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
/
Computer Science Technical Report Archive
/
USC Computer Science Technical Reports, no. 763 (2002)
(USC DC Other)
USC Computer Science Technical Reports, no. 763 (2002)
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
1
A Lightweight Mechanism to Control Misbehaving
Flows using AQM-based Aggregate Marking
Abhimanyu Das, Debojyoti Dutta, Ahmed Helmy
University of Southern California, Los Angeles, CA 90089, USA
Abstract
In this paper, we present a lightweight scheme based on stateless aggregate fair packet marking at the network edge followed by
simple RIO queueing at core nodes, to control misbehaving flows. We develop a new class of AQM-based fair, stateless and aggregate
packet markers that fairly mark packets of misbehaving flows entering an edge node, which are then dropped preferentially at a RIO-
enabled (RED with IN/OUT) router during congestion. Our edge markers imitate fair AQM schemes such as CHOKe and CSFQ to
mark flows fairly, which dampens misbehaving flows at the bottleneck router during congestion. We show, through extensive packet
level simulations that even simple and stateless edge-based marking schemes followed by differential dropping can effectively damp
misbehaving flows at congested routers, and provide much better performance to best-effort TCP flows. Our markers are stateless
and operate on edge aggregates, and we do not introduce any complicated AQM algorithm at the network core. Thus, our solution is
orthogonal to the more complex issue of fair bandwidth sharing.
I. INTRODUCTION
Fair sharing of network resources and controlling misbehaving flows at congested routers are important thrusts of Internet
research. Resource misappropriation by high-bandwidth, non congestion-responsive flows can lead to degraded performance of
congestion-reactive flows, and add to network congestion [11]. This threat is more pronounced in the face of growing UDP,
gaming, voice and video traffic which can quickly squeeze out other TCP( [25], [16]) flows, and can even lead to a congestion
collapse[19] in the Internet. Hence, the identification and the subsequent punishing of misbehaving flows (non-responsive flows
who deviate substantially from the fair bandwidth share) may be very crucial for better performance of the network.
Abhimanyu Das is with the Department of Computer Science at the University of Southern California, and Mahi Networks Inc. He can be contacted at
abhimand@usc.edu
Debojyoti Dutta is with the Department of Computer Science at the University of Southern California / Information Sciences Institute. His work is supported by
DARPA and the Space and Naval Warfare Systems Center San Diego (SPAWAR) under Contract No. N66001-00-C-8066. He can be contacted at ddutta@isi.edu
Ahmed Helmy is with the Department of Computer Engineering at the University of Southern California. He can be contacted at helmy@ceng.usc.edu. His
work is partially supported by NSF Career Award 0134650.
2
One way to control misbehaving flows is to ensure fair bandwidth allocation among flows entering a congested router. Numer-
ous fair AQM [9] schemes have been proposed to solve this problem at congested routers. However, achieving fair bandwidth
allocation in a network is a complex problem and cannot be accurately solved without making non-trivial changes to the routers
deployed. We therefore raise a question about whether it is useful in a best-effort Internet to develop schemes which may not
come close to providing exact max-min fairness in bandwidth allocation, yet may reduce unfairness among flows to a large extent
and are simple and easy to deploy.
In this paper, we present a lightweight architecture to control misbehaving flows at a congested router with minimal changes to
existing infrastructure. Our scheme, in spite of its simplicity, has the potential to drastically contain the effect of non-responsive,
high bandwidth flows on well-behaved best effort traffic. We use fair sharing algorithms to mark packets of incoming flows at all
the edge nodes. Our scheme is primarily edge-based, and the network core is kept as simple and unaltered as possible and is only
required to run RIO [4] as its queue management policy. We design fair packet markers which ensure that all flows that misbehave
(with respect to other flows entering through the same edge node) automatically get more OUT tokens, and hence, get penalized
preferentially by RIO, during congestion. We introduce a new class of Active Queue Management-inspired marking schemes for
this purpose, which are simple, stateless and scalable in design. We call them AQM-based markers. Note that now the problem
of fair sharing of resources at the core nodes is related to the fair allocation of tokens at the edge markers. The simplicity of
our architecture comes from the fact that we do not aim to solve the fair bandwidth allocation problem at each node, but only to
penalize or dampen misbehaving flows considerably.
We make a couple of assumptions in this paper. The first assumption is that misbehaving flows are always grouped together
with other flows in the edge to form an aggregate of a large number of flows before they enter the core. Our aggregate edge
marking scheme may not be able to effectively penalize misbehaving flows if they are not grouped together with a sufficiently
large number of flows. Our second assumption is that most core routers today have already implemented RIO or variants like
WRED for providing differentiated services [28]. Under this assumption, our scheme can be easily deployed, since we keep
the core virtually unchanged from its current version. Note that our scheme is designed for best-effort traffic only, and we do
not consider flows which make use of QoS mechanisms such as Diffserv [2] and Intserv [31], and which require specific QoS
guarantees and bandwidth requirements. In such a scenario, our definition of misbehaving flows therefore includes all those flows
that are not congestion-reactive and generate traffic at a considerably higher rate than other flows.
Our architecture may appear similar to the differentiated services model for obtaining QoS, but the resemblance is limited. Our
goal is to penalize misbehaving flows at congested routers. Unlike Diffserv [2], we do not provide any QoS guarantees. Secondly,
even though we use the concepts of packet marking and RIO as in Diffserv, we only segregate our packets into two classes, and we
3
use aggregate marking at edge-core boundaries. We validate our architecture using detailed packet level simulation on different
topologies and for various traffic patterns. We also show that our architecture can indeed control misbehaving flows using very
simple and informal game theoretic arguments.
The structure of the paper is as follows. We consider related work in Section II. In Section III, we describe our architecture.
Section IV describes the three markers in greater details. In Section V, we evaluate the performance of our architecture. We
use two different topologies and a variety of network traffic to show how we contain misbehaving flows. Finally, we discuss our
directions for future work and conclude in Section VI.
II. RELATED WORK
Our work leverages techniques from different related research areas such as queueing and scheduling and traffic markers. An
exhaustive survey of related work is beyond the score of this paper. In this section, we will try to highlight previous work that
is directly relevant to our problem of controlling high bandwidth, misbehaving flows. A superset of the problem of controlling
misbehaving flows is the fair bandwidth allocation problem. We can therefore divide previous related work into those that address
the problem of explicit fair resource allocation in routers, and those that look primarily at identifying and containing misbehaving
flows.
A general scheme for fair resource allocation is to ensure fair scheduling or fair buffer-allocation at the router. However
scheduling techniques to address fair allocation (such as Fair Queueing [7]) are complex, and not scalable. Some good advances
have been made recently in developing active queue management algorithms which provide fair buffer allocation and hence contain
the effects of misbehaving flows. CSFQ [26] is one such stateless scheme which uses flow rates embedded in packet headers to
approximate max min fairness at the core. It however requires an extra field in the packet header and uses a complex fair rate
computation algorithm at the core routers. FRED [17] is an enhancement to RED to ensure fair buffer sharing by maintaining
state for active flows at the router.
Other AQM schemes like as CHOKe [23] and RED-PD [18], directly address the problem of identifying and penalizing
misbehaving flows, and are less complex to implement than fair queueing. In [24], the authors develop the CHOKe algorithm,
where they modify RED to identify and penalize misbehaving flows by looking at an incoming packet and comparing its flow-id
with a random packet in the packet queue. Ott et. al. [22] proposed a variation of RED, called SRED, to identify candidates for
high bandwidth flows from a cache of recently seen flows. RED-PD [18] uses a list of previous RED packet drops to identify
misbehaving flows and control them using different drop probabilities.
However, the above mentioned AQM mechanisms still need considerable changes to the forwarding engines (such as additional
state or more complex queueing computation) of most existing core routers. Most current core routers are only enabled with RED
4
or RIO (or variants like WRED [28]) queue management algorithms. Therefore we believe many of these new AQM mechanisms
will be difficult to deploy without the disruption of existing network infrastructure.
Packet Marking
In our work we introduce a novel class of AQM based fair aggregate marking. Our traffic markers are an important component
of our work. Traffic marking is an essential component of the Diffserv [4], [20], [2] architecture too. However the fair marking
schemes that we propose in this paper are different from most other Diffserv markers developed - in terms of being aggregate-
based, stateless and fair at the same time. In the following paragraphs,we explore some related work in the field of traffic marking,
all of which come from the Diffserv domain.
Markers can be classified into two groups: per-flow markers and aggregate markers. Most sophisticated markers are per-flow
based [30], [10] and are thus not very scalable. Many of these markers only address TCP flows. Most current aggregate traffic
markers [13], [14], [30], [27] mark individual packets of an aggregate using simple token bucket-like schemes and do not provide
fairness. A more complete review of the marking techniques can be found in [5] In [15] the authors propose using an FRED
algorithm to create a fair aggregate traffic marker. However this require per-active flow computation at the marker and is also
more complex to implement than our markers.
III. OUR APPROACH
In this paper, we present a lightweight architecture that marks misbehaving flows equitably at the edges without maintaining
state. The core is kept as simple as possible, and is required to only run RIO as its queue management policy. In fair AQM
strategies, the router uses fair bandwidth allocation algorithms to contain misbehaving flows. We push this intelligence of fair
allocation toward the edges in the form of fair IN token allocation for aggregate marking.
We leverage queueing and dropping policies of existing stateless AQM algorithms to design our stateless, aggregate packet
markers, which we call AQM-based markers. To do this, we observe that a target rate can be specified by a token bucket (TB).
This TB specification can be considered to be equivalent to a queue with the Token-Bucket burst size as the maximum queue size
and the token arrival rate as the queue’s link bandwidth. Marking an arriving packet as IN is equivalent to queueing a packet for
transmission, and marking it as OUT is equivalent to dropping an incoming packet from the queue. The average queue size and the
average token bucket size give an indication of the congestion level at the queue and TB respectively. However, a small average
queue size is equivalent to a large average token bucket size and vice versa. The problem of fair token distribution at a marker
using a token bucket traffic specification among packets of an aggregate of flows, can therefore be viewed as being equivalent
to efficient buffer management and scheduling of incoming packets at a queue to provide fair bandwidth distribution among the
5
flows arriving at the queue. This novel approach of our scheme makes it easier to adapt well studied AQM techniques at edge
markers, than to actually implement those AQM techniques in each core router.
We use these AQM-based fair markers at network edges to implicitly identify and control the effect of misbehaving flows
among flow aggregates of an edge domain, before they enter the network core, as shown in Figure 1. A fair packet marker at each
network edge fairly distributes the total number of IN tokens among flows and classifies the packets of the flow aggregate as IN
or OUT. IN packets are those packets which adhere to their flow’s fair share of the aggregate bandwidth while OUT packets are
the extra packets beyond the fair share. A misbehaving flow thus has a much higher fraction of OUT packets compared to that of
other responsive flows under any fair marking scheme. The RIO queueuing policy at a congested router preferentially drops OUT
packets over IN packets during congestion. Hence, if we can fairly mark packets in this manner to allocate a much greater number
of OUT packets to high bandwidth flows of an aggregate, then during congestion at a node running RIO, these OUT packets will
be dropped before IN packets. This will ensure that, at a congested router, misbehaving flows have a much lower bandwidth
utilization than what they would have obtained in the absence of such fair marking. Fair token allocation at the edge will therefore
dampen the effects of misbehaving flows (by forcing them to share IN tokens with other flows in their aggregate) and reduce their
impact on other well-behaved flows at the congested router.
To ensure approximate bandwidth sharing at the core router, we need to set the marking token-rate at the edge router such that
the sum of the token rates at all ingress edge markers is approximately equal to the bottleneck router bandwidth. As a first step,
we assume that these can be configured using Service Level Agreements (SLA) between domains. Later we relax this assumption
and use little information about the number of flows to dynamically calculate what the target token rate at each edge router should
be. However we are still studying the issue of optimal edge token-rate assignment in greater depth.
Markers, like AQMs are either stateless or stateful. The stateful, per-flow fair markers have the same scalability problems as
their AQM counterparts. Hence, we choose algorithms from a few stateless fair AQM schemes to mark packets in our architecture.
However we feel that almost all fair AQM techniques (like SRED, BLUE, RED-PD) and even fair scheduling algorithms(like Fair
Queueing ) can be adopted to create fair aggregate markers to control misbehaving flows at the congested routers.
We devise three such sample AQM-markers which are based on well known AQM techniques:
1) Probabilistic Aggregate Marker(PAM): This is influenced by RED [12].
2) Fair Stateless Aggregate Marker (F-SAM): Influenced by CSFQ [26], it ensures almost fair distribution of IN packets.
3) CHOKed Aggregate Marker (CAM): As the name suggests, we use an adaptation of the CHOKe [23] algorithm to ensure
that misbehaving flows are punished by marking them as OUT. It should be noted that it was non-trivial to imitate CHOKe,
and hence, the design of an effective CAM marker design was challenging. Thus, the CAM marker is one of the original
6
contributions of this paper apart from our lightweight architecture to contain misbehaving flows. We believe that our CAM
marker can also be effectively used in the context of providing QoS in a Diffserv environment.
Our approach as discussed above only dampens the effect of misbehaving flows, and does not guarantee max-min fair bandwidth
allocation of goodputs. It is easy to show that fair packet marking of an edge aggregate may not translate into exact fair bandwidth
allocation at a congested bottleneck router. However, our fair token allocation does ensure that a high bandwidth flow will get
a higher number of OUT tokens, and thus its packets will observe a higher drop rate than in the absence of any such marking.
The dampening of the high bandwidth flow increases with the number of total flows in its aggregate, since the higher the number
of flows in the aggregate, the lesser the number of IN tokens which a fair marking scheme will allocate to the misbehaving flow.
However our scheme may not be able to yield globally fair allocations as we do not maintain any state. We then relax the stateless
assumption and ask whether we can provide fairer allocation and better dampening of misbehaving flows, with some additional
information about the ratio of the number of flows at an edge router to the number of flows at the congested router. Our results
show that we can indeed damp misbehaving flows more effectively with the use of such feedback from the congested routers.
RIO
RIO
RIO
RIO
RIO
RIO
RIO
RIO
EM
EM
EM
Edge Domain
Core Domain
EM Edge AQM Marker
Core Nodes with RIO
Edge Nodes
token allocation at edge
AQM-Marking based fair
Fig. 1. AQM marking +RIO architecture to control misbehaving flows at congested router
IV. STATELESS AGGREGATE AQM-BASED MARKING
The core building block of our architecture are stateless, aggregate traffic markers that can ensure approximate fair allocation
of IN packets among different flows at the edge. In this section, we describe our three markers, PAM (Probabilistic Aggregate
Marker), F-SAM (Fair Stateless Aggregate Marker) and CAM (CHOKed Aggregate Marker) that satisfy the above requirements.
Das et. al. [5], introduce PAM and F-SAM in more detail, in a DiffServ context. Note that our PAM marker does not achieve fair
token allocation in its marking. However we have included it, because it is derived from the most common AQM strategy, RED,
and is a baseline against which we compare our more advanced fair aggregate markers.
7
Traffic is specified using a token bucket defined by the average rate and a burst size. If the traffic conforms to the token bucket
specifications, it is marked as IN or in-profile. Else it is marked as OUT or out-of-profile. Thus, the fairness problem in the
marking domain is to fairly distribute the tokens among all the flows.
A. Probabilistic Aggregate Marker (PAM)
Our Probabilistic Aggregate Marker (PAM) is based on RED (Random Early Detection) [12]. Intuitively, the idea is as follows.
The aggregate traffic consumes tokens from the token bucket at a certain rate. We maintain an exponentially weighed moving
average (EWMA) of the number of tokens in the token bucket. On every incoming packet, we look at this average token bucket
size. If this bucket size falls below a certain threshold
, all packets are to be marked as OUT. If the bucket size varies
between
and
, we mark the packet as OUT based on a probability function that depends on the exponential weighted
moving average of the token bucket size. If the token bucket size exceeds
, we mark the packet as IN. More formally, our
probability function of marking a packet as OUT, can be written as follows.
!"
#%$&(’*)#+$,.-
/10325476
) /987:;476< % >= ? ! @ A! (1)
where
is the average size of the token bucket. Our probability function for marking a packet as OUT, is therefore a modification
of the RED probability function for accepting a packet into the queue. [5] contains a more detailed explanation of the PAM
marker.
B. Fair Stateless Aggregate Marker (F-SAM)
Stoica et al. [26] proposed a Core Stateless Fair Queueing algorithm to approximate max-min fairness while buffering incoming
packets at a queue, by using a probabilistic dropping function based on the average rate of a flow to which the packet belongs.
This rate information, instead of being calculated at the queue using per-flow techniques is calculated near the source of the flow
and inserted in every packet header. We adapt CSFQ in our F-SAM marker (Fair Stateless Aggregate marker).
In F-SAM, the rate information in each packet header is calculated and filled by the ingress node from where the flow originates.
Since each ingress node is responsible for maintaining the rate of only the flow that enters through it, there is no scalability issue
involved in this per-flow rate calculation. At the edge domain, the edge marker needs to calculate the fair rate, B , allocated to the
flows, and then calculates the token allocation probability (marking probability) of a packet as
%CED F , where G is the rate of the
corresponding flow. The expected rate of tokens or IN packets allocated to a flow of rate G , is then
HG
C B , which corresponds
8
to the max-in fair rate. As in [26], the fair-rate B , is calculated by an iterative method based on the average aggregate arrival rate
of packets , estimate of the aggregate token allocation rate of packets , and the token bucket rate . For more details, see [5].
This scheme ensures that tokens are allocated among the various flows fairly, by approximating a max-min fairness token
allocation strategy, based on the rate information in the packet headers. This fair token allocation is performed at the marker
without any per-flow measurements or state.
C. CHOKe’d Aggregate Marking (CAM)
In this section, we describe CHOKed Aggregate Marking (CAM), one of the primary markers introduced in the paper. As the
name suggests, it was heavily inspired by CHOKe [24]. However there are significant differences and we will enumerate them in
this section.
CHOKe is an AQM strategy that attempts to approximate fair bandwidth allocation among flows. The basic idea of CHOKe is
as follows: CHOKe uses a FIFO buffer and a transfer function like RED parameterized by the tuple
C C / A [12]. When a packet arrives at the router, the algorithm checks whether the average buffer length is greater than
. If it is
not, the packet is allowed to pass. If it is greater, then its flow-id (which could be defined as the packet’s source and destination
address tuple) is compared to that of a random packet in the queue. If the two flow-ids match, then both the packets are dropped,
else only the new packet is dropped according to the RED transfer function. If the average buffer length is greater than
,
the packet is dropped with a probability of 1.
Adapting a CHOKe-like scheme in a marker is not trivial. The main problem is as follows. CHOKe compares the incoming
packet, and, drops both if their flow-ids match. Now, we simply mark packets and leave the dropping to the congested router.
Hence we cannot penalize or remark packets that have already been injected into the network. So we need to use a slightly
different algorithm that is still based on the principles of CHOKe.
Intuitively, instead of dropping both the random packet in the buffer and the incoming packet when their flow-ids match, as in
CHOKe, we mark the current packet as OUT and remember its flow-id. We call these flows criminals. When our marking scheme
encounters the first packet belonging to a criminal flow, it marks it as OUT. Hence we need to maintain two queues - one to store
the to be sampled packets denoted by , and, the other for keeping track of criminals denoted by . Note, we need to maintain
only flow-ids of packets in both the queues, and not the entire packets themselves.
When any packet enters the CAM marker, we meter and calculate the average token bucket size. The average is computed using
EWMA. Then we check whether it is a criminal. This is done using a lookup of the queue. If it is indeed a criminal, the packet
is immediately marked as OUT. Also, we remove its identity from the queue. If the packet is not a criminal, we calculate
the probability of marking the packet as OUT using Equation 1, denoted by
/10
F . If
/90
F @ , we mark the packet as IN. If,
9
Algorithm 1: CAM marker
Data : : Queue of packet history;
: Queue of criminals;
: current time;
: current token bucket size;
foreach Incoming packet: pkt do
5G
;
;
if
5G then
G >
;
5G
;
else
G ! B G C C "
C % C ;
if
>
@ then
G >$#%
;
else
&>
G 5
’ ;
if
B ( >
& B (
then
G )
;
5G
;
else
mark
as
with prob
;
*
@ /10
F,+
, we sample the queue and choose a random victim. If the flowid of the incoming packet is same as that of the
victim, we mark this packet as OUT and enqueue this packet into the list of criminals , else we mark the incoming packet as
OUT with a probability
/10
F and IN with a probability
= /10
F . At the end of the algorithm, we enqueue the incoming packet
into the . The details are given in Algorithm 1
This algorithm is essentially like CHOKe. The only difference is that, instead of dropping a packet from the queue when the
flow-ids match during the random sampling process, we penalize the next packet belonging to the same flow. One might argue
that this dropping of the next packet belonging to the same flow is harmful for TCP flows due to TCP’s bias against multiple drops
in a single window. However in our simulations we observe no such bias of CAM against TCP flows.
Also note that even though the token distribution of IN packets to all the TCP and misbehaving flows is equitable, this will
not translate to an exactly equal division of throughput at the network core among a TCP and a misbehaving flow, if the core
only implements simple RIO queueing. This is because the loss of any packet of a TCP flow would result in its source cutting
back on its sending rate, while it does not affect the sending rate of the misbehaving flow. However, preventing large flows from
getting much more IN tokens than TCP flows would significantly increase the throughput obtained by the TCP flows in presence
of misbehaving traffic. Since this is a probabilistic marker, it will also be more TCP-friendly in terms of removing any bias against
bursty flows.
In the next section, we show that our CAM and F-SAM markers, along with RIO queueing at the core can penalize high
10
bandwidth flows at congested routers and achieve a much more fair bandwidth allocation among flows.
V. RESULTS
n0
n1
nk
s0
s1
si
d0
d1
d2
dj
1 Mbps
5ms 5ms
10Mbps
10Mbps
5ms
10Mbps
5ms
e0
e1
e2 c
e3
Fig. 2. Simulation scenario: A simple core with two edges. The inject traffic of different classes into the network while the
s are the sinks.
In this section we evaluate our approach to controlling high bandwidth flows at congested routers with the help of our AQM
markers at the network edges followed by differential dropping at the congestion point using simple RIO queueing. Using packet
level network simulations, we investigate the validity of our claims made in the previous sections about containing the effect
of misbehaving flows. We model misbehaving flows as high-bandwidth CBR flows over UDP. First, we study the end-to-end
throughout performance for traffic aggregates getting marked at a single edge domain and then consider a mixture of traffic
aggregates from multiple edge domains, interacting at the core. In each of the above cases, we study simple scenarios with UDP
flows and then move on to consider complex scenarios with both TCP as well as misbehaving UDP traffic.
A. Experimental Setup
We simulated the effect of our architecture with the ns-2 [29] network simulator The topology used is depicted in Figure 2. The
source nodes are
8 and the destination or sink nodes are
8 . The two edge nodes are and connected to . The source nodes
to
8 and
to
are connected to the edge nodes C respectively and inject traffic into the network core,
. The core is
connected to the edge which is further connected to the sink nodes
8 s. The source node
is used to generate background
traffic in form of many TCP flows carrying bulk traffic. Note that the bottleneck link is
with a capacity 1Mbps. All the other
links have a capacity of 50Mbps. Each link has a propagation delay of and a buffer size of 50 packets. We use TCP Reno
with a window size of 20 packets. The PAM, F-SAM and the CAM markers are configured with a total # of the bottleneck
link bandwidth. Note that in all the graphs, we compare the performance of our scheme with that of standard RED AQM policy.
The metrics used to evaluate our results are end-to-end goodput and a max-min fairness index. Suppose there are
flows and
each flow
has a goodput of
8 . Then, the fairness index [6] # for the set of flows is given by the following equation:
# 8 8 % 8 8 (2)
11
For the best possible allocation, # . This means that each flow gets equal bandwidth. For a very unfair allocation # is close
to
@ . Unless stated, we compare the performance of our various AQM-markers and RIO queueing approach with the current RED
based dropping at bottleneck links, without any marking.
First we conduct some experiments without any traffic originating from nodes
to
. That is, all traffic aggregates at edge
. The first four set of experiments are performed under this topology. This topology represents the case where most of the
bottleneck traffic (including misbehaving flows) comes from one edge domain. In this case we show that our fair aggregate
marking approximates fair bandwidth allocation by around 80 to 90% (in terms of fairness indices) and ensures that misbehaving
flows get checked. Then we consider the more general scenario where the bottleneck router is congested by flows from multiple
edge domains. In this scenario, TCP traffic and misbehaving flows originates from nodes
to
too, which aggregates at edge
@ , in addition to the earlier traffic aggregate from . We therefore apply our fair AQM-marking at two edge markers separately.
In this more general scenario, even though our fair aggregate marking does not translate into a globally fair token marking
allocation, we still show that the effect of misbehaving flows at bottleneck routers get dampened considerably (by as much as as
a factor of 10 for our CAM and F-SAM markers). Finally, we consider a scenario where the edge routers have information about
the bottleneck bandwidth and the approximate number of flows at the congested router and at its own input queue. We show that
if we tune the target marking rate using this information, we obtain a much more globally fair bandwidth distribution than in the
previous experiments.
B. All CBR Traffic
We first study the end-to-end performance of a collection of CBR flows over UDP when some of them are pushing traffic at a
rate much higher than the rest of the group. This is the most simple, open-loop test of the efficacy of our architecture. We consider
the case when everyone is pushing traffic faster than their fair share would permit. We take a varying number of CBR flows each
offering 2Mbps traffic into a bottleneck link with a capacity of 1Mbps. Out of these some of them (the misbehaving flows) have
10Mbps offered load.
The results are depicted in Figure 3. The first column shows the results for scenarios with 1 misbehaving flow in the group,
the second column depicts the results for 2 misbehaving flows and so on. The first row shows the standard deviation (the average
being the same) of the bandwidth allocation among the UDP flows. The second row calculates the fairness index. As we can
see from the figures, our architecture yields much superior performance compared to RED. The standard deviation of CBR flow
throughputs in our scheme with CAM and F-SAM is 300% less than for RED in Figure 3 (i), (ii) and (iii). Note that the results
of PAM is somewhat similar to those of RED as the marker uses a RED-like algorithm to distribute token. On the other hand,
both CAM and F-SAM distribute the tokens in a fair manner and the resultant end-to-end performance achieves an approximate
12
6 8 10 12
Number of UDP flows
0
50000
1e+05
1.5e+05
Stddev from fair share (bps)
RED
PAM
F-SAM
CAM
6 8 10 12
Number of UDP flows
0
50000
1e+05
Stddev from fair share (bps)
RED
PAM
F-SAM
CAM
6 8 10 12
Number of UDP flows
0
20000
40000
60000
80000
Stddev from fair share (bps)
RED
PAM
F-SAM
CAM
(i) (ii) (iii)
6 8 10 12
Number of UDP Flows
0.5
0.6
0.7
0.8
0.9
Fairness Index
RED
PAM
F-SAM
CAM
6 8 10 12
Number of UDP Flows
0.5
0.6
0.7
0.8
0.9
1
Fairness Index
RED
PAM
F-SAM
CAM
6 8 10 12
Number of UDP Flows
0.5
0.6
0.7
0.8
0.9
1
Fairness Index
RED
PAM
F-SAM
CAM
(iv) (v) (vi)
with 1 misbehaving flow with 2 misbehaving flows with 3 misbehaving flows
Fig. 3. Comparison of the fairness in goodput among UDP flows in scenarios with only UDP flows. (i), (ii), (iii) depict the standard deviation from the fair share
while (iv), (v), (vi) depict fairness indices as number of flows increases
fair bandwidth allocation. The fairness index plots in Figure 3 (iv), (v) and (vi) show that as the number of misbehaving flow
increases, the performance of our CAM and F-SAM degrades slightly. In CAM, for example, it is due to the limited buffer to keep
track of the victims and criminals. Note that the end-to-end performance is heavily influenced by the fair marking of the packets
at the edge.
C. Fair Bandwidth Allocation with TCP flows
In this section, we investigate how our markers behave in scenarios with several TCP flows and a few misbehaving flows. In
Figure 4 (i), the offered load of the UDP is 1Mbps. The TCP flows get more than three times the goodput with CAM and F-SAM
marking than with RED. Then we increase the malicious traffic to 3 CBR flows each pushing 0.5Mbps. We notice that the TCP
performance with our fair marking architecture becomes two orders of magnitude better than with RED as depicted in Figure 4
(ii). This is because, with higher UDP load, TCP connections get timed out due to excess packet losses in RED. However, both
CAM and F-SAM can mark the packets originating from the misbehaving flow as OUT with a greater probability. Thus, the TCP
13
flows get their fair share of IN packets. Hence only a few of their packets get downgraded and dropped on congestion. On the other
hand, most packets of the CBR flows exceeding the fair share gets downgraded to OUT. Thus the OUT queue is predominated by
the packets from these misbehaving flows. On congestion, these packets get dropped.
0 20 40 60 80
Number of TCP flows
0
10000
20000
30000
40000
50000
60000
70000
Average bandwidth of TCP flows (bps)
RED
PAM
F-SAM
CAM
0 20 40 60 80
Number of TCP flows
0
20000
40000
60000
80000
Average bandwidth of TCP flows (bps)
RED
PAM
F-SAM
CAM
(i) 1 misbehaving flow (ii) 3 misbehaving flows
Fig. 4. Effect of our marking based architecture in scenarios with several TCP flows and a few misbehaving UDP flows: Performance of TCP flows
0 10 20 30 40
Flows
1000
10000
1e+05
1e+06
Per-flow Goodput
RED
PAM
F-SAM
CAM
0 10 20 30 40
Flows
100
1000
10000
1e+05
1e+06
Per-flow Goodput
RED
PAM
F-SAM
CAM
(i) 1 misbehaving flow (ii) 3 misbehaving flows
Fig. 5. Effect of our marking based architecture in scenarios with several TCP flows and a few misbehaving UDP flows: Distribution of bandwidth among all the
flows
The drastic improvement in TCP goodput with our architecture suggests that AQM-based marking with RIO can contain the
impact of misbehaving non-responsive flows. This is further justified in Figure 5. The two graphs show the individual bandwidth
obtained by every flow under different marking schemes, when the number of misbehaving flows are one and three respectively.
Note that we use a logarithmic scale in the y-axis. These charts clearly show that CAM (and F-SAM to a lesser extent) ensure
14
that all flows get almost equal bandwidth, thus making the allocation fair. Note that using our scheme, we can effectively clamp
the CBR traffic, thus improving the TCP performance. It is interesting to note that our results are as good as those mentioned
in [24]. We have used an equivalent simulation setup with similar parameters. This enforces our claim that our architecture can
be as effective as the complex AQM strategies that can be implemented in routers. Note that we get good results with a buffer size
of only 50 packets at our marker, while the CHOKe paper [24] mentions a buffer size of 300 packets at the bottleneck link.
D. Mice and Elephants
0 20 40 60 80
Number of elephant TCP flows
0
20000
40000
60000
80000
Average goodput of elephants (bps)
RED
PAM
F-SAM
CAM
Fig. 6. Effect of our architecture on TCP flows. In this figure, we plot the average goodput of TCP flows.
Since short term traffic, mice, constitutes most of the connections in the Internet [18], it is important to check whether our
architecture can take care of mixed scenarios of both mice and elephants (long lived TCP flows). We took the web-traffic model
of ns-2 and we ensured that the total amount of short flow traffic was, on an average, 12-16% of the total bottleneck link capacity.
Then we changed the number of elephants in the scenario. We plot the average goodput of the long lived TCP flows in Figure 6.
We see that our scheme does not alter the bandwidth of the long term TCP flows more than 2%. Since the toal utilization of link
was almost 100%, its is clear, that the aggregate short flow traffic does not get affected by our marking, followed by RIO dropping.
E. Dampening Effect
As we mentioned in Section III, a locally fair allocation of marked IN packets do not directly yield a globally fair allocation,
when bottleneck congestion is caused by flows from multiple domains. In this experiment, we introduce 20 new TCP flows
aggregated at node flowing into the bottleneck link. These flows are not marked. We simultaneously vary the number of TCP
flows aggregated at node in the presence of a single UDP flow with a rate of 2Mbps. Figures 7(i),(ii),(ii) depict the average
goodput of TCP flows, total goodput of TCP flows and the total goodput of UDP flows respectively. We can see, in Figure 7(ii),
15
0 20 40 60 80 100
Number of TCP flows
0
5000
10000
15000
Avg. Goodput of TCP flows (bps)
RED
PAM
F-SAM
CAM
0 20 40 60 80 100
Number of TCP flows
0
1e+05
2e+05
3e+05
4e+05
5e+05
6e+05
7e+05
Total goodput of all TCP flows
RED
PAM
F-SAM
CAM
0 20 40 60 80 100
Number of TCP flows
0
2e+05
4e+05
6e+05
8e+05
1e+06
Goodput of the misbehaving UDP flow (bps)
RED
PAM
F-SAM
CAM
(i) Avg. goodput of TCP (ii) Total goodput of TCP (iii) Goodput of misbehaving flow.
Fig. 7. Our architecture helps to dampen the misbehaving flows. In this scenario, we have 20 TCP flows from , and a variable number of TCP and 1
misbehaving flow from . Only enables packet marking.
that with our markers, the total TCP goodput keeps on increasing as the number of flows increase. At the same time, we can see
that the corresponding goodput of the misbehaving flow decreases as we increase the number of flows. On the other hand, with
RED, the total goodput of the TCP flows hardly change with increasing number of TCP flows.
Now, we increase the flows aggregating at to 32 and introduce another misbehaving UDP flow of 2Mbps from edge . We
mark this new edge aggregate with the same type of marker as in and perform the above experiment. As shown in Figure 8,
the improvements in TCP goodput due to our AQM-marker based approach are even more dramatic than those in Figure 7. The
total misbehaving flow throughput is dampened by as much as a factor of 10 using our CAM marking and a factor of 7 using
our F-SAM marking, when compared to the normal RED queueing case. Note that Figure 7and Figure 8 shows that with RED,
the bottleneck link is fully saturated with the UDP flows thus squeezing out TCP. On the other hand, CAM and F-SAM restrict
the UDP flows from getting more IN tokens than their local fair share. This ensures that they are preferentially dropped at the
congested router. We note that our results do not degrade even on increasing the number of TCP flows from 20 to 32 at .
We should note that with pure fair queueing, each flow should have obtained its max-min fair share, i.e.
, where
% is the
variable number of flows originating from edge and
is the number of fixed flows at . Clearly we can see that even CAM
does not provide the exact fair share. For example, for the 12 flows case, the fair share is
@ @ . From Figure 7(i)
we see that CAM, F-SAM yield only 12381, 13731 bps respectively. This is because our marker at & ensures fairness among the
12 flows aggregating at & and not among all flows in the system. Thus the misbehaving UDP flow can only get partially damped.
On the other hand, consider the 92 flows case, the fair share is
@ . From Figure 7(i), we see that CAM gives us an
average bandwidth of 7670bps while F-SAM yields 5482bps. The results are better in this case because the our marker performs
fairness among 92 flows. Thus this clearly shows that even though we cannot hope to approximate global max-min fairness in all
16
0 20 40 60 80 100
Number of TCP flows
0
5000
10000
15000
20000
25000
30000
Avg. Goodput of TCP flows
RED
PAM
F-SAM
CAM
0 20 40 60 80 100
Number of TCP flows
0
2e+05
4e+05
6e+05
8e+05
Total Goodput of TCP flows
RED
PAM
F-SAM
CAM
0 20 40 60 80 100
Number of TCP flows
0
2e+05
4e+05
6e+05
8e+05
1e+06
Total Goodput of misbehaving UDP flows
RED
PAM
F-SAM
CAM
(i) Avg. goodput of TCP (ii) Total goodput of TCP (iii) Goodput of misbehaving flow.
Fig. 8. Our architecture helps to dampen the UDP flows. In this scenario, we have 50 flows from and 1 misbehaving UDP flow, and a variable number of TCP
and 1 misbehaving flow from . Both the edges use packet marking.
0 5 10 15 20 25
Number of flows (N)
0
20000
40000
60000
80000
Avg. bandwidth of TCP flows (bps)
RED
PAM
F-SAM
CAM
0 5 10 15 20 25
Number of flows (N)
0
1e+05
2e+05
3e+05
4e+05
5e+05
6e+05
Bandwidth of UDP flow from edge 1 or node e0 (bps)
RED
PAM
F-SAM
CAM
0 5 10 15 20 25
Number of flows (N)
0
1e+05
2e+05
3e+05
4e+05
5e+05
6e+05
7e+05
Bandwidth of UDP flow from edge 2 or node e1 (bps)
RED
PAM
F-SAM
CAM
(i) Avg. goodput of TCP (ii) Goodput of misbehaving flow from (iii) Goodput of misbehaving flow from .
Fig. 9. Performance of our marker with approximate information about the bottleneck. There are TCP flows and 1 misbehaving flow from and TCP
flows and 1 misbehaving flow from scenarios, we can provide adequate protection from misbehaving flows at the congested router, using our AQM-based marking
scheme.
F . Dampening Effect with Additional Information
In this section, we analyze the two edge scenario with a very different assumption. Suppose every router could calculate the
approximate number of flows and congestion indication, and propagate this information to all the edges. Can we tune the target
marking rates and provide a more vigorous damping of the misbehaving flows? We show that with this extra information, we can
indeed restrict the misbehaving flows to its global fair share. In this context global fair share is defined as the fair share of a flow
in the subgraph induced on the network topology by the congested router, the edge routers that access it, and all the intermediate
routers that lead to the congested router.
17
We consider the same topology as in Figure 2 and we call the node @ as edge1 and the node as edge2 for convenience. We
have
= ,
% = bulk TCP flows at edge1, edge2 respectively. Apart from this, each edge has a node pushing misbehaving
traffic in the form of CBR packets over UDP at 0.5Mbps. The total traffic at the bottleneck is 1Mbps and is therefore congested
by
% flows. Thus the bottleneck link is always congested. We fix
and vary the
% from to
& @ .
If we assume that the edge routers know the bottleneck bandwidth ( 1Mbps in our case) and the value of
% , we can set
the target CIRs of the markers to
and
respectively. Then we plot the average goodput obtained by
the TCP flows in Figure 9 (i), the bandwidth obtained by the two UDP flows in Figures 9 (ii) and (iii). Clearly, both F-SAM and
CAM damp both the misbehaving flows almost equally even though the number of flows at each edge is very different (
% can
be four times
). The TCP flows get significant bandwidth in spite of the presence of the misbehaving flows. This is because
our fair markers equally divide the IN tokens among flows from its own domain, and in this case, they have the correct target rate
to assure fairness among flows in in the other domain. Compare this result with the previous results in Section V-E where the
edges did not have any information. Clearly the results with this new information demonstrate a very big improvement over that
in the previous section. However, we need to maintain the approximate number of flows at the routers, and we need a scheme to
feedback this information to the edge routers. We are currently investigating this and we discuss it in more detail in Section VI.
G. An Informal Game Theoretic Analysis
In this section, we present an intuitive game-theoretic argument in support of our architecture. We can model our problem using
non-cooperative game theory [21]. We assume that misbehaving traffic agents are players, and traffic markers and AQM schemes
determine the rules of the game. Every user has a strategy that yields a utility; for misbehaving users, this could mean goodput,
for example. All that a player cares is the maximization of his own utility. In non-cooperative game theory, Nash equilibrium is a
fundamental game-theoretic solution concept. In our model, a system is at Nash equilibrium if there is no incentive for any selfish
agents to unilaterally deviate from the current strategy. Now, we can consider the misbehaving flow to be a selfish agent. Thus the
problem of controlling misbehaving flows can be considered to be equivalent to the existence of Nash equilibria of a game dictated
by a given AQM scheme. Akkela et.al. [1] and Dutta et. al. [8] empirically show that RED does not impose a Nash equilibrium on
selfish agents as defined in the paper. It is easy to show that for schemes like fair queueing, Nash equilibria exist. Thus, we need
to ask whether Nash equilibria exist with approximate fair marking at the edges followed by RED/RIO-like dropping at the core.
The answer to the above question is an emphatic “yes” if we assume that at any given domain, there are several flows aggre-
gating at a marker. At some level of aggregation, our fair markers, CAM or F-SAM, will fairly distribute IN token among flows
in that aggregate. Thus, the total number of IN tokens allocated to the misbehaving flow remains invariant irrespective of the
(selfish) offered load of this flow. At the RIO queue, the OUT packets will get preferentially dropped. This implies that there
18
is no incentive for any misbehaving flow to push in traffic in a greedy fashion. Thus Nash equilibrium exists, and, hence, our
architecture can control misbehaving flows to the degree of fairness allowed by the number of flows at the edge domain where
this misbehaving flow originates.
VI. DISCUSSION, CONCLUSION AND FUTURE DIRECTIONS
In this paper, we introduced a lightweight architecture that can contain misbehaving flows at a congested router by marking
packets at the edges using stateless aggregate fair markers followed by simple RIO dropping at the congested router. The design
of these markers have been inspired by approximately fair AQM techniques like CSFQ and CHOKe. We are not aware of any
other work that uses AQM-based marking to achieve the above goal. We show that this seemingly simple architecture can
punish misbehaving flows effectively, and improve the throughput of well-behaved TCP flows dramatically, in the presence of
misbehaving flows. We also show that in scenarios with only UDP flows, our scheme helps to divide bandwidth among the UDP
flows more fairly. Finally, we demonstrate that our architecture does not harm short lived mice in the presence of elephants (long
lived TCP flows).
Using detailed packet level simulation, we observe that our architecture can improve TCP performance by an order of magnitude
in the presence of misbehaving flows. We show that with certain topologies, we can actually approximate fairness in bandwidth
allocation at congested routers and achieve a fairness index of around 0.8-0.9. In general, our scheme can curtail the bandwidth
allocated to misbehaving flows by an order of magnitude. We also show, via an informal game theoretic analysis, that it is indeed
possible to control misbehaving flows using our architecture.
This paper demonstrates the novel idea of marking packets by adaptation of known AQM techniques, followed by preferential
RIO dropping at congested routers, can approximate the results provided by sophisticated per-router AQM techniques (at least in
simple topologies). For example, we illustrate that using our CAM marker, it might be possible to yield performance comparable
to the implementation of the real CHOKe AQM at core routers, in certain scenarios. The advantage of our marking based scheme
is that the core routers do not need to be modified. We were also pleased by the fact that we could leverage previously proven
AQM techniques to design better traffic markers.
Our scheme raises the following question: whether it is really necessary to attain highly accurate levels of fair bandwidth
allocation among flows in the Internet. Or can one compromise on the level of fairness, and develop simple lightweight schemes
which cannot provide exact max-min fairness but can severely dampen misbehaving flows.
There are other directions in this area that we are currently working on. First and foremost, we are investigating as to how
we can configure the correct token bucket rate at each edge router dynamically. In the previous section, we show that if each
edge router has information(either statically configured or dynamically obtained) about the ratio of the number of flows passing
19
through it versus the number of flows at the bottleneck, it can use this ratio to calculate the ideal token bucket rate for its marker.
This can maximize the fairness in bandwidth allocation under our scheme, and indeed obtains surprisingly good results in terms
of fairer throughput for each flow. However this information does not seem to be obtainable at edge routers using completely
stateless schemes. One possible way would be for a bottleneck router, at the onset of congestion, to transmit a sample of its packet
queue to all edge routers, so that they can obtain the ratio of the number of flows using some form of random sampling. We are
currently working on the mechanisms and algorithms to dynamically obtain this information at edge routers, using a minimum
amount of state. We are also looking at the possibility of modeling this problem using control theory. In [3] Chait et. al. used a
control theoretic scheme to configure the target marking rates for individual TCP flows but they do not handle unresponsive flows
and maintain per-flow state.
Another direction of future work is to adapt other AQM algorithms like BLUE, SRED and RED-PD for our aggregate fair
markers, and also develop analytical bounds on the fairness guarantees which our marking scheme can provide as compared to
actual fair queueing. We also intend to study the effect of different traffic flow distributions and edge/core domain topologies on
the fairness levels, which our schemes provide.
A very important and exciting research question that is a natural extension of this work can be stated as follows: Can we design a
lightweight, stateless, edge-marking based architecture that can provide global fairness among different flows from different edge
domains without making significant modifications to core routers? This is equivalent to developing a scheme that can result in
globally fair allocations using local computations. This problem is hard when we restrict ourselves to stateless schemes. However
we are looking at relaxing the stateless assumption, and find out an ideal balance between the accuracy of fairness results a scheme
can obtain and the amount of state required by the scheme.
REFERENCES
[1] A. Akella, R. Karp, C. Papadimitrou, S. Seshan, and S. Shenker. Selfish behavior and stability of the internet: A game-theoretic analysis of TCP. SIGCOMM,
2002 (to appear).
[2] S. Blake, D. Black, M. Carlson, E. Davies, Z. Wang, and W. Weiss. An architecture for differentiated services. RFC 2475, 1998.
[3] Y . Chait, C. Hollot, V . Misra, D. Towsley, H. Zhang, and J. Lui. Providing throughput differentiation for TCP flows using adaptive two color marking and
multi-level AQM. IEEE INFOCOM, 2002.
[4] D. D. Clark and W. Fang. Explicit allocation of best-effort packet delivery service. IEEE/ACM Transactions on Networking, 6(4):362–373, 1998.
[5] A. Das, D. Dutta, and A. Helmy. ”fair stateless aggregate marking techniques using active queue management techniques”. In IEEE/IFIP MMNS, October
2002.
[6] Bruce S. Davie and Larry L. Peterson. Computer Networks: A Systems Approach. Morgan Kauffmann, second edition, 2000.
[7] A. Demers, S. Keshav, and S.J. Shenker. Analysis and simulation of a fair queueing algorithm. Sigcomm, 1989.
[8] D. Dutta, A. Goel, and J. Heidemann. Oblivious AQM and Nash Equilibria. IEEE INFOCOM (to appear), April 2003.
[9] B. Braden et al. Recommendations on queue management and congestion avoidance in the internet. RFC 2309, 1998.
20
[10] A. Feroz, A. Rao, and S. Kalyanaraman. A tcp-friendly traffic marker for ip differentiated services. Proc. of the IEEE/IFIP Eighth International Workshop
on Quality of Service - IWQoS, 2000.
[11] S. Floyd and K. Fall. Promoting the use of end-to-end congestion control in the Internet. IEEE/ACM Transactions on Networking, 7(4):458–472, 1999.
[12] S. Floyd and V . Jacobson. Random early detection gateways for congestion avoidance. IEEE/ACM Transactions on Networking, V .1 N.4, 1993.
[13] J. Heinanen and R. Guerin. A single rate three color marker. RFC 2697, 1999.
[14] J. Heinanen and R. Guerin. A two rate three color marker. RFC 2698, 1999.
[15] L. Wood I. Andrikopoulos and G. Pavlou. A fair traffic conditioner for the assured service in a differentiated services internet. Proceedings of ICC 2000,
vol. 2 pp. 806-810, 2000.
[16] V . Jacobson. Congestion avoidance and control. ACM Computer Communication Review; Proceedings of the Sigcomm ’88 Symposium in Stanford, CA,
August, 1988, 18, 4:314–329, 1988.
[17] D. Lin and R. Morris. Dynamics of random early detection. SIGCOMM ’97, pages 127–137, september 1997.
[18] R. Mahajan, S. Floyd, and D. Wetherall. Controlling high-bandwidth flows at the congested router, 2001.
[19] J. Nagle. Congestion control in ip/tcp internetworks, 1984.
[20] K. Nichols, V . Jacobson, and L. Zhang. A twobit differentiated services architecture for the internet. RFC 2638, 1999.
[21] M. J. Osborne and A. Rubenstein. A Course in Game Theory. The MIT Press, Cambridge, Massachusetts, 1994.
[22] T. J. Ott, T. V . Lakshman, and L. H. Wong. SRED: Stabilized RED. In Proceedings of INFOCOM, volume 3, pages 1346–1355, 1999.
[23] R. Pan, B. Prabhakar, and K. Psounis. ”CHOKe - A stateless queue management scheme for approximating fair bandwidth allocation”. In INFOCOM, 2000.
[24] R. Pan, B. Prabhakar, and K. Psounis. A stateless active queue management scheme for approximating fair bandwidth allocation. IEEE INFOCOM 2000,
2000.
[25] J. Postel. Transmission control protocol. Internet Request for Comments RFC 793, September 1981.
[26] I. Stoica, S. Shenker, and H. Zhang. Core-stateless fair queueing: Achieving approximately fair bandwidth allocations in high speed networks. Sigcomm,
1998.
[27] H. Su and Mohammed Atiquzzaman. Itswtcm: A new aggregate marker to improve fairness in difserv. Globecomm, 2001.
[28] Cisco Systems. Technical specification from cisco, weighted random early detection on the cisco 12000 series router. March 2002.
[29] UCB/LBNL/VINT. The NS2 network simulator, available at http://www.isi.edu/nsnam/ns/.
[30] I. Yeom and A. L. Narasimha Reddy. Adaptive marking for aggregated flows. Globecomm, 2001.
[31] L. Zhang, S. Deering, and D. Estrin. RSVP: A new resource ReSerVation protocol. IEEE network, 7(5):8–?, September 1993.
Linked assets
Computer Science Technical Report Archive
Conceptually similar
PDF
USC Computer Science Technical Reports, no. 752 (2002)
PDF
USC Computer Science Technical Reports, no. 877 (2006)
PDF
USC Computer Science Technical Reports, no. 831 (2004)
PDF
USC Computer Science Technical Reports, no. 753 (2002)
PDF
USC Computer Science Technical Reports, no. 758 (2002)
PDF
USC Computer Science Technical Reports, no. 764 (2002)
PDF
USC Computer Science Technical Reports, no. 755 (2002)
PDF
USC Computer Science Technical Reports, no. 811 (2003)
PDF
USC Computer Science Technical Reports, no. 757 (2002)
PDF
USC Computer Science Technical Reports, no. 747 (2001)
PDF
USC Computer Science Technical Reports, no. 769 (2002)
PDF
USC Computer Science Technical Reports, no. 779 (2002)
PDF
USC Computer Science Technical Reports, no. 674 (1998)
PDF
USC Computer Science Technical Reports, no. 770 (2002)
PDF
USC Computer Science Technical Reports, no. 772 (2002)
PDF
USC Computer Science Technical Reports, no. 644 (1997)
PDF
USC Computer Science Technical Reports, no. 781 (2002)
PDF
USC Computer Science Technical Reports, no. 726 (2000)
PDF
USC Computer Science Technical Reports, no. 857 (2005)
PDF
USC Computer Science Technical Reports, no. 657 (1997)
Description
Abhimanyu Das, Debojyoti Dutta, Ahmed Helmy. "A lightweight mechanism to control misbehaving flows using AQM-based aggregate marking." Computer Science Technical Reports (Los Angeles, California, USA: University of Southern California. Department of Computer Science) no. 763 (2002).
Asset Metadata
Creator
Das, Abhimanyu
(author),
Dutta, Debojyoti
(author),
Helmy, Ahmed
(author)
Core Title
USC Computer Science Technical Reports, no. 763 (2002)
Alternative Title
A lightweight mechanism to control misbehaving flows using AQM-based aggregate marking (
title
)
Publisher
Department of Computer Science,USC Viterbi School of Engineering, University of Southern California, 3650 McClintock Avenue, Los Angeles, California, 90089, USA
(publisher)
Tag
OAI-PMH Harvest
Format
20 pages
(extent),
technical reports
(aat)
Language
English
Unique identifier
UC16270880
Identifier
02-763 A Lightweight Mechanism to Control Misbehaving Flows using AQM-based Aggregate Marking (filename)
Legacy Identifier
usc-cstr-02-763
Format
20 pages (extent),technical reports (aat)
Rights
Department of Computer Science (University of Southern California) and the author(s).
Internet Media Type
application/pdf
Copyright
In copyright - Non-commercial use permitted (https://rightsstatements.org/vocab/InC-NC/1.0/
Source
20180426-rozan-cstechreports-shoaf
(batch),
Computer Science Technical Report Archive
(collection),
University of Southern California. Department of Computer Science. Technical Reports
(series)
Access Conditions
The author(s) retain rights to their work according to U.S. copyright law. Electronic access is being provided by the USC Libraries, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright.
Repository Name
USC Viterbi School of Engineering Department of Computer Science
Repository Location
Department of Computer Science. USC Viterbi School of Engineering. Los Angeles\, CA\, 90089
Repository Email
csdept@usc.edu
Inherited Values
Title
Computer Science Technical Report Archive
Description
Archive of computer science technical reports published by the USC Department of Computer Science from 1991 - 2017.
Coverage Temporal
1991/2017
Repository Email
csdept@usc.edu
Repository Name
USC Viterbi School of Engineering Department of Computer Science
Repository Location
Department of Computer Science. USC Viterbi School of Engineering. Los Angeles\, CA\, 90089
Publisher
Department of Computer Science,USC Viterbi School of Engineering, University of Southern California, 3650 McClintock Avenue, Los Angeles, California, 90089, USA
(publisher)
Copyright
In copyright - Non-commercial use permitted (https://rightsstatements.org/vocab/InC-NC/1.0/