3 voice streams worth of bandwidth; this represents the limit on the number of simultaneous speakers. For ST-II we show the resources reserved when each source establishes an Independent Stream to all receivers, and network links are modeled as having unlimited capacity. The small slope of the RSVP plots highlights the efficiency in adding participants using the Shared Tree reservation style. Resources are reserved only along new links required to “splice” into the distribution mesh.4 As a group becomes more “dense” (the group membership covers a higher percentage of the total network nodes) the average number of new links required to splice into the distribution tree decreases, resulting in a smaller overhead per new member. In contrast, adding a participant under the ST-II model requires splicing into N — 1 existing distribution trees and setting up an independent distribution tree from the new participant to all existing members. The disparity between the Independent Streams and Shared Tree plots ''Note that the total resource allocation under Shared Tree reservation is based upon the union of the links in all distribution trees, w hile it's based upon the sum under Independent Stream s. 31 represents a resource over-allocation, which is shown to rapidly diverge as the group size increases. Allocating an independent resource reservation for each ST-II source effectively places an upper bound on the maximum group size that can be supported. For a group of size N a participant must allocate N — 1 reservations to receive from all sources. In the common scenario of a host at the network periphery with a. single access link, all N — 1 reservations must be accommodated on the same link. Making the optimistic assumption that the packet service algorithm can maintain QoS guarantees at 100% link utilization, the group size is thus limited to M a x im u m Group Size Hoilh nr ok L in k Bandw idth Single Stream Resource Request participants. Repeating the ST-II simulations presented in Figure 3.1 with link bandwidth limited to 1.5Mb/s confirmed that resource allocation requests begin to be rejected (this is termed call blocking in the telephony literature) for group sizes greater than 24.5 RSVP Shared Tree reservations do not encounter this scaling problem. The maximum resource reservation across all links is limited to the number of simultaneous sources requested, which is independent of the size of the group. One final observation is to note that the total resource requirements of the RSVP Shared Tree reservation are bounded, while ST-II resource requirements are un bounded. Under RSVP once there is a participant at each network node, resources for the complete distribution mesh have been allocated and no further resources need to be allocated to accommodate additional group members. This is evident in the plots in Figure 3.1 by the zero slope line when going from 60 to 65 group participants. ST-II always requires allocating an independent reservation from the new participant to all existing members. 5 I l _ .5 A _ /_ 6 /a I i 1 _ 2 4 [ 6 4 K b / s J ^ 32 3.2.2 Supporting H eterogen eou s Groups In a global-scale internetwork, receivers as well as the paths used to reach the re ceivers can have very different properties from one another. Network and host tech nologies are likely to span several orders of magnitude in terms of bandwidth and processing capabilities. In this environment it may not be reasonable to assume that all receivers in a group possess the same capacity for processing incoming data or desire the same QoS from the network. Applications involving wide-spread distribu tion services such as cable-TV distribution or broadcasting of an audio/video lecture may be able to accommodate additional participants by incorporating support for heterogeneous receiver capabilities. An application may employ a hierarchical cod ing scheme or provide multiple data streams utilizing different media encodings to present varying signal quality levels to the receivers. Each receiver may then deter mine its QoS requirements based on local constraints. ST-II and RSVP accommodate heterogeneity very differently. Under the ST-II service model a data source must view the entire stream as a homogeneous distri bution path. After stream setup the source must conform to the minimum resource allocation forcing all participants to suffer with the least capable or least demand ing receiver. To satisfy the most demanding receiver the source must allocate the maximum requested resources along all links. RSVP’s receiver-initiated reservation scheme propagates reservation requests from a, receiver up the sink tree toward the source “splicing” into the distribution tree. This reservation establishment process reserves the minimum resources on each link required to satisfy the QoS requirements of all downstream receivers. Thus, RSVP incorporates support for heterogeneous reservations directly in the protocol in a manner transparent to both end-points. In this section we compare the total network-wide resource requirements to sup port a heterogeneous mix of receivers listening to an audio lecture. The 60-node network introduced in Section 3.2.1 is used again and the lecture is modeled as a single data source transmitting a “high quality” 64Kb/s audio stream that also contains a sub-band 16Kb/s “low quality” audio stream. Two alternatives for sup porting this application are to send the “high quality” and “low quality” components on separate multicast trees, or to send the entire data stream over a single multi cast tree. Sending the entire data stream on a single multicast tree and forwarding D a ta S o u rc e L o w Q u a lity R e c e iv e r L o w B W R e se rv a tio n H ig h Q u a lity R e c e iv e r Figure 3.2: Link reservations installed by RSVP in support of a heterogeneous audio lecture using Independent Trees reservations (40 receivers total, 20 “low quality” ). only the components recjuired to satisfy all downstream receivers provides the most efficient support of the application. This is the model we investigate.0 Figure 3.2 shows the link reservations installed by RSVP to support 10 randomly selected receivers of the audio lecture using an Independent Trees reservation, 20 re ceivers request the full 64I\b/s stream and 20 receivers request the 16Kb/s “low quality” audio sub-channel. This diagram depicts RSV P’s ability to install a hetero geneous resource reservation across the data distribution tree. Only those branches leading to receivers requesting the "high quality” audio stream require the high band width reservation, this can resnlt in significant resource savings. For the scenario illustrated only 29 of the 46 links in the multicast distribution tree require a high bandwidth reservation, resulting in a 27.7% savings in network resource allocation when compared to a homogeneous distribution tree. '^Specification o f the m echanism s to en cod e/d ecod e this stream , and the filter to select the sub-band audio are outside the current discussion. N um ber of ST-II RSV P Low Resource Resource Quality A llocation Allocation Receivers (K b/s) (K b/s) 0 2944 2944 10 2944 2656 20 2944 2176 30 2944 1600 40 736 736 Table 3.1: Network-wide resource requirements for ST-II and RSVP protocols in support of a 40 receiver audio lecture, with various numbers of low quality receivers. Table 3.1 presents the total network-wide resource requirements for both ST-II and RSVP to support the 40 receiver audio lecture, with various numbers of low quality receivers. The ST-II stream exhibits an “all-or-nothing” effect due to the protocol’s limitation of treating the stream as a homogeneous distribution path. As long as there is at least one demanding receiver the maximum resources must be allocated along all links; this ensures the QoS for the demanding receivers is met. Under RSVP the total network-wide resources reserved reflects the minimum allo cation required along the distribution tree to satisfy all receivers QoS requests. As the number of low quality receivers increases additional branches in the distribution tree shed their high quality resource reservation resulting in a gradual decrease in total network-wide resource allocation.' 3.2.3 Supporting C hannel S election In large multiparty conferences a receiver may be unable to accommodate data streams from all active participants simultaneously but would like the ability to select dynamically a subset of the sources to receive at any time. This restriction on number of simultaneous sources may be due to bandwidth limitations, display or codec hardware, or the inability of the user to assimilate information from all sources concurrently; we term this communication style channel selection. From the 'N ote that in the worst case scenario o f a linear network R SV P allocation is identical to the ST-II case, w hile the best case scenario o f a fully connected network yields allocations that are linear in the number o f low quality receivers. user’s perspective there are two possible service models, assured channel selection and non-assured channel selection. A key characteristic of assured channel selection is that once a receiver has established its reservation it should be guaranteed that a change request will not be denied. The non-assured channel selection model does not provide such a guarantee, and a change request may be denied. The traditional method to provide assured channel selection is to allocate an independent reservation for each source, which is just the Independent Streams reservation style discussed in Section 3.1.1. The receiver can then switch between channels by selecting the desired incoming stream. The channel selecting, or filtering of incoming data, is done entirely at the receiver. RSVP introduces the Dynamic Filter reservation style, which allocates sufficient resources on each link so that the receiver can always select, without failure, any set of m sources (where m is the maximum number of simultaneous sources). Once the resource allocation is fixed a receiver may dynamically modify the associated filter, which chooses the packets that get to use the resource. Thus, the filtering is done within the network. The actual resource allocation on each link is limited to the maximum number of non-overlapping reservations; this is the sum of all downstream receiver requests limited by the number of upstream sources. A third channel selection alternative is to make a new reservation every time a new channel is selected (and then to tear down the old reservation). This provides the non-assured service because the new request may be blocked. We call this the Chosen Source reservation style since it only reserves for the currently chosen sources. Resources are reserved along the distribution tree from each source to the set of receivers that are currently tuned into that source, and the trees from different sources are independent. Because the Chosen Source reservation style reserves for only the currently selected sources it provides a useful lower bound for the resource consumption required by the assured channel selection service. Table 3.2 presents the total network-wide resource allocation required by each of the channel selection mechanisms to support the participants of the N-way con ference introduced in Section 3.2.1. While the Chosen Source reservation style does not provide assured switching among the sources, it is presented to quantify the overhead in the assured channel selection schemes, as indicated in the “Overhead 36 Chosen Source (4 Reservations) Dynamic Filter (4 Reservations) Independent Streams (N -l Reservations) Resource Resource Resource Group Allocation Allocation Overhead Allocation Overhead Size (K b /s) (K b /s) Ratio (K b /s) Ratio 5 3200 3200 1.00 3200 1.00 10 6592 8704 1.32 12032 1.83 15 9728 13184 1.36 23808 2.45 20 11840 18432 1.56 36224 3.06 25 14400 22720 1.58 52480 3.64 35 20160 32704 1.62 89152 4.42 45 26368 42048 1.59 140352 5.32 60 36416 57024 1.57 226432 6.22 Table 3.2: Network-wide resource allocation required by each of the channel selection mechanisms in support of N-way conference. Ratio” column. For the simulations conducted, the resource overhead incurred us ing the Independent Streams mechanism can be quite substantial, and it increases as the group size is increased. The resource overhead in providing assured channel selection is much smaller using the Dynamic Filter mechanism; most importantly, for the class of graphs simulated the overhead appears to be bounded as the group size is increased. The only requirement on the reservation protocol to support the Independent Streams or Chosen Source channel selection mechanism is that a fixed resource allocation can be established from each selected source to the receiver. Both the ST-II stream model and the RSVP Independent Trees reservation style provide this service. For the Dynamic Filter channel selection mechanism a distinction must be made between a resource allocation and the packet filter; this distinction is currently provided only by the RSVP Dynamic Filter reservation style. 3.3 D y n a m ic A n a ly sis In Section 3.2 we compared the resource allocations of the ST-II and RSVP pro tocols to support a fixed set of group members. In a real, large scale internetwork environment there may be frequent dynamic events that must be accommodated by the reservation protocol. These events include both network dynamics such as link and router failure/recovery and group membership dynamics (participants join and leave the multicast group). It is extremely important that the group membership dynamics be supported efficiently as membership change is expected to be a common occurrence, whereas topology change represents an exceptional event. In this section we describe the ST-II and RSVP mechanisms for supporting network dynamics and compare the protocol overhead associated with accommodating group membership dynamics. 3.3.1 N etw ork D yn am ics Section 3.1 described the mechanisms incorporated into the ST-II and RSVP pro tocols to provide reliability and robustness in the face of network dynamics. ST-II utilizes a reliable control message protocol and a Hello protocol to monitor neighbor ST agent health, while RSVP uses a datagram control message protocol in conjunc tion with a soft state refresh mechanism. The difficulty in conducting a comparison of the dynamics of the two protocols is that both rely heavily on timers (ST-II Hello interval and RSVP refresh period), which have a great effect on the protocol overhead and recovery period, and no explicit timer values are m andated by the pro tocol standards. Instead, we compare the design philosophies behind the dynamics support in the two protocols. The integration of support for network dynamics in ST-II and RSVP are sub stantially different in terms of both implementation and design philosophy. ST-II incorporates a failure detection mechanism using Hello, Status, and Notify messages, and these add considerable complexity to the protocol.8 In contrast, RSVP relies on the soft state refreshes to automatically adapt without additional protocol com plexity. RSVP could be modified to incorporate a failure detection mechanism to trigger refreshes as an optimization; however, there are more fundamental differ ences that distinguish the protocols. The key difference between the two protocols is in where recovery takes place. ST-II requires that the network be responsible for correctness by either restoring itself or reliably contacting the source; this leads to complex protocols with strange failure modes. Clark [21] notes that systems relying on distributed state are difficult to build and few truly provide protection against 8 Partridge and Pink [93] note th at much o f the functionality is overlapping. 38 failure. RSVP leaves the final responsibility for maintaining reservations with the ends; this is consistent with the current Internet philosophy of “fate-sharing” among the end-points.9 Note that even in steady state (no network or group dynamics) there is an over head associated with both protocols. Under ST-II this overhead is a result of each ST agent periodically exchanging one Hello message with each active neighbor. Requir ing the agent to track peers separately from streams may pose a slight complication in data structure organization; however, it results in a protocol that scales indepen dent of the number of active streams. Protocol overhead in RSVP results from the periodic Path and Reservation refreshes. This would seem to imply that RSVP over head scales directly with the number of participants; however, RSVP incorporates a protocol overhead reduction mechanism called “merging” to reduce this overhead. The merging process insures that only a single reservation message is propagated over a link per refresh period. W ith a Shared Tree reservation there is only a single reservation on each link for the entire group, for an Independent Trees reservation there is one reservation for each source forwarding along a link, while a Dynamic Filter requires a separate reservation per receiver (bounded by the total number of upstream sources). Thus, RSVP protocol overhead scales with the number of reservations. 3.3.2 G roup M em bership D ynam ics Large multicast groups such as global distribution of a conference or lecture are likely to encounter frequent membership dynamic “events.” These events are a result of participants tuning into and leaving the conference. In a correctly functioning internet, group membership changes are much more common than network dynamic events. It is important that the reservation protocol be able to accommodate these membership dynamics efficiently. Protocol efficiency can be evaluated in terms of messaging overhead and latency in adapting to changes. In this section we compare the protocol overhead for ST-II and RSVP to adapt to group membership changes. 9Clark [21] characterizes the fate-sharing m odel as gathering the critical sta te inform ation at the end-point o f the net, in the entity which is utilizing the service o f the network. It is then acceptable to lose the sta te inform ation associated w ith the entity if, and only if, the entity itself has failed at the sam e tim e. 39 Dynamic addition of receivers under ST-II requires the generation of Connect and Accept messages between source and receiver. The end-to-end messaging of ST-II results in an overhead on each link proportional to the number of downstream receivers. This results in links closer to the source becoming “hot spots,” in that they incur a higher overhead in terms of bandwidth and protocol processing overhead. Also, the explicit source interaction required for every group membership dynamics could result in a processing bottleneck at the source. The receiver initiated reservations in RSVP result in a very different join overhead model. Assuming homogeneous receivers the join overhead is reduced to one protocol message on each link in each direction. This represents a single Path message sent by the source to build the reverse path state, and a single reservation request sent by each receiver. The key to R SV P’s reduced join overhead is the merging function; as soon as the reservation request splices into an existing distribution branch the request can be merged (discarded). The situation becomes only slightly more complicated when heterogeneous receivers are introduced. In this case the merging function must ensure the request splices into the distribution tree and there are sufficient resources allocated. This may result in multiple reservation messages being propagated over a link if a more demanding request is received after a less demanding reservation has already been installed. The use of receiver initiated reservations and reservation merging in RSVP result in two distinct advantages over the end-to-end protocol of ST-II. First, the implo sion of messages at the sender causing “hot spots” is eliminated; second, the total network-wide protocol overhead is reduced. Figure 3.3 shows the total network-wide protocol overhead for ST-II and RSVP for various numbers of homogeneous receivers independently joining the audio lecture first described in Section 3.2.2. This graph shows that the RSVP merging function is indeed highly effective in reducing pro tocol overhead. In fact, RSVP becomes more efficient as the group becomes more “dense” due to the average number of hops to splice into an existing distribution branch decreases.10 10N ote that the current assum ption o f hom ogeneous receivers result in a best case scenario o f one protocol m essage on each link in each direction. T h e w orst case is encountered in a heterogeneous environm ent when the receivers join in order from least dem anding to m ost dem anding, resulting in an overhead proportional to the num ber o f dow nstream receivers on each link. 40 80000 ST-II - RSVP 70000 60000 50000 s 40000 > o 30000 o o o o 20000 0. 10000 0 5 10 15 20 25 30 35 40 45 50 Number of Receivers Figure 3.3: Network-wide protocol overhead for ST-II and RSVP supporting homo geneous receivers independently joining a multicast group. In addition to protocol overhead, another important measure of group dynamics support is the latency in reacting to group changes. RSVP latency can be “tuned” by adjustment of refresh timers making direct comparison of latency times difficult; however, we can make some general observations regarding the two protocols. Un der ST-II the reservation setup and teardown times for a target are nominally one round trip time between source and receiver and one end-to-end delay respectively. Latencies in RSVP are much less precise. Adding a new receiver may involve an initial delay in waiting for a Path refresh if the receiver is on a new branch in the multicast distribution tree; reservation setup time is also variable from as little as one hop up to an end-to-end delay depending upon whether an existing reservation can be “spliced.” When a receiver leaves, explicit reservation teardown can release the resources immediately. 41 3.4 C h a p ter S u m m ary We have described how the ST-II and RSVP protocols provide resource reservation establishment in support of an Integrated Services Packet Network. Both protocols utilize multicast data distribution to improve network efficiency for multipoint com munication; however, we argue that a richer service model is required for the ISPN environment. Our simulations show that RSV P’s support for heterogeneous receiver requests and multiple reservation styles can be exploited to obtain significant im provements in network-wide resource allocation for several common applications. If these application classes make up a significant fraction of the resource demands in an ISPN, then incorporation of RSVP could result in a substantial reduction in net work resource requirements and improve scaling in terms of the number and size of groups that can be accommodated. Both ST-II and RSVP use timer based mechanisms to provide robustness in adapting to network dynamics; however, the design philosophies are quite different. ST-II requires that the network be responsible for correctness, leading to increased protocol complexity. RSVP uses a soft state mechanism, leaving end-systems re sponsible for refreshing state. We also showed that the receiver-initiated reserva tion and merging in RSVP reduces the load on links closer to the source, reduces source-receiver interactions, and reduces the network-wide protocol overhead when compared to ST-II. 12 C h a p ter 4 A sy m p to tic R e so u r c e C o n su m p tio n in M u ltica st R eserv a tio n S ty le s T he addition of resource reservations in the ISPN changes the concept of resource consum ption in the netw ork. W ith o u t reservations, the usage of resources in these netw orks is tied directly to sending packets; i.e., if you haven’t sent any packets you haven’t consumed any resources since you haven’t denied or affected anybody else’s service. W ith reservations, adm ission control will deny access if there are not sufficient unreserved resources available; reservations, even if unused, can therefore prevent other flows from reserving resources. Thus, reservations them selves can be seen as consuming resources som ew hat independently1 from the actual usage of those reservations. In this chapter, we will focus on the reservation of resources rather than their actual use. A lthough audio and video applications often do not have a fixed quality of service requirem ent (i.e., they can operate over a broad range of data rates providing varying degrees of perceived quality), we assum e th a t a reservation m echanism is required to ensure m inim al signal quality levels. A nother class of applications whose requirem ents are not handled efficiently by the In te rn et’s point-to-point best-effort service are those th at require the sam e data to be sent to several receivers. This occurs in teleconferences and rem ote lectures, where voice and video from one individual go to m any other participants. Using the l T he degree o f independence depends on the nature o f real-tim e service provided. Service which provides vvorst-case delay bounds m ust base all adm ission control decisions on the reservation param eters w ithout regard for the actual usage. Service which provides lower quality assurances, like the predictive service in [‘ 22, 102], can base adm ission control decisions at least in part on actual usage. traditional point-to-point service, a separate packet is sent to each receiver (we call this sim ultaneous unicasts); m ultiple copies of packets are sent over the overlapping portions of the routes to individual receivers. M ulticast routing [27, 29, 113] solves this problem by having the netw ork, whenever the routes diverge, send a single copy along each path. As we quantify later, this leads to trem endous efficiency gains. In fact, m ulticast (as em bodied in the M bone [16]) has been crucial in enabling the widespread distribution of video and voice in broadcasting Internet Engineering Task Force m eetings. B roadcasting these m eetings, which at tim es have several hundred listeners, would sim ply have been im possible w ithout m ulticast given the current lim ited bandw idth on m any Internet links. Adding m ulticast and real-tim e service will transform th e Internet from < point- to-point best effort netw ork into one th a t also offers point-to-m ultipoint real-tim e service. This will greatly expand the range of applications whose needs can be efficiently m et by the In tern et. We should note, however, th a t there are some real tim e applications th a t are best described as m ultipoint-to-m ultipoint; an exam ple of this is an N-way teleconference where every participant needs to see and hear every other participant. Such applications can be dealt w ith as a set of indepen dent point-to-m ultipoint applications; when the paths from two different sources to the sam e receiver overlap, resources are reserved for both sources independently. This approach is typically sufficient when the traffic from these sources, and the receiver’s desire to see th a t traffic is relatively independent.2 However, as recog nized in [55, 85, 123], when the sources are not independent this approach leads to inefficiencies. The RSV P resource reservation protocol [123] introduces the con cept of reservation styles. Reservation styles represent different ways of aggregating the resource requirem ents for each source on a single link. Table 4.1 sum m arizes the reservation styles investigated in this study. T he detailed definition of these reservation styles, and the m echanism s th at im plem ent them , have been described elsewhere [13, 123]. O ur purpose in this chapter is to evaluate the extent to which these reservation styles increase efficiency; in particular, we focus on asym ptotic resource usage as it (the num ber of participants in the m ultipoint application) gets large. Earlier, in 2By this we m ean that my desire to see source A is independent o f m y desire to see source B. 41 R eservation Style D escription Independent Trees A separate and independent reservation is allocated for each source distribution tree. Per-link reservation is based on the num ber of upstream senders (N up_src). Shared Tree A shared reservation is allocated on each link in the dis trib u tio n mesh for use by any source. Per-link reservation is based on the num ber of upstream senders, lim ited by the num ber of sim ultaneous sources th a t will transm it at any one tim e (M IN [ N up^ rc, N sim_src\). Chosen Source A separate and independent reservation is allocated along the distribution tree from each source to only the set of receivers th a t are currently tuned into th a t source. Per-link reservation is based on the num ber of upstream senders th a t have been selected by at least one down stream receiver (Nup_sei_STC ). D ynam ic Filter A set of shared resources is allocated on each link to ac com m odate the m axim al dow nstream resource dem and. Each reservation has a receiver-controlled “filter” allow ing dynam ic selection among sources. Per-link reserva tion is based on the num ber of upstream senders, lim ited by the num ber of independent reservations required to al low all dow nstream receivers to make independent source Selections ( M I N^NupsTci (Ar do\vn-rcvr * N sim-chan )] )• Table 4.1: Sum m ary of R SV P reservation styles. Section 3.2, we introduced these issues in a m ore informal m anner; th e present work is an a tte m p t to quantify these savings in a m ore system atic and rigorous fashion. We find th a t the RSVP reservation styles achieve very significant savings for large n. T hus, if m ultipoint-to-m ultipoint applications represent a sizable portion of the future netw ork load then it will be im portant to include these reservation styles in the basic Internet service model. We look at two cases w here these reservation styles are useful; “self-lim iting” applications and channel selection. T he details of these application models were presented in Section 3.2.1 and Section 3.2.3. Each of these application models violate the "independence" assum ption, and thus the reservations for traffic from separate senders should not necessarily be treated independently. We investigate the asym ptotic resource savings achieved by the R SV P reservation styles in three sim ple netw ork topologies (see Figure 4.1): linear, m -tree, and star. These topologies are not m eant, of course, to be realistic. R ather, by restricting ourselves to sim ple and tractab le topologies, our intent is to provide a more rigorous and system atic analysis of resource consum ption in large m ultipoint-to-m ultipoint applications. Most of our results are analytical, although we do rely on sim ulations for some quantities th a t have so far defied direct calculation. T he rem ainder of this ch ap ter is organized as follows: we first, in Section 4.1, de fine the basic resource consum ption model and describe the three sim ple topologies. We then, in Sections 4.2 and 4.3, describe the two reservation styles and analyze, through both analytical m odeling and sim ulations, their asym ptotic resource con sum ption. We conclude with a sum m ary of our results in Section 4.4. 4.1 N etw o rk M o d e l We consider m ultipoint-to-m ultipoint applications running on a set of n network hosts. Each host is both a sender of d ata and a receiver of data. Each host gener ates an equivalent traffic stream , which consumes (or at least requires the reservation of) some given am ount of bandw idth. T he q u an tity of interest is the total reserved bandw idth needed to support a given size application. We will set the am ount of bandw idth reserved to be the unit of bandw idth, so th a t every independent reserva tion consumes one unit of bandw idth3. All reservations are unidirectional in nature. We consider the three netw ork topologies depicted in Figure 4.1: linear, m -tree, and star. W hile none of these topologies are particularly good m odels for a real netw ork, they do represent a wide spectrum of possibilities. M any of our results are relatively independent of topology, which suggests th a t perhaps our results are 3.\o te that we are using a rather prim itive m odel o f reservations, using only bandwidth to describe the reservation. In practice, the flow specification [94, 113] will likely be som ew hat more com plex. Legend: O = Host ^ = Router “ ~ — = Link m t Linear m-trce (m =2) Star Figure 4.1: Network topologies considered in resource consum ption analysis. relevant to m ore general netw orks. Each link is bi-directional, with separate reser vations for bandw idth in each direction. We consider the capacity of each link to be unlim ited. Each source sends its d a ta to all other hosts. R outing is done via m ulticast. Since these are acyclic topologies, there is no am biguity in the routes. T here is a m ulticast distribution tree from each source to all other hosts. Similarly, there is a reverse tree going from each receiver to all other hosts; this describes the paths taken by d a ta arriving at th a t host. In our topologies, the distribution tree and the reverse tree are always identical. In fact, for all hosts they both are the entire network in all of our topologies (although links m ay be traversed in different directions in different trees, each link is traversed exactly once in each tree). A distribution mesh is the union of the distribution trees. For our networks, the distribution mesh is always the entire netw ork with every link traversed in both directions. For each netw ork topology, we consider a netw ork w ith n end hosts and let the network grow as the num ber of hosts does. T here are several quantities th at will be relevant to our later analysis. For a given size netw ork n, we can consider: Total Links L T he total num ber of links in the topology. D iam eter D T he m axim um host-host distance, in num bers of hops. Average P ath A The average host-host distance, in num bers of hops. This does not count a host connecting to itself. 47 T o p o lo g y L D A Linear 77— 1 7 7 — 1 n+1 3 m-Tree — — 7 ( 7 7 - 1 ) 7 11 — 1 ' 1 2 logm n 2[(m — 1 )n logm 71— 71+ 1] ( 7 7 — 1 ) ( 7 7 7 — 1 ) Star n 2 2 Table 4.2: Topological properties sum m ary. Let us now briefly discuss the three topologies. T he linear topology has n — 1 links, with each connecting two hosts. Clearly D = L — n — 1. A slightly more nontrivial calculation'1 reveals th a t A — In the m -tree topology, the hosts are at the leaves of the tree and the tree has a constant branching ratio of m. Here, n = rn,l„ where d is the depth of the tree. T he longest path is one th a t traverses to and from the root of the tree, so D = 2d = 2 log 7 7 .. We also have L = - ^ ( n - 1), and A = "~"+l1 ■ o m H I — 1 V n (71 — 1 ) ( 111 — 1) T he star configuration has a central hub, and there is a link connecting each host to this hub. Here, D = A — 2, and L = n. Notice th a t the star topology is merely the lim iting case of the m -tree topology w ith d = 1 and in — n. These results are sum m arized in Table 4.2. For later com parisons, we now com pu te the resource usages of m ulticast and sim ultaneous unicasts. T he quantity of interest is the total num ber link traversals, which counts each tim e a packet traverses a link as a separate use of the link. To do this com putation (and later com putations), we define two quantities for a given direction along a given link:5 N U p_src is the num ber of upstream sources th a t include the link in their m ulticast distribution tree. Ndown-rcvr is the num ber of dow nstream hosts th a t receive d ata along this link. 4 Here, as elsewhere in (lie chapter, we will spare the reader the derivation o f these essentially com binatoric form ulae. A lso, all o f these form ulae are only valid for 7 1 > 1 ami for values of 77 that represent a com plete topology. T his is not an issue for the linear or star topologies, but is relevant for the m -tree, where 77 = rnd are the only valid values for 7 7 . 5Even though links are bidirectional, when referring to the reservations along a link we typically are referring to a single direction. For the topologies we consider, these two num bers m ust always sum to n, N up_src + Ndown^rcvr = n , since every link is on every distribution tree. Furtherm ore, considering the reverse direction of the link m erely reverses these two num bers. W hen using sim ultaneous unicasts, the total num ber of packets traversing a particular link is given by N up_src * Ndown_T C V r. W hen using m ulticasts, the total num ber of packets traversing a particular link is given by N up_S T C , because duplication for different receivers is elim inated. Sending a packet from each source to each destination w ithout using m ulticast involves n(n — 1)/1 link traversals; the d a ta from a single source travels over n — 1 paths, of average length A, and there are n such sources. Using m ulticast involves merely nL link traversals, since no link is traversed m ore th an once by any packet, and all links are traversed in exactly one direction in each m ulticast tree. Thus, the ratio of (? ? . — l)/l to L is an estim ate of resource savings due to m ulticast. For the linear netw ork, these savings are 0(n), for m -trees the savings are 0 (lo g m »), and for a star the savings are 0 (1 ). We should note th a t these savings are calculated for link traversals of data. In the rest of the chapter we are interested in savings in term s of reserved resources. Reser vation styles do not affect the actual num ber of link traversals, only the resources reserved. 4.2 S elf-L im itin g A p p lica tio n s We now consider self-lim iting applications. This application style was introduced earlier in Section 3.2.1. To recap, these are m ultipoint-to-m ultipoint applications w ith application-level constraints th a t inhibit d a ta sources from tran sm ittin g sim ul taneously. An audioconference is one exam ple of this, where the social inhibitions tend to discourage sim ultaneous speaking.6 A nother rath er different exam ple is satellite tracking. Here there are a num ber of large antennae, and when the satellite is w ithin their range the data, is downloaded and then sent to the other sites. If the ranges of the antennae do not overlap so the satellite is only within range of a G N ote th at a vidoconference is not self-lim iting, since video is independent o f what other par ticipants are doing. Topology N um ber of Res Independent Trees ervations Shared Tree R atio Linear n(n — 1) 2 ( n - l ) n 2 Tree nm(n-l) m — 1 2m(n— 1 ) m— 1 1 1 2 Star n2 2 n ii 2 Table 4.3: Sum m ary of resource allocations for self-lim iting applications with N simsrc — 1 ■ single antenna at any one tim e, then the traffic is self-limiting because two sources are never active sim ultaneously. M ore generally, we can describe a self-lim iting ap plication by the m axim al num ber of sources th a t will transm it at any one tim e; we will denote this quantity by N ai,n_arc. T he traditional approach is to m ake separate and independent reservations for each distribution tree. We will call this the Independent Trees approach. On every link, the num ber of units of bandw idth reserved is given by N up_STC . T he total bandw idth reserved over the whole netw ork is given by nL = n(n — 1)^4, since there is an independently reserved path from every sender to its n — 1 receivers, and there are n senders. However, as was noted in [123], the application requirem ents are entirely m et if, on every link, there are m erely sufficient reserved resources to accom m odate the m axim um num ber of upstream sources th a t will sim ultaneously transm it. Thus, on each individual link we need only reserve M I N [ N up_arc, Naim_arc\, which is in contrast to the N up_arc units reserved by the Independent Trees approach. RSVP has defined a reservation style which we call Shared Tree, in which only this necessary reservation of M I N[Nup_src, /V sim_src] is made; these resources are shared between the upstream sources in the sense th at traffic from any of them can use this reserved bandw idth. For sim plicity, we focus on the case where N si,n_arc = 1, so in the Shared Tree reservation style each link has either 0 or 1 units of bandw idth reserved. T he differ ence in reservation styles can then be concisely captured by noting th a t the Indepen dent Trees reservations are based upon the sum. of all links in all distribution trees, while the Shared Tree reservations are based upon the union of the links across the distribution mesh. In all of our topologies the mesh consists of all the links traversed 50 in both directions, whereas each distribution tree consists of all links traversed in only one direction. T hus, the ratio of bandw idth consum ed betw een these two ap proaches is always j in our th ree topologies. We have sum m arized these results in Table 4.3. In fact, w henever the distribution mesh is acyclic the ratio of Independent Trees to Shared Tree resource usage is exactly Consider the distribution tree from source A, and assum e th a t it does not touch some link (in either direction) in the distribution m esh. Since this link is in the distribution m esh, it m ust lie 011 the path between two sources; assum e th a t this link is on the path from B to C. Then, the path from A to B to C - to A contains a cycle, since it can only traverse the missing link once (going from B to C). T hus, if the distribution mesh is acyclic then every distribution tree touches every link once and only once. It then directly follows th at the distribution mesh touches every link in both directions. Therefore, whenever the distribution mesh is acyclic th e ratio of resource usage is exactly Note th at in cyclic networks this result need not hold. For instance, in a fully connected network, the Independent Trees and the Shared Tree resource dem ands are exactly the same. For purposes of com parison, it is interesting to note th a t m u lticast’s advantage over sim ultaneous unicasts ranged from O(n) in linear netw orks, to O(login n) in m -tree netw orks, to 0 (1 ) in sta r networks. In contrast, the Shared Tree reservation style has an advantage of | in all networks w ith acyclic distribution meshes. Observe also th a t the results for the Shared Tree and Independent Trees reservation styles are consistent with the intuition th a t the resource recpiirements of Independent Trees scale as O(nL) w hereas those of Shared Tree scale as 0(L). 4.3 C h a n n el S ele ctio n We now consider ‘'channel selection” applications [25]. This application style was introduced earlier in Section 3.2.3. To recap, these are applications in which the traffic from each sender is independent, but the receiver only wishes to receive d ata from a lim ited num ber of senders at any one tim e. T he eponym ous exam ple is th at of television, where one w ants access to many channels but only wants to receive one at a tim e. Similarly, large m ultiparty videoconferences are som etim es an exam ple of this, in th a t a receiver may be unable to accom m odate d a ta stream s from all active participants sim ultaneously, but desires the ability to dynam ically select a subset of the sources to receive at any tim e. This restriction on the num ber of sim ultaneous sources m ay be due to bandw idth lim itations, display or codec hardw are, or the inability of the user to assim ilate inform ation from all sources concurrently. In general, we can characterize a channel selection application by the m axim um num ber of channels Nsimian it wishes to receive at any one tim e. From the user’s perspective there are two alternative service models: assured channel selection and non-assured channel selection. In assured channel selection, the user is guaranteed th at the resources will be available to view the selected chan nel. Thus, assured channel selection involves pre-reserving resources for all of the channels. Assured channel selection is the service th a t is appropriate for the ex am ples cited above. In non-assured channel selection, no such guarantee is made, and the request can be denied by admission control. We consider this case only because it provides a convenient lower bound to the resources required for assured channel selection. T here is a tradeoff betw een the e x tra assurance of the assured service m odel, and its presum ably higher cost due to ex tra resource consum ption; one of the goals in this section is to exam ine quantitatively this ex tra resource con sum ption. We now describe the reservations required to support these assured and non-assured services. T he m ost direct way to support this non-assured service is to m ake a new reserva tion every tim e a new channel is selected (and then to tear down the old reservation). We will call this the Chosen Source reservation style, since it only reserves for the currently chosen sources. Resources are reserved along the distribution subtree from each source to the set of receivers th a t are currently tuned into th a t source. The trees from different sources are independent. T hus, the reserved am ount on a link is given by N up_sei_src which is the num ber of senders upstream th at have been selected by at least one dow nstream receiver. The Chosen Source reservation style, because it only reserves for the currently selected sources, provides a lower bound for the resource consum ption required by assured service. We can provide assured service using two different reservation styles. T he traditional way to provide assured channel selection is to reserve independent trees for each source, which is ju st the Independent Trees reservation style discussed in Section 4.2. This provides sufficient resources for all sources to sim ultaneously arrive at the receiver. T he receiver can then switch betw een channels by selecting the desired incoming stream . T he channel selecting, or filtering of incom ing data, is done entirely at the receiver (m uch like the signals for all TV channels arrive at the cable set-top box, and the tu n er selects one). RSVP [123] introduced the idea, inspired by a com m ent from Jon Crowcroft [25], th at this selecting or filtering process can occur w ithin the netw ork rather than ju st at the receiver. Each reservation on a link is accom panied by a filter th a t determ ines which packets get to use the reserved resources. One of the novel aspects of RSVP is th a t even while the reservation is fixed this filter can change dynam ically in response to signals from the receivers. RSV P offers a Dynamic Filter reservation style which reserves enough bandw idth on each link so th a t the receiver can always select, w ithout failure, any set of Nsim_ciian sources. RSV P provides a m echanism whereby receivers inform their upstream routers which sources they wish to receive, and the filters are then set to only allow packets from those sources to pass. T he resource requirem ents can be expressed as follows; on every link the am ount reserved is given by M I N [ N up_src, (N lloiun_rcvr * N.iim_ckan)}, recalling th at Ndoum-rcvr > s the num ber of dow nstream hosts th a t receive d ata (from any source) along the link (i.e., the num ber of receivers for which this link is in the reverse tree), and N up_src * K the num ber of upstream sources th a t include the link in their m ulticast distribution tree. This form ula m erely expresses the insight th a t one need not reserve more chan nels than the num ber of upstream sources, nor m ore than the m axim al num ber of dow nstream requests. As one can see directly from the expression for per-link reservations, on every link th e resources required for D ynam ic F ilter reservations is bounded above by the Independent Trees reservation and below by the Chosen Source reservation. We now proceed to analyze the asym ptotic resource consum ptions of these var ious reservation styles. For sim plicity, we choose N sim_cilan = 1, so every receiver receives only one channel at a tim e. W ith this choice, the reservation in the two directions on a link are identical, since reversing directions merely reverses the m ean ings of N up_src and A (^0 (i;7 ? _ 7 .cl,r . N um ber of R eservations Topology Independent Trees D ynam ic F ilter R atio Linear n(n — 1) n2 2 2(n-l) n Tree mn(n— 1 ) 2n logm n m(n-l) m — 1 2(m — l)logm n Star 7 ? 2 2 n n 2 Table 4.4: Sum m ary of resource allocations for assured channel selection with N s in i-c h a n — 1 ■ 4.3.1 A ssured C hannel S election A ltern atives We now com pare the two assured channel switching reservation styles. The Inde pendent Trees reservation case was already considered in Section 4.2; recall th a t the total resource consum ption is given by nL. The D ynam ic Filter reservation style requires a reservation of M I N [ N up_src, A^O U )n _rc„r] on every link; the presence of the M I N function makes this style som ew hat harder to characterize and com pute. In the linear case, the reservation needed on a link in the Fth position is M IN [i, n — t]. For 77. odd, this sum s to and for n even it sum s to —■ (in Table 4.4 we only show the result for n even). For the m -tree topology, the expression M I N [ N up_src, Ndow,i.rcvr] reduces to the num ber of hosts below the link on the tree (assum ing the root of the tree is “up"). T here are m ‘ links at depth i in the tree, and there are m'l~l nodes below the links at depth i. Thus, the resource consum ption at every level is ju st 2???r f (the factor of 2 is because each link has two directions). Since there are d levels, the total resource consum ption is ju st 2dmd. In term s of n, this becomes 277 logm ?r. T he star topology result can be calculated by merely setting ? n = 7 ? in the m -tree result, yielding a resource consum ption of 2n. These results are sum m arized in Table 4.4. They are consistent with the intuition th a t the worst case of Chosen Source, and hence Dynam ic Filter, scales as 0 ( n D ), in contrast to Independent Trees scaling as O(nL).' 'W h ile we expect that the worst case o f Chosen Source will scale as 0 ( n D ) in m ore general topologies, we doubt th at D ynam ic Filter will continue to be equal to the worst case o f Chosen Source in m ore general topologies. 54 4.3.2 D ynam ic F ilter vs. N on-assured S election O verhead As m entioned earlier, our interest in the Chosen Source reservation style is prim arily because it represents the m inim al resources needed to support the currently selected sources. Thus, we can use this reservation style to quantify th e overhead incurred in providing the ex tra assurance in the assured selection service (as opposed to the non-assured service provided by the Chosen Source style). However, the total re source requirem ents for Chosen Source depend not only on netw ork topology and participant distribution8 but also on the set of sources selected by each receiver. W hen describing the resource consum ption of Chosen Source, we therefore need to characterize the set of source selections. Consequently, we define three classes of Chosen Source behavior; worst, case (C S worst) occurs when all receivers correlate their source selections to m axim ize the to tal resource consum ption; average case (C Savg) is the average result when each receiver perform s an independent and ran dom source selection; and best case (CSbest) is achieved when receivers correlate their source selections to m inim ize the total resource consum ption. For each of these Chosen Source behaviors we com pute, either through analysis or sim ulation, the total resource consum ption and com pare these asym ptotic resource requirem ents with those of the Dynam ic F ilter reservation style. 4 .3 .2.1 Chosen Source W orst Case (C S W 0TSt) T he worst case for Chosen Source results when each receiver selects a distinct source (resulting in no overlap in distribution trees) such th a t the set of selections maximizes the total point-to-point distance. For the linear topology, C S worst is obtained when each receiver selects th e host ^ distance away (assum ing for convenience th a t n is even). T he total resource requirem ents for this case can be easily calculated to be Y - For m -tree, C S worst is obtained when each receiver selects a host th a t requires traversal of the root node, this results in 2D link reservations per source giving a total requirem ent of 2nD = 2 n lo g m ?i. For the star topology, C S worst is obtained whenever each receiver selects a distinct source, resulting in 2n reservations. '’T hese com pletely determ ine the resource requirem ents for the Independent Trees and D ynam ic Filter reservation styles. Topology N c s’ ^ W O TSt um ber of R eservation C Savg s C Sf,est Linear T l2 2 0 [ n 2) sim ulation n Tree 2n logm n 0 ( n lo g m n) sim ulation m(n+l) — 2 m~ 1 Star 2 n 0[n) sim ulation n + 2 Topology R a C - S•!!-;; / f Sworst tio L Sl,zsi j C Sworsl Linear 0.53 sim ulation 2 n Tree 0.77 sim ulation 1)— 2 2 r< (rn — 1)Iogm n Star 0.S2 sim ulation n-f 2 2 T i Table 4.5: Sum m ary of resource allocations for non-assured channel selection with N s im - c h a n — 1 • These results are reproduced in Table 4.5. Surprisingly, for all the topologies studied the ratio of C S worst to Dynam ic Filter is always exactly 1. T hat is. in these topologies providing assured channel selection requires absolutely no additional resources when com pared to th e worst case of the non-assured channel selection. We do not yet know how fully general this result is. We do know th a t it does not hold for the fully connected netw ork (where Dynam ic Filter requires n(n — 1) reservations and C S worst requires only n). 4.3.2.2 C hosen Source A verage C ase (CSaiy) We now consider the average case perform ance of the Chosen Source reservation style when all receivers m ake an independent and random source selection. We have been unable to solve this case exactly, and so instead we use sim ulation to com pute C S i J n v g • O ur experim ental m ethodology was to sim ulate each of th e th ree network topolo gies for various values of n. For each value of n we perform ed random source select ion for each receiver; selecting a Chosen Source from am ong th e n — 1 other participants 56 1 0 0.8 G S & c 1 06 < O J g °-4 < u os 0.2 0 100 200 300 400 500 600 700 800 900 1000 Number o f Hosts (n) Figure 4.2: R atio of chosen source average and worst case for selected topologies. with uniform probability. T hen we calculated the exact num ber of link reservations recjuired by the Chosen Source reservation style. We repeated this process m ultiple tim es and used the sam ple m ean to predict C SaVg■ Even though the total num ber of perm utations for source-receiver selection grows as (n — l ) n we found th a t repeating the random source selection process ju st 500 tim es for each n resulted in an estim ate of CSavg w ith less th an 1% relative error at a 95% confidence level. R ather than displaying C.S'tt„i:;’s absolute perform ance, we show how C S worst com pares to C S avg.9 In Figure 4.2 we plot the ratio of the sim ulated C S avg re source requirem ents against those of C Sworst for the linear, m -tree (w ith m = 2 and in = 4), and star topologies. Note th a t the ratio appears to asym ptotically approach a nonzero constant for all topologies investigated (the constant depends on the topology, but in each topology the ratio appears to asym ptote to a constant). This implies th a t Dynam ic F ilter overallocates only a fixed percentage of resources Linear Topology ------ M-tree Topology (m=2) ------ M-trce Topology (m=4) ........ Star Topology °N ote that C S wor, t is equivalent to D ynam ic Filter thus this also represents the ratio in per form ance o f D ynam ic Filter assured channel selection to average case Chosen Source 1 1 0 1 1 -assured channel selection. when com pared to Chosen Source average case. Thus, not only does the assured ser vice of Dynam ic F ilter not require any overallocation when com pared to the worst case of the non-assured service of Chosen Source, but it also only requires a fixed percentage of overallocation when com pared to th e average case of the non-assured service of Chosen Source. Again, we do not know how general these results are, but they do hold in each of our three topologies. 4.3.2.3 Chosen Source B est Case {CSbeat) Chosen Source best case is when the source selections m inim ize th e to tal resources required. This occurs when all receivers but one select the sam e source (a receiver cannot select itself as its source) and the exceptional receiver selects a nearest source. This results in a single m ulticast distribution tree from the source to all but one of the receivers plus the path from the exceptional receiver to its nearest neighbor. T hus the total resources required are L + 1 for the linear topology and L + 2 for the m -tree and star topologies. For each topology, CS'best scales as O(n). In contrast, Dynam ic F ilter scales as 0 ( n 2) in the linear topology and as 0(?? logm n) in the tree topology (and as O(n) in the star topology). Therefore, only when com paring D ynam ic F ilter to the best case of Chosen Source in the linear and tree topologies do we find an asym ptotic scaling advantage for Chosen Source. T he extent of this advantage scales as 0{D), since Dynam ic Filter scales as O(nD) and Chosen Source best case scales as O(n). 4 .4 C h ap ter S u m m ary In this chapter we studied the asym ptotic resource consum ption of various RSVP reservation styles in three sim ple topologies. To our knowledge, this is the first analytic comparison of the relative m erits of these approaches. For self-lim iting applications, the Shared Tree reservation style achieves savings of | over the tra d i tional Independent Trees reservation style in any topology with an acyclic distribu tion mesh. For channel selection applications, the D ynam ic F ilter reservation style achieves substantial savings over the Independent Trees reservation style in the m- tree and star topologies. More surprisingly, the Dynam ic Filter reservation style uses 58 exactly the sam e resources as the worst case of the Chosen Source reservation style, and appears to be only a constant factor worse th an the average case of the Chosen Source reservation style. These results suggest th a t, at least for large m ultipoint applications, the RSV P reservation styles of Shared Tree and Dynam ic F ilter offer substantial savings in resource consum ption over the traditional Independent Trees reservation style. 59 C h a p ter 5 R eso u rce R eserv a tio n D y n a m ics and T h ra sh in g Admission control, or equivalently resource reservation, introduces a new form of resource contention for the shared netw ork resources. Work in other dom ains in corporating resource allocation (e.g., database system s) has uncovered phenom ena - usually called thrashing - having great detrim ental effects on overall system per form ance and throughput. In this chapter we study thrashing in the context of an ISPN architecture such as th a t proposed in [1‘ 2]. This work represents a first study of w hat we believe to be an im portant phe nom ena. We observe several necessary conditions to induce thrashing, including inter-reservation resource dependencies, and allowable reservation setup and tear- down delays. W ith reservations, adm ission control will deny access if there are not sufficient unreserved resources available. O nce reservation blocking occurs, the end- user or application m ay exhibit several different, styles of behavior in regard to if and how reservation requests are requeued. We look at the effects different user behavior can have on system perform ance. We then look at more complex m ulticast reser vations and m ultipoint-to-m ultipoint applications, which have not been studied in previous network reservation setup investigations [1, 109]. Finally, we propose end- user and application behavioral characteristics for reservation request retry backoff which result in significant im provem ents in system stability. T he rem ainder of this chapter is organized as follows: we first, in Section 5.1 define the model of thrashing we consider and introduce our intuition as to its cause. Next we introduce the netw ork m odel, topologies studied, and our sim ulation m ethodology in Section 5.2. We then, in Sections 5.3 thru 5.5 evaluate reserva tion system throughput using sim ulations. We begin by establishing the underlying 60 principles of the thrashing phenom ena and dem onstrate its existence using a sim ple uni-directional point-to-point reservation m odel (Section 5.3). In Section 5.4 we look at the effects of m ore complex reservations and application styles, and finally at m ethods to im prove system stability (Section 5.5). O ur current results represent a sim ple progression through several initial thrashing scenarios; it is far from a com plete understanding of th e entire netw ork resource reservation thrashing space. In Section 5.6 we sum m arize our findings from the current scenarios studied. 5.1 T h ra sh in g Thrashing has been observed in a num ber of different dom ains where there is con tention for a set of shared resources [1, 79, 112]. Tay [112] showed th a t, in databases with a fixed transaction length and total resources, scaling up the resource dem and results in an initial increase in throughput to a point and then a steady decrease due to contention. We believe th a t this model is m ost sim ilar to the allocation of hop-by-hop reservations in th e ISPN architecture. O ur rationale as to why thrashing m ay be exhibited by the reservation m echanism of the ISPN architecture is based upon the fact th a t resources are wasted when they are reserved b ut not used, we ci.ll these resources provisional. Provisional resources can be accum ulated in several ways: such as during the setup process due to long end-to-end propagation and adm ission control delays, or due to long teardow n delays after a request has been blocked or a receiver has com pleted service. A ccum ulation of any provisional resources increases the system usage level while not contributing to throughput, and in turn increases the probability of blocking for other independent reservation requests. In essence, blocking begets m ore blocking. In addition to the accum ulation of provisional resources due to initial blocking, we also believe th a t circularity in th e resource dependencies of independent reservations can have a detrim ental effect. We m ean, looking at the set of network resources required to establish a set of n independent reservations {/q, r2, ..., /■ „} then /q may require resources held by 7 ' 2 , r 2 may require resources held by 7 ’3 , ..., rn may require resources held by 7q. This is the fam iliar circular wait condition described in the deadlock detection literatu re [95], and can in fact lead to deadlock am ong the set of reservations involved in the cycle. 61 O ur thesis is th a t the accum ulation of significant levels of provisional resources can occur, with the effect of significant throughput degradation as the resource dem and is increased. We follow a strategy of first finding the basic phenom ena in sim ple if perhaps unrealistic settings and then studying its dependence on various factors. In later sections we show th a t thrashing can arise in m ore realistic, but more com plicated, situations. All results presented rely on reservation teardow n delay as the prim ary source of delays introduced into the system , we have perform ed some analysis of the effects of propagation and adm ission control delays but much work is left to future research. 5.2 N etw o rk M o d e l In this work we once again assum e the Integrated Services Packet Network architec ture of [12]. T he details of the com ponents of this architectural model were presented earlier in Section 1.1 and Section 2.1. We consider a reservation to be a uni-directional point-to-m ultipoint stream using a source-rooted m ulticast distribution tree. Reservation requests are receiver initi ated as in the RSVP reservation protocol [13, 123], the request is m erged with the m ulticast distribution tree at th e first branch where sufficient resources are already allocated for the requested stream . We arbitrarily set the am ount of bandw idth requested for each reservation to be the unit bandw idth, th a t is each independent reservation consumes one unit of bandw idth.1 We also arbitrarily select a 60 second holding tim e for all successful reservation requests. The effect of varying the reser vation holding tim e is to p ertu rb the total netw ork resource dem and for a specific request arrival rate, but this does not affect w hether thrashing occurs. T he underlying building block in our investigations is the individual resource reservation, however we also consider more com plex scenarios th a t include m ulticast d a ta distribution and m ultipoint-to-m ultipoint applications. W hen discussing these complex scenarios we often find it useful to refer to the grouping of all reservations associated w ith the application. We use the term session throughout the chapter 'N o te that we are using a rather prim itive m odel o f reservations, using only bandw idth to describe the reservation. In practice, the flow specification [94, 113] will likely be som ew hat more com plex. when referring to a group of related source and receivers and the set of related individual reservations. In addition, a session typically exhibits a specific behavior in regard to the coordination of establishm ent of all the com ponent reservations. We define the details of these behavior in Section 5.4 where we consider the session models in detail. All sim ulations reported in this chapter were perform ed using a discrete event sim ulation package im plem enting a receiver-initiated soft sta te reservation protocol sim ilar to th at specified in [13]. T he term soft-state was first used by Clark in [21] and, in our context, refers to reservation sta te m aintained at each network switch which is periodically refreshed by end applications: in the absence of refresh messages, such as in case of route changes or end host crashes, the reservation state tim es out and removes itself. T he soft-state approach can add both adaptivity and robustness to reservation protocols, however at the added overhead of periodic refreshing messages. Therefore to keep the overhead low the refresh period should not be too short, and the tim eout period also needs to be set accordingly. In our sim ulation a ‘‘teardow n” delay is introduced to model the reservation removal delay. Explicit reservation teardow n requests result in zero teardow n delay; absence of explicit requests leads to a teardow n delay greater than zero seconds. We consider two distinct classes of netw ork topology, cyclic and acyclic. The cyclic network is composed of four switching nodes each connected to two neighbors, forming a sim ple square topology. Each of the interconnection links in the cyclic network is provisioned with sufficient capacity to accom m odate 23 sim ultaneous reservations.2 We assum e m inim al link propagation and node processing delays of 1 millisecond each, thus our current work focuses exclusively on th e effects of teardown delays. Source and receivers were placed at each of the four sw itching nodes with all reservations being betw een a source and receiver at opposite diagonals of the network, thus all reservations are of length two hops. For the acyclic network we consider a binary tree topology of depth four, again all links have a capacity of 23 sim ultaneous reservations. Source and receivers are placed only at th e 16 leaf nodes of the tree with reservations betw een source and receiver at each pair of leaf nodes, this results in a mix of 2-, 4-, 6-, and 8-hop paths. 2V V e have also performed prelim inary sim ulations on networks w ith higher degrees o f m ultiplex ing and we found sim ilar thrashing behavior. 63 W hen studying a single netw ork and homogeneous reservations the resource de m and can typically be quantified by session arrival rate alone. In this study we look at m ultiple network topologies, heterogeneous num bers of sessions and path lengths, m ultipoint-to-m ultipoint sessions and m ulticast distribution trees. Assum ing all reservation requests are for th e unit bandw idth, we can calculate the Norm alized Loading (NL) of the network as the product of the session arrival rate and the num ber of link reservations required to successfully allocate the session distribution mesh. Calculating the NL for a specific sim ulation scenario results in an estim ate of the average num ber of new link reservation requests injected into the network per unit of tim e. We use the N orm alized Loading m easure throughput this chapter when presenting results th at are dependent on netw ork loading level. The norm alized loading m etric tells us th a t for a larger session size fewer arrivals are required to m aintain a fixed loading level, for this reason we scale the length of our sim ulation runs for scenarios with larger session sizes. All of our sim ulation runs are for 10,000, 20,000, or 40,000 sim ulated seconds depending on the session size. We found these sim ulation lengths to be sufficient to ensure consistent sim ulation results. W hen reporting any results we discard the d a ta from the first half of the sim ulation run and calculate all statistics from the d a ta following the w arm up period. 5.3 P o in t-to -p o in t R eserv a tio n s The basic reservation model of the ISPN architecture is th a t of a uni-directional point-to-m ultipoint stream . However, we begin our investigation by looking at the sim plest case, th a t of uni-directional point-to-point reservations. W hen a reserva tion request is blocked the end-user or application does not a tte m p t to requeue the request; th at is, there are no reservation request retry attem pted. O ur intuition tells us th a t in a system with acyclic ordering of resources there should be a guaranteed level of successes; as resource dem and is increased we expect system throughput to be strictly non-decreasing, and if throughput asym ptotes we expect the value to be non-zero. We believe th a t in order to obtain dependencies among independent reservations th a t lead to thrashing requires a circularity in the resource dependencies of the individual reservations. Note th a t under the sim ple 64 4000 Teardown=0s Teardown=30s --a—■ -+—-Tearc(Qwn-60s 3500 t / 5 t / 3 < D 3 C T O O S c .2 £ o c/i o O S 3000 2500 2000 C / i C /1 CD CD o 3 C / 0 1500 1000 0 2 3 4 5 6 Network Load (link reservation requests / s) 7 8 Figure 5.1: Num ber of successful reservation requests on the square topology for uni-directional point-to-point reservations. uni-directional point-to-point reservation model this circularity in inter-reservation resource dependencies can occur only if the network itself contains cycles.3 Figure 5.1 and Figure 5.2 present the num ber of successful reservation requests for sim ulations of the uni-directional point-to-point reservations on the square and binary tree topologies. T he larger size and capacity of the binary tree topology required much higher loading to induce blocking, however the im portant distinction is in the different shapes of the plots in the two figures. We see th a t as predicted the cyclic resource dependencies introduced by the cyclic topology can result in a decline in network throughput as the load is increased, while the acyclic topology always shows increasing throughput. T he results in Figure 5.1 do establish th a t the thrashing phenom ena can occur w ithin the simple uni-directional point-to-point reservation scenario investigated, however the degradation was only observed for extrem ely long delays and network 3We note that in fact in the real world m ost topologies are likely to contain cycles to avoid the single point of failure problem s of a purely acyclic network, however the cyclic resource dependency m ay still be rare although the possibility is alw ays there. 14000 Teardown=0s - Teardown^Kls-* Tcardo>yn<K)£ - - Teardpwn=6()s 12000 10000 8000 6000 a t/j C /5 < D O O D 4000 2000 0 5 10 15 20 25 30 35 40 Network Load (link reservation requests / s) Figure 5.2: Num ber of successful reservation requests on the binary tree topology for uni-directional point-to-point reservations. overload. For system s w ithin reasonable operational ranges it seems quite stable. One situation where these long delays may actually be encountered m ight be w ithin a soft state protocol th a t requires large tim er values to control protocol overhead. A lesson to be learned here is th a t a soft sta te m echanism should not rely on tim ers alone, explicit messages should be incorporated to effect sta te changes. Soft state tim ers should only be relied upon to m aintain consistency in exceptional cases such as when messages are lost or system s crash. We note th a t in fact the RSVP soft state reservation protocol does employ this model with explicit teardow n messages. 5.3.1 P oin t-to-P oin t R eservations w ith B locked R eservation R etry In the sim ple reservation scenario explored in Section 5.3 we noted th at whenever a reservation request was blocked no further action was taken by the end-user or application. An extension to this scenario is to recognize that, a common m ode of operation might be for the end-user or application to retry its reservation setup 66 4000 T%ardowu=0s— « — _ _TeaFdowa=Q.5s. .-.t.v.z Teardown=0.95s a — T eardow n=ls * Teardown=60s 3500 C /5 C/3 1 ) 3 cr o 0 6 c .2 c d £ O J t/J <L> O h 3000 2500 2000 1500 a V J 00 ( U O O 3 00 1000 500 1.5 0.5 2 2.5 3.5 3 4 Network Load (link reservation requests / s) Figure 5.3: Num ber of successful reservation requests on the square topology for uni directional point-to-point reservations with a one second reservation request retry interval. request after a short delay in th e hope th a t netw ork conditions have changed in the interim . In this section we investigate the effect of adding reservation request retries to the uni-directional point-to-point reservation model. We arbitrarily selected a reservation request retry interval of one second.'1 We assum e th at each reservation request retry a ttem p ts to build upon the provi sional resource allocation obtained during earlier requests if the reservation is still in place; this leads to two distinct regions of operation. W henever the teardow n delay is greater th an the reservation request retry interval, each retry can build upon the previously established partial path reservation. If the teardow n delay is less than the retry interval, then the provisional resources have already tim ed out and each new request m ust once again contend for resources along every link in the end-to-end path. 4T he effect o f the selection o f the reservation request retry interval is shown to partition the system perform ance into two d istinct operational regions, however it does not affect our results in term s of w hether the network can be m ade to exhibit thrashing. 67 Network Total Blocked Successful Load R equests R equests R equests 0.7 1725 0 1725 0.8 43929 42293 1636 0.9 577949 577949 0 Table 5.1: Sum m ary of reservation success and failures on the square topology for uni-directional point-to-point reservations with a one second reservation request retry interval and 60 second teardow n delay. Figure 5.3 presents the num ber of successful reservation requests for sim ulations on the square topology for uni-directional point-to-point reservations with a one second reservation request retry interval. It m ight seem logical th a t the perform ance in the region where retries a tte m p t to build upon earlier provisional allocations (i.e., the teardown delay is greater th an the retry interval) would be superior to th a t of the region where each request m ust contend for new resources at every link, however this is obviously not true. T he problem with attem p tin g to build upon the earlier partial path reservation is th a t once a sufficient blocking level is reached every retry continues to block and the provisional allocation is held forever. Thus we see th at system perform ance for all teardow n delays greater th an the retry interval is exactly identical, the entire system deadlocks and throughput im m ediately drops to zero. N ote th at because the allowable teardow n delay is directly dependent on the retry interval, system instability can be induced for arbitrarily small teardow n delays by aggressive applications. For the operational region where earlier provisional allocations are released the deadlock condition can be avoided. We observe for teardow n delays only slightly less than the retry interval thrashing can still occur, while sm aller delays result in strictly non-decreasing throughput. This transition from thrashing to non-decreasing throughput is dependent on a num ber of factors including network topology, capacity and delays, and resource dem ands. We have not com pletely modeled this transition phenom ena at the present tim e. Table 5.1 and Table 5.2 sum m arize the total num bers of reservation request success and failure on the square and binary tree topologies for uni-directional point- to-point reservations with a one second reservation request retry interval and 60 second teardow n delay. We see th a t the total reservation requests behavior is sim ilar 68 N etw ork Total Blocked Successful Load Requests R equests R equests 4 1528 0 1528 5 6057 4146 1911 6 163105 161007 2098 Table 5.2: Sum m ary of reservation success and failures on the binary tree topology for uni-directional point-to-point reservations with a one second reservation request retry interval and 60 second teardow n delay. 2400 Tcardown^Os Tcardqwri=10s ~i~ Teprdown^Os ,7eardbwn=60s * ■ 2200 2000 1800 1600 1400 1200 a 1000 800 600 400 200 3 2 4 5 6 7 Network Load (link reservation requests / s) Figure 5.4: N um ber of successful reservation requests on the binary tree topology for uni-directional point-to-point reservations with a one second reservation request retry interval. 69 under both network topologies. As the netw ork load is increased and blocking begins to occur the total num ber of reservation requests injected into the system begins to increase exponentially. Figure 5.4 presents the num ber of successful reservation requests for sim ulations on the binary tree topology for uni-directional point-to-point reservations with a one second reservation request retry interval. N ote th a t even with the much greater reservation request rate introduced by the retry policy, th e system throughput is still non-decreasing. These binary tree results differ significantly from the throughput perform ance observed for the cyclic topology (see Figure 5.3) where thrashing was observed as the netw ork load increased. This further re-enforces our initial intuition th a t the inter-reservation resource dependency circularity is a necessary condition to induce thrashing. 5.4 M u ltip o in t S essio n M o d els In Section 5.3 we looked at scenarios where each source and receiver represented an independent uni-directional point-to-point reservation. In fact, the ISPN reser vation model directly supports more complex point-to-m ultipoint reservations asso ciated with m ulticast d a ta distribution. A dditionally we believe th a t m ultipoint- to-m ultipoint applications m ay become quite common and these may require coor dination in establishm ent of m ultiple reservations. We call the set of source and receivers composing a m ultipoint-to-m ultipoint application a session and assum e the existence of a session controller elem ent which applies specific policies to the coordinated resource reservation requests.5 In the rem ainder of this section we look at the effects of the session model on the reservation system perform ance. We assum e an N-way conferencing session model w ith a session controller th a t requires all reservation requests to succeed before session establishm ent is com pleted. We believe this model is m ost appropriate in capturing the effects of small video teleconferencing sessions.6 We will quantify the 5N ote that, the m ultipoint application and session controller issues are actually independent. One could theoretically im agine a p oin t-to-m u ltipoin t application w ith a session controller that required all end-points to succeed, however such applications appear to be less com m on. 6We recognize that there are other classes o f applications, particularly ones w ith very large m em bership, that do not have this strict m odel o f success. A nalysis o f these session m odels is another area for future investigation. 70 4000 Teardown=0s — T eardow n=0^5s^ ’ + — Teardowj*f0.5s Q - Tcardoyi/fi=0.95.s * Trffirdown=ls ~4~ - Jfeardow n=60s -*■— 3500 3000 2500 2000 ( A (A < U c / 3 1500 a 1000 500 0 2 3 6 4 5 7 8 9 10 li Network Load (link reservation requests / s) Figure 5.5: N um ber of successful session reservation requests on the binary tree topology for 2-way sessions with a one second reservation request retry interval and retry-all-receivers session retry policy. effects of the more complex resource dependencies inherent in the session model and the effects of increased session sizes. 5.4.1 B i-d irection al Session M odel T he sim plest extension to the reservation model is to pair two uni-directional point- to-point reservations to form a bi-directional session. T h a t is, if there is a reservation requested from A to B then there m ust also be a reservation requested from B to A and both requests m ust successfully com plete for the session to be established. This session model is likely to be quite common (e.g., telephone conversations). Note the m ajor im plication inherent in this sim ple session model extension, now for any two independent sessions traversing a link there is an inter-session resource dependency circularity. O ur conjecture is th at thrashing is now possible even on the acyclic network topology. Figure 5.5 presents the num ber of successful session reservation requests on the binary tree topology for 2-way sessions with a one second reservation request retry 71 interval. We assume that when a session is blocked and provisional resources are to be released (i.e., the teardown delay is less than the retry interval) that the session controller causes both end-points of the session to simultaneously release their provisional allocations. System performance is now similar to that observed earlier in Figure 5.3 for the uni-directional point-to-point reservations on the cyclic topology. The session model has introduced resource dependency circularities, which in turn induces thrashing and system deadlock for teardown delays greater than the retry interval. This is quite significant in the fact that with the introduction of the session model it is now possible to induce thrashing on any network topology, not just those with physical cycles. 5.4.2 Scaling Session Size The simplest session model, as presented in Section 5.4.1, is the combination of two uni-directional point-to-point reservations to form a bi-directional session, however we expect larger N-way sessions to be quite common too. A consequence of these larger session models is that each source must now establish a multicast distribution tree to each of the other (N — 1) session members, and N of these reservations must be obtained before session establishment is completed. Even with the increased effi ciency in multicast distribution over multiple point-to-point connections, the effect of the larger session sizes is to greatly increase the total number of link reservations required for session establishment. For example, for all combinations of 2-wa,y ses sions on the binary tree topology we find that the average number of link reservations required for session establishment is 13, for 4-wa,y sessions this increases to 54, and for 8-way sessions 172 link reservations are required. The net effect of larger session sizes is that each session arrival: 1) represents a larger independent resource demand and 2) involves many more admission control decisions, failure of any one can result in blocking of the entire session. For these reasons we believe that as the session size is increased the probability of session blocking will also increase for a fixed network loading level. Table 5.3 shows the blocking probability for various N-way session sizes on the binary tree topology when there are no blocked reservation retry attem pts and immediate blocked reservation 72 N etw ork Load 2-way Sessions Blocking P robability 4-way Sessions Blocking P robability 8-way Sessions Blocking P robability 1 0.00 0.00 0.01 5 0.00 0.07 0.17 10 0.13 0.32 0.42 15 0.27 0.51 0.59 20 0.38 0.60 0.65 Table 5.3: Blocking probability for various N-way session sizes on the binary tree topology for no blocked reservation retry and immediate blocked reservation tear down. teardown. As expected, the blocking probability is significantly increased for fixed load level as the session size is increased. In the remainder of this section we look at the effects on system performance of this higher blocking probability for larger sessions. 5.4.2.1 R etry-all-receivers Session R etry Policy As noted in Section 5.4.1 during simulation of the 2-way sessions, one possible session retry policy is for the session controller to force all receivers to simultaneously release their provisional resource allocations after a session block. We call this session retry policy retry-all-receivers since the retry policy applies to all session receivers independent of whether their individual requests succeeded. Figure 5.6 presents the number of successful session reservation requests on the binary tree topology for 4-way sessions with a one second reservation request retry interval and retry-all-receivers session retry policy. We see that the shape of the plots are similar to those observed for the 2-way sessions in Figure 5.5, that is both 0.95 and 1 second teardown induce thrashing while the smaller teardown delays do not. However, note the effect of the higher blocking probabilities in the 4-way sessions, the onset point for thrashing has been significantly reduced. For a one second teardown delay we see that the maximum network loading has been reduced from a load level of 8 down to a load level of 4. An interesting observation was discovered after histogramming the tree depths at which blocking was occurring. As the session size is increased the total number of flows traversing the network backbone links significantly increases. This would 73 1 400 Teardown=0s . T eardow ^O ^vj^*^ Teardown^Ss o-- Teardown^u.95s ■ ■ • * Teajdown=ls - JC0iaraown=6O s - 1200 c /3 C /3 < a 3 cr o Q £ c o £ D c /1 i) oc 1000 800 600 a C / l C / 1 0 ) O O 3 00 400 200 2 3 4 5 Network Load (link reservation requests / s) 6 7 Figure 5.6: Number of successful session reservation requests on the binary tree topology for 4-way sessions with a one second reservation request retry interval and retry-all-receivers session retry policy. Tree D epth 2-way Sessions Blocking P robability (NL = 8) 4-way Sessions Blocking P robability (NL = 4) 8-way Sessions Blocking P robability (NL = 4) 1 0.16 0.00 0.00 2 0.45 0.20 0.07 3 0.33 0.61 0.72 4 0.06 0.19 0.21 Table 5.4: Session blocking probability at thrashing onset point for various tree depths and N-way session sizes with a one second reservation request retry interval, one second blocked reservation teardown delay and retry-all-receivers session retry policy. 74 lead one to believe that in a homogeneous network, such as the one we simulated, the backbone becomes more of a bottleneck as the session size is increased. Further, one might assume that scaling the capacity of links closer to the backbone would be beneficial in maintaining uniform utilization levels throughout the network. In fact, we found that the opposite is true. Table 5.4 shows the percentage of session reservation requests blocked at the various tree depth levels (level 1 is at the root of the tree while larger values of tree depth are towards the leaves) for different N-way session sizes at the network loading level that induced thrashing. We see that as the session size is increased each receiver requests a larger aggregate reservation on its local access links, pushing the bottleneck and associated blocking out towards the leaves, and thereby reducing blocking on the backbone links. 5.4.2.2 R etry-blocked-receivers Session R etry Policy An alternative to the retry-all-receivers session retry policy investigated in Sec tion 5.4.2.1 is to incorporate a session controller that coordinates reservation retry for only those session receivers that had their previous reservation request blocked, all successful receivers maintain their resource allocations while waiting for the other receivers to complete session establishment. We call this session retry policy retry- blocked-receivers. The retry-blocked-receivers session retry policy might appear ben eficial to developers of multipoint applications because, as we noted earlier, as the session size increases it becomes much more difficult to establish a complete session as a single request. Retry-blocked-receivers eliminates the need for all session re ceivers to re-contend for resources for every retry request, while incrementally adding receivers to complete session establishment. Figure 5.7 presents the number of successful session reservation requests on the binary tree topology for 2-way sessions with a one second reservation request retry interval and retry-blocked-receivers session retry policy. Although the retry-blocked- receivers session retry policy intuitively seemed better suited to establishing multi point sessions than the retry-all-receivers policy, we see this is not true. Similar to the efTect observed in the uni-directional point-to-point retry model (see Section 5.3.1) where teardown delays greater than the retry interval led to system deadlock, we see that, the effect of successful receivers holding onto their resource allocations is to 3500 Teardown=0s -® — :ardown=0.25s - eardown=0.5s - e j-- ■keardown=ls * 3000 C Yi c n u 3 cr o O S c o 2500 2000 1500 a C/3 C/3 o o 3 1000 500 0 0 4 6 0 2 8 10 12 14 Network Load (link reservation requests / s) Figure 5.7: Number of successful session reservation requests on the binary tree topology for 2-way sessions with a one second reservation request retry interval and retry-blocked-receivers session retry policy. significantly increase the level of provisional resources, resulting in further system degradation. In fact, now we see that even if there are no delays in blocked reser vation teardown for those receivers performing retries, the system deadlocks. The exact same effects were observed in the 4- and 8-way session simulations with the incorporation of the retry-blocked-receivers session retry policy. The performance results presented in Figure 5.7 show that the retry-blocked- receivers session retry policy is very detrimental to the network. However, the retry- blocked-receivers session retry policy is very beneficial to the individual sessions that adopt it. The incremental addition of session receivers always results in fewer retries, and therefore lower session establishment delays, than retry-all-receivers. 5.5 R eserv a tio n R e tr y B ack off In the previous sections we have seen that reservation thrashing can occur due to the backlog of users retrying their reservation requests. Previous work on arbitrating access to shared resources (e.g., Ethernet CSMA/CD, TC P congestion control) has 76 3800 — T eardaw n=l s..: rt: + - - Teardown=60s o 3600 c / > C / i CD 3 cr o > O i 3400 3200 c _o 3000 c a t CD C / i ID O S 2800 2600 C /S C/S CD O O 3 on 2400 2200 2000 0 2 3 4 5 6 7 Network Load (link reservation requests / s) Figure 5.8: Number of successful reservation requests on the square topology with uni-directional point-to-point reservations with exponential retry backoff. shown an exponential retry backofT policy to be highly effective in improving system stability. In this section we investigate the effects on reservation system stability when an exponential retry backoff policy is added to the reservation retry scenarios explored in earlier sections. Our model for incorporating exponential reservation retry backofT is to assume a set of cooperating end-applications each of which doubles its retry interval timer each time a blocked reservation request indication is received. 5.5.1 U nidirection al P oin t-to-P oin t R eservations w ith E xponential R etry Backoff Figure 5.8 presents the number of successful reservation requests for simulations on the square topology with uni-directional point-to-point reservations and the expo nential retry backoff. Comparing the results to those found earlier for the similar scenario with fixed retry interval (see Figure 5.3) we see that the retry interval back off has a dramatic effect on improving system stability. With the retry backoff in 4000 Teardown=0§. Teardowi Tcardowff j j p f s -•+ — i=60s v j? . -' 3500 3000 2500 c d t o > C/5 1 ) a: c o 2000 1500 a c r t C/3 1 > U o 3 CO 1000 500 0 2 4 6 10 12 8 Network Load (link reservation requests / s) Figure 5.9: Number of successful session reservation requests on the binary tree topology for 2-way sessions with exponential retry backoff and retry-all-receivers session retry policy. place, once a reservation request begins to block the retry interval is quickly pushed to a level greater than the teardown delay, thus eliminating the deadlock problem. Interestingly, measurements of the average setup delays showed both scenarios to be nearly equivalent. The total number of reservation request retries was significantly lower with the backoff, however the exponential increase in retry interval resulted in virtually identical elapsed times. 5.5.2 Session R eservations w ith E xp on en tial R etry Backoff Figure 5.9 presents the number of successful session reservation requests for simula tions of the binary tree topology with 2-way sessions, the exponential retry backoff and retry-all-receivers. Once again when compared to the earlier results for the similar scenario with fixed retry interval (see Figure 5.5) we see a substantial im provement in system stability. Throughput is now strictly non-decreasing even for very long teardown delays. In fact, for reasonable teardown delays (e.g., less than 1 78 4000 Teardown=0s Teardown=ls 3500 3 cr < u 06 c o 3000 2500 c d e i> < n < u 06 c o 2000 'In in u G O 1500 a 1000 500 0 2 6 4 12 8 10 Network Load (link reservation requests / s) Figure 5.10: Number of successful session reservation requests on the binary tree topology for 2-way sessions with exponential retry backoff and retry-blocked- receivers session retry policy. second) we see that throughput is almost identical to the immediate teardown case in Figure 5.5. We also saw similar results for the larger session sizes. Figure 5.10 presents the number of successful session reservation requests for simulations of the binary tree topology with 2-way sessions, the exponential retry backoff and retry-blocked-receivers. We see that in this case the exponential retry interval backoff is insufficient to stabilize the system. The problem here is that the backoff policy is only applied to those receivers performing retries. Once a receiver has successfully obtained its reservation request it continues to consume resources while any remaining session members are still attem pting to obtain their requested reservations. This accumulation of provisional resources is sufficient to interfere with the requests of additional session arrivals. 79 5.6 C h a p ter S u m m a ry We wish to stress once again that the current results presented are just an initial study. However, we think that our simple progression of reservation scenarios un covers a number or interesting dynamics introduced into the ISPN architecture by resource reservations and multipoint applications. Our earliest results establish that even the simplest reservation scenarios with no blocked reservation retries can exhibit throughput degradation. Because of the excessively long delays required this result most likely pertains only to those de velopers of soft state protocols, who should heed the advice to incorporate explicit, state change messages. We also found that the larger resource demand and number of admission control decisions as N-way conference session size is increased results in significant decreases in maximum throughput. The primary findings of our study is related to the effects the end-user and application behavior can exert on the network stability. It is reasonable that an application be allowed to retry its reservation request after a failure, however we found that the retry interval selected has a direct influence on the allowable delays before entering the thrashing region. Aggressive applications can induce thrashing in a system with arbitrarily fast reservation setup and teardown. The retry-blocked- receivers session retry policy provides another example of a behavior that is advan tageous to an application but quite detrimental to the network. We showed that retry-blocked-receivers also induced thrashing for arbitrarily small delays. We assumed an environment with cooperative applications employing an expo nential reservation request retry backoff and found significant improvement . Expo nential backoff resulted in improved system stability, maximized throughput, and setup delays consistent with the non-backoff scenario. We did find that the expo nential backoff was not sufficient to overcome the negative effects of every possible user behavior, the retry-blocked-receivers still showed significant degradation. SO C h ap ter 6 S u m m ary an d F u tu re W ork In this final chapter, we first briefly summarize the contributions of this research, and then provide a practical perspective of our work. We conclude witli a discussion of several directions for future pursuit. 6.1 S u m m ary o f C on trib u tion s This dissertation studied issues in resource reservation within the framework of an Integrated Services Packet Network. Our work focuses on two major topics: resource reservation protocol scaling, and reservation dynamics. Our analysis employed both analytical and simulation modeling techniques. 6.1.1 R eservation P rotocol Scaling Our interest in studying reservation protocol scaling was sparked by our belief that scaling is the critical constraint faced today in support of services across a global in ternetwork. Our thesis was that new and novel reservation protocol mechanisms are required to efficiently support the ISPN architecture. We established that the result ing network utilization and efficiency depends to a great extent on the reservation protocol’s service model and dynamic response. We conducted our analysis within the context of two proposed reservation proto cols, ST-II and RSVP. ST-II borrows many concepts from traditional circuit switched network setup protocols; including source initiation and reliable messaging. RSVP 81 proposes several novel mechanisms; including receiver initiation, heterogeneous reser vations, and soft-state. 6.1.1.1 Service M odels We developed analytical and simulation models to evaluate the static resource re quirements to support specific application classes. This work established that the reservation protocol service model can have a significant effect on network utilization. We found that the mechanisms incorporated into RSVP to support self-limiting ap plications. heterogeneous groups, and channel selection result in significant resource savings in the scenarios investigated. For self-limiting applications our initial study on a number of simulated topolo gies found that the Independent Streams reservation style can result in limits on maximum group size and unbounded resource requirements. The RSVP Shared Tree reservations bound maximum resource requirements. Using analytical modeling techniques we proved that the Shared Tree reservation results in significant resource savings compared to the Independent Streams across a broad class of topologies. In fact, we found for any network with acyclic distribution mesh that Shared Tree resource savings grow linearly in the group size. We contrasted this by noting that adoption of multicast data delivery results in at best 0(n) savings over simultaneous unicast, only in the limited case of linear networks. Initial simulations of channel selection applications using Dynamics Filter reser vations showed significant resource savings over Independent Streams reservations, while it appeared to incur a bounded overhead compared to the less demanding non assured channel selection. Analytical modeling of the channel selection alternatives showed that Dynamic Filter consistently results in resource savings when compared to Independent Streams for assured channel selection service. However, the magni tude of the savings was highly dependent on the network topology. For the tree and star network topologies the Dynamic Filter savings level increased with group size, while the linear network produced a constant savings level. We confirmed our earlier intuition that Dynamic Filter incurs only a fixed overhead when compared to the average case behavior of non-assured channel selection for the classes of topologies investigated. Surprisingly, we also found that the worst-case behavior of non-assured channel selection is exactly the same as Dynamic Filter. Only when comparing Dy namic Filter to the best case of Chosen Source in the linear and tree topologies did we find an asymptotic scaling advantage to non-assured channel selection service. Finally, support for heterogeneous reservations in RSVP can better accommodate the range of bandwidths and processing capabilities expected in a global internet. Heterogeneous reservations provide for shedding high reservation levels along paths leading to less demanding receivers. This can result in significant resource savings compared to a protocol that requires a homogeneous distribution tree. 6.1.1.2 D ynam ic Response We were unable to perform a direct comparison between the ST-II and RSVP proto col overhead in adapting to network dynamics. This was due to both protocols being highly dependent on timer settings. We did note several fundamental differences in design philosophy. The RSVP soft state mechanism reduces protocol complexity and is consistent with the current Internet philosophy of fate-sharing. Building reliability and recovery into the network results in much more complex protocols. Group membership is likely to be more dynamic than network topology changes in a correctly functioning internet. Our simulations showed the receiver-initiated reservations and merging of RSVP to be more efficient in terms of protocol overhead when adapting to group membership changes. In addition, end-to-end messaging “hot spots” near the source and possible source processing bottlenecks are elimi nated. 6.1.2 R eservation T hrashing We established our intuition on reservation thrashing, based on provisional reserva tion accumulation and circular resource dependencies among independent reserva tions. Our simulations of uni-directional point-to-point reservations demonstrated the thrashing phenomena and confirmed our intuition on the effects of resource de pendency circularity. We also established that end-user and application behavior can have a significant effect on system performance and thrashing onset. We believe our study of the multipoint session model represents the first reser vation dynamics analysis of the ISPN model. We established that the session model introduces thrashing into a broader range of topologies, and scaling the N-way ses sion model results in reduced throughput. We showed the retry-blocked-receivers retry policy, which is advantageous to the application, can be disastrous to network stability. We showed an exponential retry backofF policy can be beneficial in im proving network stability, however greedy users can still negatively impact system performance. 6.2 S ign ifican ce o f W ork A great deal of research has been going on in developing the components of the ISPN architecture. Much of the work has focused on the design and analysis of high bandwidth physical medium, packet service, and admission control and policing al gorithms. There has been surprisingly little documentation in the research literature on design of resource reservation protocols to support the ISPN, and almost no anal ysis to quantify the performance of the alternatives. This dissertation represents a first attem pt at developing tools and techniques for the study of resource reservation protocols for the ISPN. We hope that this dissertation encourages further research and discussion related to design and analysis of novel reservation mechanisms, and the analysis of resource reservation dynamics. There are several practical applications to our results presented. As noted several times, many of the mechanisms incorporated into the RSVP protocol were novel and not well understood. Our protocol scaling analysis in effect “validates” many of these mechanisms. We have quantified the performance of a number of RSVP components, including Shared Tree and Dynamic Filter reservation styles, heterogeneous group support and reservation merging, and shown that they do in fact result in significant improvements in network efficiency and scaling. Our analysis can also be used when considering the inclusion of controversial components within the protocol, such as Dynamic Filter support. The results from the channel selection analysis could then be used to quantify the expected Dynamic Filter resource savings, while contrasting against the increase in protocol and implementation complexity required. Our work on resource reservation dynamics has several practical applications. The application to soft state protocol design is quite clear; soft state protocols should always incorporate explicit messages to effect state changes. Application developers can benefit from an understanding of the effects their retry behavior can have on network performance. Longer retry intervals, an exponential retry backoff, and release of provisional resources are all beneficial. Application developers must understand that greed can result in reduced performance for themselves and all other users of the network. More fundamentally, we believe our initial work on resource reservation dynamics opens up a new area for study. Our establishment of thrashing, even for the simple scenarios, demonstrates that this could be a serious problem in the ISPN. With the addition of the session models the dependencies become even more complex. This demonstrates that a complete understanding of the dynamics, and development of mechanisms to “protect” the network may be necessary to ensure efficient operation. Independent of the actual results presented, our study has also resulted in the development of tools and techniques to study scaling and dynamics of resource reser vation protocols. The application scenarios and analytical scaling models could be used by others as the basis for analysis of other reservation protocols. The simu lation tools developed represent a considerable level of time and effort, and can be reused in other studies of resource reservation consumption. In addition, the simu lation tools could be incrementally modified to add new reservation styles or other mechanisms to be analyzed. 6.3 F u tu re W ork This dissertation opens up a number of avenues for future work under both the reser vation protocol scaling analysis and reservation dynamics topics. In the remainder of this section we outline several areas for further study. Under each topic we iden tify a number of simulation and analysis scenarios meant to broaden the scope of our current analysis, as well as more fundamental issues requiring more in depth research and development of additional analysis techniques. 6.3.1 R eservation P ro to co l Scaling There are a number of reservation protocol scaling and RSVP related topics open for further investigation. 6.3.1.1 D etailed Q uestions The following simulation and analysis scenarios identified would broaden the scope of the current analysis. Protocol Scaling • Current simulations and analysis of receiver initiated reservations and reservation merging, heterogeneous groups, and group dynamics were con ducted on a very limited number of topologies. Additional simulations across a broader range of topologies, group membership distributions, and application styles are required to more completely capture their perfor mance behavior. Soft S tate Scaling • We noted that a local fault detection and refresh trigger mechanism could be incorporated as an optimization. How would this affect protocol com plexity and recovery latency? • Timer settings control adaptation latency and have a large effect on proto col overhead. Is it possible to dynamically adapt timers to measured net work performance to reduce protocol overhead? Alternatively, network wide signaling overhead could be set to a constant utilization level by scaling refresh timers dependent on link bandwidths and the number of active reservations.1 How would this affect rate of adaptation to dynamic events and maintenance of QoS requirements? Reservation Styles • For the analytical models of Shared Tree and Dynamic Filter reservation styles, what are the effects of N sim_c/ian > 1, N sim_src > 1, and allowing the number of senders and receivers to be different? 'T h is was first, suggested by Van Jacobson and draws directly from his work on scaling session announcem ent frequency in the Session D irectory (S D ) application to m aintain a fixed signaling overhead independent o f number o f total sessions advertised. 86 • The current Chosen Source average case performance results were ob tained via simulation. Is it possible to formulate the average case perfor mance analytically, allowing direct confirmation of the bounded overhead of Dynamic Filter? 6.3.1.2 Fundam ental Q uestions We see two fundamental topics, related to our initial protocol performance and scal ing analysis, for further study. The first topic for consideration recognizes that as new applications are defined they may exhibit source and/or receiver interdependen cies differing from those supported by the Shared Tree or Dynamic Filter reservation styles. W hat additional reservation styles and mechanisms are required to efficiently support future ISPN applications? This task may be an ongoing project as the applications and ISPN service model evolve over time. Secondly, the current reservation scaling analysis results were all derived from a limited number of simulation scenarios or oversimplified topologies. It is important to explore to what extent these results apply to real networks. However, this question is more subtle than it first appears, since our results describe the large n limit of a network, where both the network and the number of resident hosts are growing. Two questions must be addressed. How can one characterize “real” networks? Certainly randomly generated networks are no more real than the simple topologies considered here. To some extent, real networks are the product of chaotic growth at the edges and planned growth in the interior. Assuming one can characterize more realistic networks, how can one explore the asymptotic limit? Should one hold the density fixed, or the ratio of the diameter to number of hosts, or is there some other criterion. In the simple topologies explored in the current scaling analysis the answer was clear, but in more general networks this is a completely unexplored issue. These issues are relevant to any investigation that depends on network topology. 6.3.2 R eservation D ynam ics There are a number of directions in which the reservation dynamics and thrashing investigation can be expanded. 6.3.2.1 D etailed Q uestions The following simulation and analysis scenarios would broaden the scope of the current analysis. Scaling Studies • In the current scenarios all significant delays are attributed to the tear- down delay. In large networks end-to-end propagation and admission control may also introduce significant delays. The effect of these other delay components on network throughput also need to be quantified. Sev eral scenarios using longer propagation delays have exhibited significant reductions in throughput, but h o deadlock. Are there scenarios that lead to thrashing? • Results for simulations of 4- and 8-way session models were surprising in that the larger sessions showed very little degradation. These similarities may have been an artifact of the dense distribution of session members across the 16 leaf nodes in the current acyclic network model. Simulations on larger networks with sparse membership distributions are required to determine if there is any differentiation in performance of the larger session models. Additionally, the N-way success session model does not scale to very large groups, and there are many applications that require only a subset of members to succeed. Simulations with further scaling of the session size and using a partial success model are required to quantify the dynamics of this scenario. Session M odels • The effects on application setup delay from retry-all-receivers and retry- blocked-receivers session retry polices is dependent on the behavior of the global set of session. W hat is the effect of a single greedy user, or all users acting in a, greedy manner? • Provided that the retry-blocked-receivers session retry policy is beneficial from an applications point of view, it may be useful to consider a hybrid retry policy incorporating both the exponential backoff and a session idle threshold. Under this scheme alter a specified number of failed reser vation request retry attempts, all session receivers would initiate retry in an attem pt to preempt any deadlock conditions. W hat is the effect of this hybrid scheme on system stability, throughput, and session setup duration? 6.3.2.2 Fundam ental Q uestions We see two fundamental questions to be answered related to our initial thrashing study. The first issue is related to the dependence between deadlock and thrashing. The majority of the scenarios exhibiting thrashing in the current study were induced in the operational region in which the system deadlocked. Thrashing was observed in a limited number of scenarios without deadlock, however the phenomena was produced consistently. The question to be answered is whether the deadlock region is the only scenario to guarantee thrashing or can other scenarios and retry policies also lead to thrashing consistently? Additionally, a greater understanding of the transition region between thrashing and non-decreasing throughput is required to accurately predict whether thrashing may occur for a specific teardown delay. Secondly, we have investigated the problem of thrashing from a technical perspec tive; however, there is an important underlying incentive issue. If we were looking at this as a unified design problem, where we could design the behavior of end users as well as the network, then there is little question that we could easily prevent thrashing. When sessions use the retry-all-receivers policy, and the retry interval is significantly longer than the teardown delay, we never observed thrashing. The problem is that sessions can determine their own session retry policy, and own retry timing, and these decisions need not be taken with the overall health of the network in mind. The retry-blocked-receivers session retry policy yields lower delays for the individual sessions adopting it. Reducing the retry interval also lowers delays. The result of each session optimizing its own performance is to send the network into a thrashing state. The ultimate question, therefore, is whether it’s possible for the network to actually proscribe specific user behavior by employing incentive mech anisms, or by isolating individual sessions from each other similar to that done by Fair Queueing [38] for transport data streams. R eferen ce L ist [1] J. M. Akinpelu. The overload performance of engineered networks with non-hierarchical routing. A T & T Bell Laboratories Technical Journal, 63(7), September 1984. [2] D. Anderson, R. Herrtwich, and C. Schaefer. SRP: A resource reservation protocol for guaranteed performance communication in the Internet. Technical Report TR-90-006, Computer Science Division, University of California at Berkeley, February 1990. [3] D. Anderson, S. Tzou, R. Wahbe, R. Govindan, and M. Andrews. Support for continuous media in the DASH system. Technical Report UC'B/CSD 89/537, Computer Science Division, University of California at Berkeley, October 1989. [4] Frederick T. Andrews. The evolution of digital loop carrier. IE E E Communi cations Magazine, 29(3), March 1991. [5] G. R. Ash, J. Chen, A. E. Frey, B. Huang, C. Lee, and G. L. McDonald. Real-time network routing in the AT&T network - improved service quality at lower cost. In Proceedings of IE E E Global Telecommunications Conference, December 1992. [6] ATM Forum. ATM user network interface specification, version 3.0. 1993. [7] ATM Forum. ATM user network interface specification, version 3.1. 1994. [8] U. Bagchi and R. Sullivan. Dynamic, non-preemptive priority queues with gen eral, linearly increasing priority function. Operations Research, 33, November 1985. [9] A. J. Ballardie. Core based trees (CBT) multicast: Architectural overview. Internet Draft, draft-ietf-idmr-cbt-arch-02.txt, June 1995. [10] A. J. Ballardie, P. F. Francis, and J. Crowcroft. Core based trees. In Proceed ings of AC M SIG CO M M ’ 93, 1993. [11] A. J. Ballardie, N. Jain, and S. Reeve. Core based trees (CBT) multicast: Protocol specification. Internet Draft, draft-ietf-idmr-cbt-spec-02.txt, June 1995. 90 [12] Anindo Banerjea, Domenico Ferrari, Bruce Mah, Mark Moran, Dinesh Verma, and Hui Zhang. The Tenet real-time protocol suite: Design, implementa tion, and experiences. Technical Report TR-94-059, International Computer Science Institute, Berkeley, California, November 1994. [13] F. Bernabei, M. Listanti, and D. Matrullo. Performance analysis of an ATM switch supporting a circuit emulation technique. Computer Communications, 14(9), November 1991. [14] R. Braden, D. Clark, and S. Shenker. Integrated services in the Internet architecture: an overview. Internet Request for Comments, RFC-1633, June 1994. [15] R. Braden, L. Zhang, D. Estrin, S. Herzog, and S. Jamin. Resource reservation protocol (RSVP) - version 1 functional specification. Internet Draft, draft-ietf- rsvp-spec-07.txt, July 1995. [16] Lee Breslau. Adaptive Source Routing of Real-Time Traffic in Integrated Ser vices Networks. PhD thesis, University of Southern California, December 1995. [17] R. Bubenik, J. DeHart, and M. Gaddis. Multipoint connection management in high speed networks. In Proceedings of IE E E Infocom ’ 91, April 1991. [IS] S. Casner and S. Deering. First IETF Internet audiocast. A C M SIG CO M M Computer Communications Review, 22(3), July 1992. [19] CCITT. Recommendation X.25: Interface between data terminal equipment (DTE) and data circuit-terminating equipment (DCE) for terminal operating in the packet mode on public data networks. Orange Book, VIII.2, 1980. [20] CCITT. Integrated services digital network (ISDN). Red Book, 3, 1985. Rec ommendations of the I-series. [211 CCITT Study Group XI. Specification of signaling system no. 7. Blue Book, VI, 1989. Fascicle VI.8. [22] A. Chesnais, E. Gelenbe, and I. Mitrani. On the modeling of parallel access to shared data. Communications of the ACM, 26(3), March 1983. [23] Israel C'idon and Inder S. Gopal. Paris: An approach to integrated high-speed private networks. International Journal of Digital and A nalog Cabled Systems, 1, 1988. [24] D. Clark. The design philosophy of the DARPA Internet protocols. In Pro ceedings of A C M SIG C O M M ’ 88, August 1988. 91 [25 [26 [27 [28 [29 [30 [31 [32 [33 [34 [35 [36 D. Clark, S. Shenker, and L. Zhang. Supporting real-time applications in an integrated services packet network: Architecture and mechanism. In Proceed ings of A C M SIG CO M M ’ 92, August 1992. Ron Cocchi, Deborah Estrin, Scott Shenker, and Lixia Zhang. A study of priority pricing in multiple service class networks. In Proceedings of A C M SIG CO M M ’ 91, August 1991. E. Coffman, M. Elphick, and A. Shoshani. System deadlocks. Computing Surveys, 3(2), February 1971. J. Crowcroft. personal communication. S. Deering. Host extensions for IP multicasting. Internet Request for Com ments, RFC-1112, August 1989. Stephen Deering. Multicast routing in internetworks and extended LANs. In Proceedings of A C M SIG CO M M ’ 88, 1988. Stephen Deering, Deborah Estrin, Dino Farinacci, Van Jacobson, Ching-Gung Liu, and Liming Wei. An architecture for wide-area multicast routing. In Proceedings of A C M SIG C O M M ’ 9Jt, August 1994. Stephen Edward Deering. Multicast Routing in a Datagram Internetwork. PhD thesis, Stanford University, December 1991. Steven Deering, Deborah Estrin, Dino Farinacci, Van Jacobson, Chinggung Liu, and Liming Wei. Protocol independent multicast (PIM): Motivation and architecture. Internet Draft, draft-ietf-idmr-pim-arch-01.txt, January 1995. Steven Deering, Deborah Estrin, Dino Farinacci, Van Jacobson, Chinggung Liu, and Liming Wei. Protocol independent multicast (PIM): Protocol speci fication. Internet Draft, draft-ietf-idmr-pim-spec-01.txt, January 1995. Mikael Degermark, Torsten Kohler, Stephen Pink, and Olov Schelen. Advance reservation for predictive service. In Proceedings of Fifth International Work shop on Network and Operatinq Systems Support for Distributed Audio and Video, April 1995. J. DeHart, M. Gaddis, and R. Bubenik. Connection management access pro tocol (CMAP) specification. Technical Report WUCS-92-01, Department of Computer Science, Washington University, 1992. [37] L. Delgrossi and L. Berger. Internet stream protocol version 2 (ST2): Protocol specification - version ST2+. Internet Draft, draft-ietf-st2-spec-04.txt, July 1995. 92 [38] Luca Delgrossi, Ralf Guido Herrtwich, and Frank Oliver Hoffmann. An imple mentation of ST-II for the Heidelberg transport system. [39] Luca Delgrossi, Ralf Guido Herrtwich, Frank Oliver Hoffmann, and Sibylle Schaller. Receiver-initiated communication with ST-II. [40] Luca Delgrossi, Sibylle Schaller, Hartm ut Wittig, and Lars Wolf. Issues of reserving resources in advance. In Proceedings of Fifth International Workshop on Network and Operating Systems Support for Distributed Audio and Video, April 1995. [41] A. Demers, S. Keshav, and S. Shenker. Analysis and simulation of a fair queueing algorithm. Journal of Internetworking: Research and Experience, 1, 1990. [42] Z. Dziong and L. Mason. An analysis of near optimal call admission and routing model for multi-service loss networks. In Proceedings of IEEE lnfocom '92, 1992. [43] Deborah Estrin, Scott Shenker, and Daniel Zappala. Routing support for RSVP. Internet Draft, draft-ietf-rsvp-routing-OO.txt, June 1995. [44] D. Ferrari and D. Verma. A scheme for real-time channel establishment in wide-area networks. IE E E Journal on. Selected Areas in Communications, 8(3), April 1990. [45] Domenico Ferrari, Amit Gupta, and Giorgio Ventre. Distributed advance reservation of real-time connections. Technical Report TR-95-008, University of California at Berkeley, March 1995. [46] Sally Floyd. Link-sharing and resource management models for packet net works, September 1993. [47] Fore Systems, Inc. SPANS: Simple protocol for ATM network signalling, ver sion 2.1.1. Technical report, Fore Systems, Inc., 1992. [48] Fore Systems, Inc. ATM networks. Technical report, Fore Systems, Inc., 1993. [49] J. Forgie. ST - a proposed internet stream protocol. Internet Experimental Notes, IEN-119, September 1979. [50] M. Gardner, I. Loobeek, and S. Cohn. Type-of-service routing with loadshar- ing. In Proceedings of IE E E GlobeCom ’ 87, November 1987. [51] V. Gligor and S. Shattuck. On deadlock detection in distributed systems. IEEE Transactions on Software Engineering, SE-6(5), September 1980. 93 [52] B. Goldman. Deadlock detection in computer networks. Technical Report MIT-LCS-TR-185, Massachusetts Institute of Technology, September 1977. [53] S. J. Golestani. A stop and go queueing framework for congestion management. In Proceedings of A C M SIG C O M M ’ 90, 1990. [54] P. Green. An introduction to network architectures and protocols. IEEE Transactions on Communications, 28(4), June 19S0. [55] Amit Gupta and Mark Moran. Channel groups: A unifying abstraction for specifying inter-stream relationships. Technical Report TR-93-015, University of California at Berkeley, March 1993. [56] C. Hedrick. Routing information protocol. Internet Request for Comments, RFC-1058, June 1988. [57] Ralf Guido Herrtwich. An introduction to real-time scheduling. Technical Report TR-90-035, Computer Science Division, University of California at Berkeley, July 1990. [58] Frank Oliver Hoffmann and Luca Delgrossi. ST-II for the Heidelberg transport system: Protocol, design and implementation. [59] J. Hyman, A. Lazar, and G. Pacifici. Real-time scheduling with quality of service constraints. IE E E Journal on Selected Areas in Communications, 9(9), September 1991. [60] ITU-T. B-ISDN user network interface layer 3 specification for basic call/bearer control. Draft Recommendation Q.93B, October 1992. [61] ITU-T. Broadband aspects of ISDN (B-ISDN protocol reference model). Rec ommendation 1.121, June 1993. [62] ITU-T. Broadband integrated services digital network (B-ISDN), digital sub scriber signalling system no. 2, user network interface layer 3 specification for basic call/connection control. Draft Recommendation Q.2931, June 1994. [63] ITU-T. Broadband integrated services digital network (B-ISDN), digital sub scriber signalling system no. 2, user network interface layer 3 specification for point-point multiconnection call control. Draft Recommendation Q.298x, June 1994. [64] ITU-T. Digital subscriber signalling system no. 1 (DSS 1) - ISDN user-network interface layer 3 specification for basic call control. Recommendation Q.931, November 1994. 94 [65] ITU-T. Negotiation/renegotiation: Traffic and QoS negotiation during active phase. Draft Recommendation Q.2963, 1994. [66] ITU-T. Negotiation/renegotiation: Traffic and QoS negotiation during call establishment. Draft Recommendation Q.2962, 1994. [67] ITU-T. Negotiation/renegotiation: Traffic and QoS parameters. Draft Rec ommendation Q.2961, 1994. [68] Van Jacobson and David D. Clark. Flexible, efficient resource management for datagram networks, unpublished work. [69] S. Jamin, S. Shenker, L. Zhang, and D. Clark. An admission control algorithm for predictive real-time service (extended abstract). In Proceedings of the Third International Workshop on Network and Operating System Support for Digital A udio and Video, November 1992. [70] Sugih Jamin, Peter B. Danzig, Scott Shenker, and Lixia. Zhang. A measurement-based admission control algorithm for integrated services packet networks. In Proceedings of AC M SIG CO M M ’ 95, 1995. [71] C. Kalmanek, H. Kanakia, and S. Keshav. Rate controlled servers for very high-speed networks. In Proceedings of IE E E GlobeCom. ’ 90, 1990. [72] Gunnar Karlsson and Martin Vetterli. Packet video and its integration into the network architecture. IEEE Journal on Selected Areas in Comm unications, 7(5), June 1989. [73] F. Kelly. Loss networks. The Annals of Applied Probability 1, 1991. [74] A. Klianna and J. Zinky. The revised ARPA routing network. A C M Computer Communications Review, 1989. [75] Leonard Klienrock. Queueing Systems, Volume 11: Computer Applications. John Wiley and Sons, 1976. [76] Youngo Lim and John Kobza. Analysis of a delay-dependent priority discipline in a multi-class traffic packet switching node. In Proceedings of IE E E Infocom ’ 88, 1988. [77] Arthur Y. Lin and John A. Silvester. Priority queueing strategies and resource allocation protocols for traffic control at an ATM integrated broadband switch ing systems. Technical Report CSI-90-11-02, University of Southern California Communications Sciences Institute, 1990. [78] J. C. Luetchford. CCITT recommendations - network aspects of the ISDN. In Proceedings of the International Conference on Communications, June 1985. [79] M. Maekawa, A. Oldehoeft, and R. Oldehoeft. Operating Systems Advanced Concepts, chapter 7. Benjamin/Cummings, 1987. [80] M. Maekawa, A. Oldehoeft, and R. Oldehoeft. Operating Systems Advanced Concepts, chapter 5, page 139. Benjamin/Cummings, 1987. [81] G. Malkin. RIP version 2 - carrying additional information. Internet Request for Comments, RFC-1388, January 1993. [82] L. G. Mason. On the stability of circuit-switched networks with non- hierarchical routing. In Proceedings of the 25th Conference on Decision and Control, December 1986. [83] I. M atta and U. Shankar. Type-of-service in adaptive next-hop routing. Tech nical report, Department of Computer Science, University of Maryland, 1993. [84] D. Menasce and R. Muntz. Locking and deadlock detection in distributed databases. IEEE Transactions on Software Engineering, SE-5, May 1979. [85] S. Minzer. A signalling protocol for complex multimedia, services. IEEE Jour nal on Selected Areas in Communications, December 1991. [86] M. Moran and R. Gusella. Network support for interactive multimedia. In Proceedings of the Third International Workshop on Network and Operating System Support for Digital Audio and Video, November 1992. [87] J. Moy. Multicast extensions to OSPF. Internet Request for Comments, RFC- 1584, March 1994. [88] J. Moy. OSPF version 2. Internet Request for Comments, RFC-1583, March 1994. [89] Biswanath Mukherjee. Integrated voice-data communication over high-speed fiber optic networks. IE E E Computer, 24(2), February 1991. [90] John Nagle. On packet switches with infinite storage. IE E E Transactions on Communications, 35, 1987. [91] R. Obermarck. Distributed deadlock detection algorithm. A C M Transactions on Database Systems, 7(2), June 1982. [92] D. Oran. OSI IS-IS intra-domain routing protocol. Internet Request for Com ments, RFC-1142, February 1990. [93] A. Parekh. A generalized processor sharing approach to flow control in inte grated services networks. Technical Report LIDS-TR-2089, Laboratory for In formation and Decision Systems, Massachusetts Institute of Technology, 1992. 96 [94] C. Partridge and S. Pink. An implementation of the revised Internet stream protocol (ST-2). Journal of Internetworking: Research and Experience, 3(1), March 1992. [95] Craig Partridge. A proposed flow specification. Internet Request for Com ments, RFC-1363, September 1992. [96] J. Peterson and A. Silberschatz. Operating System Concepts, chapter 8, page 261. Addison-Wesley, 1983. [97] J. Postel. Internet protocol. Internet Request for Comments, RFC-791, September 1981. [98] J. Postel. Transmission control protocol. Internet Request for Comments, RFC-793, September 1981. [99] E. Rathgeb. A comparison of policing mechanisms for ATM networks. In Proceedings of IEEE Infocom ’ 92, June 1992. 1001 Wilko Reinhardt. Advance reservation of network resources for multimedia applications. In Proceedings of ICAW A '94, October 1994. 1011 L. Roberts. The evolution of packet switching. Proceedings of the IEEE, November 1978. 1021 In Kyung Ryu and Alexander Thomasian. Analysis of database performance with dynamic locking. Journal of the ACM , 37(3), July 1990. 103] S. Shenker, D. Clark, and L. Zhang. A scheduling service model and a schedul ing architecture for an integrated services packet network, preprint submitted to A CM /IEEE Transactions on Networking, 1993. 104] S. Shenker and C. Partridge. Specification of guaranteed quality of service. Internet Draft, draft-ietf-intserv-guaranteed-svc-01.txt, July 1995. 105] S. Shenker and C. Partridge. Specification of predictive quality of service. Internet Draft, draft-ietf-intserv-predictive-svc-00.txt, March 1995. 1061 S. Shenker, C. Partridge, and J. Wroclawski. Specification of controlled delay quality of service. Internet Draft, draft-ietf-intserv-control-del-svc-01.txt, June 1995. [1071 S. Shenker and J. Wroclawski. Network element service specification template. Internet Draft, draft-ietf-intserv-svc-template-01.txt, June 1995. 1081 Paul W. Shumate and Richard K. Snelling. Evolution of fiber in the residential loop plant. IEEE Communications Magazine, 29(3), March 1991. 97 109] Shiow-Chen Shyu and Victor O. K. Li. Performance analysis of static locking in distributed database systems. IE E E Transactions on Computers, 39(6), June 1990. 110] Sandeep Sibal and Antonio DeSimone. Controlling alternate routing in general-mesh packet flow networks. In Proceedings of A C M SIG CO M M '9Jh August 1994. 111] F. Solensky. monthly nsfnet growth chart, posted to Big-Internet mailing list (Big-Internet@munnari.OZ.AU). 11‘ 2] M. Steenstrup. Inter-domain policy routing protocol specification: Version 1. Internet Request for Comments, RFC-1479, July 1993. 113] Y. C. Tay. Locking Performance in Centralized Databases. Academic Press, 1987. 114] C. Topolcic. Experimental Internet stream protocol: Version 2 (ST-II). Inter net Request for Comments, RFC-1190, October 1990. 115] Jonathan S. Turner. New directions in communications (or which way to the information age?). IE E E Communications Magazine, 24(6), October 1986. 116] D. Verma, II. Zhang, and D. Ferrari. Delay jitter control for real-time commu nication in a packet switching network. In Proceedings of Tri.Com. '91, 1991. 117] D. Waitzman, C. Partridge, and S. Deering. Distance vector multicast routing protocol. Internet Request for Comments, RFC-1075, November 1988. 118] Liming Wei and Deborah Estrin. The trade-offs of multicast trees and algo rithms. Technical Report TR93-560, University of Southern California, 1993. 119] Liming Wei and Deborah Estrin. Multicast routing in dense and sparse modes: Simulation study of tradeoffs and dynamics. Technical Report TR95-613, Uni versity of Southern California, 1995. 120] Patrick E. White. The role of the broadband integrated services digital net work. IEEE Communications Magazine, 29(3), March 1991. 121] Ouri Wolfson. The overhead of locking (and commit) protocols in distributed databases. AC M Transactions on Database Systems, 12(3), September 1987. 122] J. Wroclawski. Integrated services WG flowspec definition. Internet Draft, draft-ietf-intserv-flowspec-fmt-00.txt, July 1995. 123] L. Zhang, S. Deering, D. Estrin, S. Shenker, and D. Zappala. RSVP: A new resource reservation protocol. IEEE Network Magazine, September 1993. 98 [124] Lixia Zhang. A New Architecture for Packet Switching Network Protocols. PhD thesis, Massachusetts Institute of Technology, August 1989. [125] Lixia Zhang. VirtualClock: A new traffic control algorithm for packet switch ing networks. In Proceedings of A C M SIG C O M M ’ 90, August 1990. 99
Linked assets
University of Southern California Dissertations and Theses
Conceptually similar
PDF
00001.tif
PDF
00001.tif
PDF
00001.tif
PDF
00001.tif
PDF
00001.tif
PDF
00001.tif
PDF
00001.tif
PDF
00001.tif
PDF
00001.tif
PDF
00001.tif
PDF
00001.tif
PDF
00001.tif
PDF
00001.tif
PDF
00001.tif
PDF
00001.tif
PDF
00001.tif
PDF
00001.tif
PDF
00001.tif
PDF
00001.tif
PDF
00001.tif
Asset Metadata
Core Title
00001.tif
Tag
OAI-PMH Harvest
Permanent Link (DOI)
https://doi.org/10.25549/usctheses-oUC11256545
Unique identifier
UC11256545
Legacy Identifier
9617123