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. 638 (1996)
(USC DC Other)
USC Computer Science Technical Reports, no. 638 (1996)
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
Multicast F eedbac k Suppression Using
Represen tativ es
Dan te DeLucia
Computer Science Departmen t
Univ ersit y of Southern California
Los Angeles CA and
Hughes Researc h Lab oratories
Malibu Can y on Road
Malibu CA v oicefax email dan teuscedu
Katia Obraczk a
Univ ersit y of Southern California
Information Sciences Institute
Admiralt yW a y Suite Marina Del Rey CA
v oice fax email k atiaisiedu
August Abstract
F or a reliable o wcon trolled m ulticast transp ort proto col to scale it
m ust a v oid the feedbac k implosion problem particularly if the pro
to col targets arbitrarily large m ulticast groups comm unicating o v er lossy
net w orks
Most existing feedbac k con trol mec hanisms based on probabilistic sup
pression address the feedbac k implosion problem b y suppressing feedbac k
using timers based on roundtrip time R TT information This approac h
requires that all receiv ers compute R TT to the data source
W e presen t an algorithm whose ma jor b enet deriv es from the fact
that it do es not need to compute R TT from receiv ers to the source and
do es not require kno wledge of group mem b ership or net w ork top ology W e
use a small set of r epr esentative receiv ers and probabilistic suppression to
limit feedbac k
W e b eliev e that our approac h will p erform w ell in real net w orks Sim u
lations using randomlygenerated net w ork top ologies of v arying sizes with
p essimistic net w ork loss rates sho w that represen tativ es considerably re
duce the amoun t of feedbac k compared to a purely suppressionbased
sc heme F or v arious m ulticast group sizes a few represen tativ es can k eep
the amoun t of feedbacklo w while not degrading feedbac k timeliness
Keyw ords Multicast feedbac k suppression represen tativ es
In tro duction
While man y existing m ulticast transp ort proto cols suc h as SRM and R TP
target dela y sensitiv e realtime applications little has b een done to address
m ultip oin t bulk data transfer services suc hasm ulticast FTP The feedbac k suppression algorithm w e presen t is part of our eorts to
build a scalable o wcon trolled m ulticast transp ort mec hanism suitable for bulk
data transfer applications Our solution addresses the problem of feedbac k
implosion while pro viding the frequen t and timely feedbac k required b y a
o w con trol algorithm Our solution do es not require kno wledge of the m ulti
cast group membershipornet w ork supp ort
Sev eral existing reliable m ulticast transp ort proto cols suc h as SRM use
R TTbased probabilistic suppression to a v oid feedbac k implosion at the source
In R TTbased suppression eac h receiv er measures R TT to the source and sets a
feedbac k timer based on its distance from the source Receiv ers nearer the source
will b e the rst to resp ond to a pac k et loss and will tend to suppress NA CKs
from receiv ers farther do wn the m ulticast tree This approachis in tended for
conferencing applications in whic h man ytoman ycomm unication tak es place
In this en vironmen t computing R TTs to the source requires little additional
o v erhead
In this pap er w e presen t a solution to the feedbac k implosion problem that
do es not rely on R TT estimation and hence has greater p oten tial for scalabilit y Although our mec hanism is not tied to an y particular application it is aimed at
applications with a single data source and m ultiple receiv ers All comm unication
is via m ulticast
W e use a small set of group mem b ers or r epr esentatives to represen t the con
gested m ulticast subtrees In addition to pro viding fast feedbac k to the source
represen tativ es suppress feedbac k from the other group mem b ers Through
sim ulations w e ev aluate our feedbac k suppression mec hanism in terms of the
amoun t and timeliness of the feedbac k generated bythe m ulticast group
F or m ulticast groups of v arying sizes comm unicating o v er randomlygenerated
net w ork top ologies sub ject to p essimistic loss rates our sim ulation results sho w
that represen tativ es can greatly reduce the amoun t of feedbac k when compared
to purely suppressiv e algorithms y et still pro vide timely feedbac k to the source
Represen tativ es also mak e the algorithm less sensitiv e to suppression timer set
tings whic h allo ws the algorithm to p erform w ell without ha ving to rely on
R TT estimation bet w een eac h receiv er and the source Not p erforming R TT
estimates cuts do wn the o v erhead of our feedbac k con trol mec hanism consid
erably Our sim ulations also sho w that represen tativ e set size scales w ell with
m ulticast group size
Related W ork
F eedbac k Con trol
T raditional reliable unicast transp ort proto cols suc h as TCP use A CKs to re
co v er from pac k et loss This approachto ac hieving reliabilit y is often referred
to as senderinitiate d since it is the resp onsibilit y of the sender to detect pac k et
losses In a m ulticast en vironmen t as group sizes increase the senderinitiated
sc heme ma y cause A CK implosion since eac h deliv ered pac k et triggers an ac
kno wledgmentfromev ery receiv er in the group
Alternativ elyin the r e c eiverinitiate d approac h to reliabilit y receiv ers detect
pac k et losses and request its retransmission b y generating a NA CK Placing the
resp onsibilit y of reco v ering from pac k et losses on the receiv er helps alleviate
the A CK implosion problem The p erformance comparison study presen ted in
conrms that receiv erinitiated m ulticast transp ort proto cols ha v e b etter
scalabilit y prop o erties than their senderinitiated coun terparts F or this reason
most reliable m ultip oin t transp ort proto cols are either pure receiv erinitiated or
use a h ybrid approachbycom bining receiv er and senderinitiated reliabilit y Ho w ev er receiv erbased proto cols also suer from the feedbac k implosion
problem esp ecially when losses o ccur higher up in the m ulticast tree in larger
groups o v er lossy net w orks In this section w e fo cus on prop osed solutions to
the feedbac k implosion problem in the con text of reliable m ulticast transp ort
proto cols
In solutions to the feedbac k implosion problem are classied as structur e
b ase d or timerb ase d Structurebased approac hes suc h as the LogBased Pro
to col rely on a designated site either a dedicated serv er in the case of the
LogBased Proto col or a preassigned group mem b er to pro cess and lter
feedbac k information
Timerbased solutions rely on probabilistic feedbac k suppression to a v oid im
plosion at the source Receiv ers in the SRM proto col whic hw as designed to
supp ort the WB distributed whiteb oard application dela y their retransmission
requests for a random in terv al uniformly distributed b et w een the curren t time
and the onew a y trip time to the source The goal is that group mem b ers closer
to the source send their feedbac k so oner suppressing feedbac k from farther a w a y
mem b ers A site uses p erio dic session messages to measure its distance based
on the resulting R TT to the other group mem b ers
The Deterministic Timeouts for Reliable Multicast DTRM algorithm
also uses R TT b et w een receiv ers and the sender to compute the receiv ers sup
pression timeouts
The feedbac k con trol mec hanism prop osed in do es not fall in to either
the structurebased or the timerbased categories In this approac h to feedbac k
con trol whic h is used b y the IVS video conferencing to ol and is la y ered atop
of R TP video sources use probabilistic p olling to select a set of receiv ers
that should pro vide feedbac k
Windo wBased v ersus RateBased Flo w Con trol
Weha v e considered t w o approac hes to o w con trol The rst is the traditional
windo wbased o w con trol sc heme used b y TCP The second approac h is the
ratebased sc heme suc h as that used in NETBLT Windo wBased Flo w Con trol
A more traditional approachto o w con trol is allo wing the sender to transmit
a certain n um b er of pac k ets or a windo w at a time The sender adv ances its
windoweac h time it receiv es an ac kno wledgmen t for an outstanding pac k et The
problem of using a windo wbased o w con trol sc heme in a m ulticast en vironmen t
is deciding ho w to adjust the transmission windo w
Because of the A CK implosion problem m ulticast transp ort proto cols use
negativeac kno wledgmen t NA CKs to signal pac k et loss Should the windo w
b e adjusted in the absence of a NA CK Ho w long do wew ait for NA CKs W e
could measure the n um ber of NA CKs coming bac k but since w e do not kno w
the size of the group coun ting NA CKs is not v ery useful
Another conceptual problem is that a windo wbased o wcon trol mec hanism
is closely tied to reliabilit y There has b een a n um ber of argumen ts in the
literature for decoupling o w con trol from reliabilit y F or these reasons w e
decided to use a ratebased o w con trol sc heme
RateBased Flo wCon trol
In ratebased o w con trol the transmission rate can be set indep enden tly of
the reliabilitymec hanism The source sends pac k ets according to the curren t
transmission rate Since our goal is to a v oid pac k et loss the receiv ers should
determine ho w m uc h queuing is taking place in the net w ork If there is to o
m uc h data queued in the net w ork receiv ers need to notify the source to slo w
do wn
Since pac k et losses ma y still o ccur the o w con trol mec hanism should re
sp ond b y m ultiplicativ ely decreasing the transmission rate when NA CKs are
receiv ed at the sender
Congestion Con trol
In Nagle sho ws that congestion can o ccur ev en in a datagram net w ork
with innite storage Jain et al prop oses a congestion a v oidance sc heme
where routers signal congestion b y setting the c ongestion avoidanc e bit in the
pac k ets net w orkla y er header Dep ending on the o v erall feedbac k receiv ed
sources decide whether to increase or increase the curren t windo w size Jains
net w orkla y er congestion a v oidance sc heme requires a new bit in the pac k et
headers as w ell as routers b eing able to set this bit
In Jacobson describ es his slo wstart o w con trol algorithm for TCP whic h gradually op ens the TCP transmission windo w as the source receiv es
ac kno wledgmen ts from the receiv ers Slo wstart uses data loss as sign of con
gestion and sh uts the windowdo wn to pac k et after a pac k et is lost Both the
T aho e and Reno distribution of BSD UNIX incorp orate Jacobsons slo w
start algorithm
In m ulticast comm unication the probabilit y of lo osing pac k ets gro ws as a
function of the group size and the cost of a pac k et loss is m uc h higher than in
a p oin ttop oin t exc hange Because reliable deliv ery is critical for data distri
bution applications our goal is to a v oid pac k et drops
Another v arian t of TCP called TCP V egas implemen ts a senderside
congestion a v oidance algorithm In Danzig et al conrms that V egas
congestion a v oidance sc heme yields higher throughput and k eeps less data in
the net w ork than Reno By computing the dierence b et w een b est and curr ent
roundtrip times R TT aV egas sender measures the amoun t of data queued
in the net w ork and adjusts its transmission windo w accordingly This congestion a v oidance approac h in a m ulticast en vironmen t do es not
scale w ell Senders need A CKs to measure R TTs whic h for large groups ma y
lead to ac k implosion There is also the unkno wn mem b ership problem If
group mem b ers clo c ks w ere sync hronize it w ould b e p ossible to measure one
w a y trip times and ha v e receiv ers compute queuing themselv es Ho w ev er clo c k
sync hronization requires an additional proto col In Section w e describ e some
tec hniques to address the scalabilit y problem
Reliable Multicast T ransp ort Proto cols
T raditional reliable unicast transp ort proto cols suc h as TCP use p ositiv e ac
kno wledgmen ts A CKs to reco v er from pac k et loss This approachto ac hieving
reliabilit y is often referred to as senderinitiate d since it is the resp onsibilit yof
the sender to detect pac k et losses In a m ulticast en vironmen t as group sizes in
crease the senderinitiated sc heme ma y cause acknow le dgment implosion since
eac h deliv ered pac k et triggers an ac kno wledgmen t from ev ery receiv er in the
group
Alternativ ely in the receiv erinitiated approac h to reliabilit y receiv ers detect
pac k et losses and request its retransmission b y generating a negativeac kno wl
edgmen t Placing the resp onsibilit y of reco v ering from pac k et losses on the
receiv er alleviates the ac kno wledgmen t implosion problem The p erformance
comparison study presen ted in conrms that receiv erinitiated m ulticast
transp ort proto cols deliv er b etter p erformance than their senderinitiated coun
terparts
Most reliable m ultip oin t transp ort proto cols are either pure receiv erinitiated
or use a h ybrid approachb y com bining receiv er and senderinitiated reliabilit y Belo w w eo v erview some of these proto cols
The Mo del
Our feedbac k suppression mec hanism fo cuses on the follo wing applicationlev el
requiremen ts and lo w erla y er services
IP Multicast
Senders transmit data pac k ets using in ternet m ulticast The currentm ul
ticast routing mo del whic h has b een in use on h undreds of routing domains
that form the In ternets MBONE
is based on D VMRP a distancev ector
m ulticast routing proto col D VMRP builds sourcero oted shortestpath
distribution trees where all leaf routers are attac hed to group mem b ers
IP m ulticast uses the In ternet Group Managemen t Proto col IGMP
to
manage group mem b ership Hosts send an IGMP join message to the
m ulticast group they w an t to join Multicastcapable routers use IGMP
messages to propagate mem b ership information among themselv es and to
p oll directly attac hed hosts for up dated mem b ership information
Although w e are assuming the currentIn ternet m ulticast mo del our pro
to col will w ork with an y of the alternate m ulticast mo dels that ha vebeen
prop osed suc h as CBT and PIM Unkno wn Group Mem b ership
T o supp ort scalabilit y it is assumed the set of receiv ers is unkno wn This
is also a reection of the seman tics of IP m ulticast where there is no
cen tralized group managemen t
Unkno wn Net w ork T op ology
No kno wledge of the underlying ph ysical net w ork top ology is assumed
Routers are not relied on to pro vide feedbac k ab out the net w ork condi
tions or lter feedbac k requests
Static Data
Our fo cus is on applications that distribute static data as opp osed to
realtime data Am ulticast le distribution service where les can b e of
arbitrary size is a t ypical target application
W e assume that application seman tics issues suc h as data consistency will b e handled b y the sp ecic application
P ersource Flo wCon trol
Although there can b e m ultiple sources in a m ulticast group o w con trol
is on a p ersource basis In other w ords w e are not designing an aggregate
o w con trol mec hanism for a m ulticast group
h ttpwwwresearc hattcomm b onefaqh tml
h ttpwwwcisohiostateeduh tbinrfcrfch tml
Application Seman tics
Applications are resp onsible for the proto col seman tics When mo ving
from a onetoone reliabilit y proto col suc h as TCP to a onetoman ypro to col that w ould b e used in m ulticast applications it b ecomes m uchmore
dicult to build a single transmission proto col that can handle all the
p ossible seman tics that migh t b e required
Ev en though w e are targeting bulk data transfer applications the mo del
is v ery general It can b e applied to anym ulticast application requiring
prompt feedbac k
F eedbac k Suppression Algorithm
Our m ulticast feedbac k mec hanism is based on the assumption that in a large
m ulticast group a small set of b ottlenec k links will cause the ma jorityof the
congestion problems W e exploit this b y nding a small set of group mem b ers to
represen t the congested m ulticast subtrees These group r epr esentatives pro vide
immediate feedbac k whic h can suppress feedbac k from other group mem bers
th us prev en ting feedbac k implosion at the source
If a receiv er nev er exp eriences an y pac k et loss or has its pac k et losses co v ered
b y a represen tativ e it will nev er generate an y messages
The rst c hallenge in selecting represen tativ es is to c ho ose them suchthat
they represen t the congested subtrees of the m ulticast tree Ideally eac h con
gested link w ould b e represen ted b y one represen tativ e in the aected subtree
The second c hallenge is to react to new congestion in a timely manner byc ho os
ing new represen tativ es and discarding those that are no longer con tributing to
the feedbackeorts The congestion a v oidance proto col ween vision relies on b oth p ositiveA CK
and negativeNA CK ac kno wledgmen ts A CKs are required to prev en t conges
tion collapse while NA CKs are required to pro vide feedbac k in the case of
congestion If a source do es not hear an y feedbac k it can assume that either
there are no group mem b ers other than itself or that there has b een some sort
of catastrophic net w ork failure
Receiv ers that ha v e b een selected as represen tativ es pro vide immediate feed
bac k to the source F eedbac k from other receiv ers is sc heduled o v er a random
in terv al and is sub ject to suppression W e explain ho w suppression timers are
set in Section A t startup there is no represen tativ e set and suppression timers are set v ery
lo osely As feedbac k comes in the source builds the represen tativ e set In
the absence of NA CKs the receiv ers whose A CKs w ere receiv ed b y the source
will b e selected as represen tativ es Since NA CKs are an immediate indication
In our sim ulation w e initially set the suppression timers to milliseconds Once the
GR TT see is computed w e use that as the basis to set timers
of congestion feedbac k suppression will giv e precedence to NA CKs o v er A CKs
Receiv ers sending NA CKs will tak e precedence o v er receiv ers sending A CKs for
consideration as represen tativ es As net w ork conditions c hange new feedbac k
is receiv ed b y the source and the represen tativ e set is up dated
The Source
The source main tains the represen tativ e set and computes the groups curren t
maxim um roundtrip time GR TT dened in Section The source is re
sp onsible for distributing the curren t represen tativ e set and the GR TT to the
group
Receiv ers
Up on receiving a data pac k et nonrepresen tativereceiv ers sc hedule a NA CK
if a data pac k et N is receiv ed without ha ving seen pac k et N
Otherwise an
A CK is sc heduled A sc heduled resp onse is held for a random p erio d of time
b efore b eing sen t If another resp onse is receiv ed b efore the sc heduled send time
and the receiv ed resp onse is dened to b e as go o d or b etter than the resp onse
sc heduled the sc heduled resp onse is suppressed In our denition of go o d or
b etter NA CKs suppress NA CKs and A CKs while A CKs can only suppress
other A CKs
Represen tativ es
Receiv ers designated as represen tativ es send feedbac k to the source immediately
forgoing an y suppression in terv al
A receiv er designates itself as a represen tativ e when it receiv es a represen
tativ e set notication in whic hitisamem b er A represen tativ e rev erts to non
represen tativ e op eration when it receiv es a represen tativ e set up date in whic h
it is not a mem b er
Represen tativ e Selection
A t startup an y receiv er that pro vides feedbac k is eligible for selection as a
represen tativ e After a full represen tativ e set has b een obtained only NA CKs
qualify a receiv er for selection as a represen tativ e T o prev entasudden c hange of
the represen tativ e set only one new represen tativema y b e selected in resp onse
to an y one pac k et When the represen tativ e set is full and a new represen tativ e
is selected an existing represen tativ e m ust be ejected from the curren t set
The b est candidate for ejection is ob viously the w orst represen tativ e but the
problem then lies in what constitutes the w orst Tok eep things simple the
Since this is not a reliabilit y mec hanism w eonlyNA CK the N pac k et W e are not
concerned with an y other previous lost pac k ets
represen tativ e that has not sentaNA CK in the longest time is selected This
criteria is based on the assumption that a represen tativethat has not sen t a
NA CK is ha ving the few est n um b er of congestion problems
Timers
Since w e cannot compute R TT b et w een the receiv ers and the source w e cannot
use the precise R TT to set up the suppression timers as done in SRM So w e are forced to resort to cruder measures W e can partially comp ensate
for using lo ose timers b y using represen tativ es to pro vide fast feedbac k Since
represen tativ es pro vide immediate feedbac k on b ehalf of the subtrees they co v er
w e need bac kup timers for those losses not co v ered b y represen tativ es
W e break our suppression times in to t w o comp onen ts The rst is a simple
wait p erio d and the second is a random interval The purp ose of the w ait p erio d
is to allo w time for resp onses b y the represen tativ es to tra v erse the group thereb y
suppressing nonrepresen tativ e feedbac k A t the same time they cannot b e to o
long so that losses not co v ered b y the curren t represen tativ e set can b e detected
in a timely fashion The purp ose of the random in terv al is to space out feedbac k
resp onses and allo w probabilistic suppression to reduce the amoun t of feedbac k
The w ait in terv al is set as p ercen tage of the estimated GR TT describ ed
b elo w
GR TT Measuremen t
Computing the maxim um group roundtrip GR TT time is a dicult and am
biguous prop osition If w e simply k eep trac k of the w orst R TT in the group
w e can get an o v erly p essimistic v alue ie a v alue inordinately large due to
transien t congestion Once a large v alue has b een established it migh tnev er b e
reduced The ob vious answ er is to decaythe v alue o v er time Again a problem
arises in in c ho osing what deca y function to use
Another alternativ e is to measure all R TTs and use a simple a v eraging lter
One then winds up with an a v erage whic h is not desirable If w e deca y only the
w orst v alues then w eha v e to dene a maxim um in terv al o v er whichto w ait for
resp onses whic h again is not what w ew an t
Our solution is for the source to keepatable of the w orst R TTs receiv ed
Eac h table en try containsanR TT measuremen t and a timetoliv e TTL Eac h
time a pac k et is sen t the TTL is decremen ted The curren t TTL valuesisset
to Wec hose this v alue to reduce rapid turno v er in the R TT set in the case
of lost pac k ets Since w e are measuring R TT for ev ery feedbackpac k et receiv ed
the R TTs should b e up dated once for ev ery data pac k et sen t
Wek eep trac k of the three w orst R TTs receiv ed When a new R TT is receiv ed
it is assigned a TTL Note that w e do not generate an y additional messages to
estimate the GR TT
Timer Setting
Eac h receiv ers w ait p erio d and suppression in terv al is dened as a p ercen tage
of the adv ertised GR TT Since A CKs pro vide little o w con trol information
they can b e suppressed to an arbitrary degree Ideally w e only need one A CK
p er pac k et Curren tly w e set the A CK w ait in terv al to GR TT
In the case of NA CKs w e are faced with the follo wing tradeo On one
hand w e seek to minimize the n um ber of NA CKs On the other w e wish to
receiveNA CKs in a timely fashion since w e need to react to them as quic kly as
p ossible In the normal case represen tativ es should b e selected in sucha w a y
as to maximize the probabilit y of resp onding to pac k et loss When represen ta
tiv es fail to co v er the pac k et loss the bac kup mec hanism of nonrepresen tativ es
sending NA CKs m ust react as quic kly as p ossible
Ov erhead
Since our feedbac k con trol mec hanism do es not rely on R TT estimation the
only bandwidth o v erhead required b y our sc heme is the distribution of a new
represen tativ e set and GR TT GR TT measuremen ts are sentaspartof a pac k et
header and hence do require separate messages
W e tried to k eep the computational o v erhead to a minim um The most
computationally in tensiv e features of our algorithm are in the source The sizes
of the represen tativeset and maxim um R TT tables are b ounded to limit the
amoun t of computational o v erhead
Similarly the space requiremen ts at the source are b ounded b y the repre
sen tativ e set and maxim um R TT table sizes
Results
W e use a simple sim ulator to ev aluate our feedbac k suppression mec hanism The
goal is to ev aluate the amoun t and timeliness of the feedbac k generated bythe
receiv ers in a m ulticast group connected using an arbitrary net w ork top ology
with loss rates and dela ys sp ecied p er link
Ev aluation Metho dology
Net w ork T op ologies
Early tests of the feedbac k suppression mec hanism utilized a simple binary tree
net w ork with the source transmitting at the ro ot of the tree and all other no des
b eing receiv ers While suc h net w orks w ere easy to generate and pro vided useful
preliminary information on the feedbac k suppression mec hanism it w as neces
sary to test the feedbac k suppression mec hanism on more realistic net w orks
T o generate more realistic top ologies w e use Stev e Hotzs Net w ork T op ology
Generator The net w orks generated w ere three lev el net w orks with lo w dela y
links at the rst and second lev els and large dela y links at the third lev el This
is in tended to b e a lo ose appro ximation of a real in ternet w ork top ology with a
high sp eed bac kb one slo w er regional net w orks and nally slo w nal links The
loss rates w ere randomly c hosen o v er the in terv al See for more details
on the generation of the net w orks
The receiv ers are all at the third lev el and hence will tend to b e lea v es but
still ha v e dieren t distances to the source The source is at the top lev el
Sim ulator
The sim ulator generates a minim um dela y spanning tree from a giv en net w ork
top ology Link dela ys and drop rates are sp ecied p er link No capacitymod eling is done ie all links ha v e innite bandwidth Endtoend dela y is simply
the sum of the link dela ys and no dela y jitter is in tro duced The mem b ership
of the group is static
Congestion is mo deled in a v ery simple w a y A t the b eginning of a sp ecied
in terv al a link is c hosen at random and giv en a high loss rate After the
sp ecied congestion in terv al in n um b er of source pac k ets has elapsed the loss
rate is returned to normal In the curren t sim ulations the congestion loss rate is
set to and the in terv al is set to source pac k ets Congestion is in tro duced
once p er congestion in terv al
The source transmits data pac k ets at a constan t rate of pac k ets a second
Wedescribe thesim ulation parameters b elo w and T able summarizes them
Represen tativ es T o measure the eect of represen tativ es w e sim ulated
groups with and represen tativ es
NA CK W ait In terv al As explained in Section the feedbac k timer
dela y has t w o comp onen ts The rst is a simple w ait in terv al and the
second is a random in terv al The w ait in terv al is a percen tage of the
estimated maxim um group roundtrip time GR TT The feedbac k time
is sc heduled b y adding the w ait and random in terv als
In our sim ulations w e used w ait times of and of the maxim um
GR TT
Maxim um Link Loss Rate
Eac h link is randomly assigned a v alue b et w een and the maxim um loss
rate The o v erall loss rate of the net w ork is a function of the link loss
rates net w ork size and top ology In our sim ulations the net w ork loss rate is v ery high The net w orks t ypi
cally ha v e a loss rate since w ew an ted to test our feedbackmec hanism
on the w orst p ossible cases
Represen tativ es NA CK w ait in terv al Maxim um link loss rate T able Sim ulation P arameters
Congestion In terv al The length of time that congestion lasts A t the
b eginning of the in terv al a link is c hosen and giv en a high loss rate A t
the end of this in terv al the loss rate returns to normal
Congestion Loss Rate The loss rate on the congested link
W e ran our sim ulator using a randomlygenerated no de top ology with
and lev el lev el and lev el no des resp ectiv ely In eachsim ula
tion run w ev ary one of the parameters in T able while k eeping all the other
parameters constan t
Ev aluation Metrics
In this ev aluation w e are concerned with t w o metrics The rst is the quan tit y
of the feedbac k and the second is timeliness
F or feedbac k quan tit y w e lo ok at the total amoun t of feedbac k receiv ed
The less the b etter Ideallyw ew an t a constan t amoun t of feedbac k p er pac k et
indep enden t of group size
F or timeliness w e normalize the feedbac k time to the maxim um GR TT o v er
the en tire sim ulation Ideally wew ould lik e the R TT to b e the minim um R TT
for a giv en congested p oin t to the source In practice it is dicult to select a
represen tativ e that is the rst represen tativ e on the far side of the congested
link If w e shorten the random suppression in terv al the probabilit y of receiving
feedbac k from the rst congested receiv er increases but the amoun t of feedbac k
is increased While this is not a problem for leaf links it can create a great deal
of feedbac k if the congested link o ccurs near the ro ot of a large m ulticast tree
Ev aluation Results
In the graphs sho wn in this section eachpoin t corresp onds to the a v erage of
a sequence of sim ulation runs where all parameters w ere k ept constan t W e
ev aluate the eects of the NA CK w ait timer represen tativ es and group size
in the quan tit y and qualit y of the feedbac k generated bythe m ulticast group
The graphs that sho w amoun t of feedbac k generated b y the group plot n um ber
of pac k ets receiv ed at the source against data pac k et sequence n um b er The
timeliness graphs plot the minim um time for the source to receiv e feedbac k
normalized to GR TT against data pac k et sequence n um ber
NA CK W ait Timer Settings
Figures and sho whowthe NA CK w ait in terv al inuences the quan tityand
the timeliness of the feedbac k receiv ed at the source Figure sho ws the amoun t
of feedbac k generated b y the group in terms of n umberofNA CKs and A CKs
and Figure sho ws the time for NA CKs to arriv e at the source normalized to
the corresp onding GR TT F or these sim ulations w e use mem b er groups represen tativ es and a maxim um link loss rate of Note that this maxim um
link loss rate resulted in an o v erall loss rate of appro ximately whic h is a
considerably higher loss rate than what is observ ed in real net w orks There
fore in our sim ulations w e are driving our feedbac k con trol algorithms with
p essimistic scenarios in terms of loss rates
Regarding NA CKs in the case where the NA CK w ait timer is zero Figure
a the a v erage n um ber of NA CKs receiv ed is at least t wice the a v erage
n um ber of NA CKs when the NA CK w ait time is GR TT Figure c This
is due to the fact that longer NA CK w ait in terv als giv e represen tativ es a b etter
c hance to suppress nonrepresen tativeNA CKs
During the rst data pac k ets while represen tativ es are rst b eing selected
w e receiv e noticeably more feedbac k The represen tativ e set is constan tly c hang
ing during the sim ulation in resp onse to pac k et loss b eing detected in subtrees
not co v ered b y the curren t represen tativ e set This accoun ts for some of the
v ariation in the amoun t of feedbac k generated
When comparing the graphs obtained for the dieren tNA CK timer settings
in the represen tativ e case with the represen tativ e case
it is clear that
represen tativ es greatly reduce the impact of ho w the NA CK w ait timer is set
This is b ecause when there are no represen tativ es feedbac k con trol relies on
pure suppression and hence timer setting is crucial Represen tativ es pro vide
immediate feedbac k and only rely on NA CK timers for pac k et losses they still
do not co v er
If the curren t represen tativ e set co v ered all pac k et losses in their m ulticast
group longer NA CK timers w ould be the w a y to go Ho w ev er since in real
net w orks congestion o ccurs at dieren t p oin ts at dieren t times NA CK w ait
timers should b e set so that nonrepresen tativ es can still resp ond to losses not
co v ered b y the curren t represen tativ es in a timely fashion
The graphs in Figure sho w the impact of NA CK timers on feedbac k time
liness There is the ob vious result that shorter NA CK timers allo w feedbackto
be receiv ed so oner While in the case of GR TT NA CK w ait timer it tak es
on a v erage GR TT for feedbac k to arriv e at the source it tak es less than GR TT when the NA CK w ait timer is Ho w ev er as w e discuss in the next sec
tion represen tativ es comp ensate for longer NA CK timers b y sending immediate
feedbac k
The represen tativ e graphs with dieren tNA CK timer settings ha v e b een omitted due
to space limitations
0
10
20
30
40
50
0 10 20 30 40 50 60 70 80 90 100
# packets
sequence #
ACK & NACKS
ACKS Sent
NACKS Sent
a NA CK w ait timer is of the
maxim um GR TT
0
10
20
30
40
50
0 10 20 30 40 50 60 70 80 90 100
# packets
sequence #
ACK & NACKS
ACKS Sent
NACKS Sent
b NA CK w ait timer is of the
maxim um GR TT
0
10
20
30
40
50
0 10 20 30 40 50 60 70 80 90 100
# packets
sequence #
ACK & NACKS
ACKS Sent
NACKS Sent
c NA CK w ait timer is of the
maxim um GR TT
Figure Eects of NA CK timers on feedbac k Sim ulation runs used mem ber m ulticast groups with represen tativ es maxim um link loss rate of
and NA CK w ait timer equal to and of the GR TT resp ectiv ely Net w ork loss rate is appro ximately
0
0.5
1
1.5
2
2.5
3
0 10 20 30 40 50 60 70 80 90 100
# RTTs
sequence #
NACK Response Times
Min NACK Response
a NA CK w ait timer is of the
maxim um GR TT
0
0.5
1
1.5
2
2.5
3
0 10 20 30 40 50 60 70 80 90 100
# RTTs
sequence #
NACK Response Times
Min NACK Response
b NA CK w ait timer is of the
maxim um GR TT
0
0.5
1
1.5
2
2.5
3
0 10 20 30 40 50 60 70 80 90 100
# RTTs
sequence #
NACK Response Times
Min NACK Response
c NA CK w ait timer is of the
maxim um GR TT
Figure Eects of NA CK timers on feedbac k Sim ulation runs used mem ber m ulticast groups with represen tativ es maxim um link loss rate of
and NA CK w ait timer equal to and of the GR TT resp ectiv ely Net w ork loss rate is appro ximately
During the rst data pac k ets w e also observ e the impact of the represen
tativ e set startup on feedbac k timeliness
Represen tativ es
Figures and Figure sho w the amoun t and timeliness of feedbac k for dieren t
represen tativeset sizes These sim ulations used mem ber groups a maxi
m um link loss rate of or appro ximately o v erall loss rate and NA CK w ait
in terv al of of GR TT Graphs a b c and d in eac h gure corresp ond
to represen tativ e set sizes of and resp ectiv ely Figure sho ws ho w represen tativ es can reduce the amoun t of feedbac k re
ceiv ed W e observ e that using represen tativ es result in a signican t reduction in
the amoun t of feedbac k generated Ov erall the represen tativecase a v erag
ing NA CKs p er pac k et sen t generates less feedbac k than the represen tativ e
case NA CKs per pac k et but still p erform eectiv e suppression The represen tativecase ho w ev er seems to b e more eectiv e at limiting the v ariance
in the amoun t of feedbac k This is due to the fact that the represen tativ e
set pro vides a b etter co v erage of pac k et losses and therefore k eeps the amoun t
of feedbac k generated b y the group more constrained
F or the nonempt y represen tativ e set graphs the amoun t of feedbackishigh
during the rst pac k ets b ecause represen tativ es w ere still b eing elected
There is a tradeo b et w een the amoun t of feedbac k generated and the time
liness of the proto col On a v erage represen tativ es lo w er the amoun t of feedbac k
without incurring unreasonable latency in getting feedbac k to the source Fig
ure a sho ws the nonrepresen tativ e case A pure suppression sc heme with the
NA CK w ait in terv al equal to results in an impro v ementof GR TT o v er the
case where the represen tativ e set in not empt y In other w ords represen tativ es
pa y a p enalt y of ab out GR TT in terms of timeliness
In our simple sim ulation mo del represen tativ es tend to b e farther from the
source This is due to the fact that the farther do wn the tree a receiv er is the
more lik ely it is to lose a pac k et The more lik ely a receiv er loses pac k ets the
more lik ely it is that it be elected as a represen tativ e Hence represen tativ es
will tend not b e the receiv ers closer to the source
In this section w e sho w ed that for a giv en m ulticast group size w e can
p erform eectiv e suppression with v ery few represen tativ es In the next section
w e showho w our algorithm p erforms for dieren t group sizes
Group Size
Figures and Figure sho w the amoun t and timeliness of feedbac k generated
b y groups of dieren t sizes W e use group sizes of and and set
the NA CK timer at of the GR TT and the maxim um link loss rate at or appro ximate o v erall loss rate W e use a represen tativ e set in all cases
0
10
20
30
40
50
0 10 20 30 40 50 60 70 80 90 100
# packets
sequence #
ACK & NACKS
ACKS Sent
NACKS Sent
a No Represen tativ es
0
10
20
30
40
50
0 10 20 30 40 50 60 70 80 90 100
# packets
sequence #
ACK & NACKS
ACKS Sent
NACKS Sent
b Represen tativ es
0
10
20
30
40
50
0 10 20 30 40 50 60 70 80 90 100
# packets
sequence #
ACK & NACKS
ACKS Sent
NACKS Sent
c Represen tativ es
0
10
20
30
40
50
0 10 20 30 40 50 60 70 80 90 100
# packets
sequence #
ACK & NACKS
ACKS Sent
NACKS Sent
d Represen tativ es
Figure Eects of represen tativ es on feedbac k W e use and rep
resen tativ es in a mem b er group NA CK w ait time is of GR TT and
maxim um link loss rate is Net w ork loss rate is appro ximately
0
0.5
1
1.5
2
2.5
3
0 10 20 30 40 50 60 70 80 90 100
# RTTs
sequence #
NACK Response Times
Min NACK Response
a No Represen tativ es
0
0.5
1
1.5
2
2.5
3
0 10 20 30 40 50 60 70 80 90 100
# RTTs
sequence #
NACK Response Times
Min NACK Response
b Represen tativ es
0
0.5
1
1.5
2
2.5
3
0 10 20 30 40 50 60 70 80 90 100
# RTTs
sequence #
NACK Response Times
Min NACK Response
c Represen tativ es
0
0.5
1
1.5
2
2.5
3
0 10 20 30 40 50 60 70 80 90 100
# RTTs
sequence #
NACK Response Times
Min NACK Response
d Represen tativ es
Figure Eects of represen tativ es on timeliness W e use and represen tativ es in a mem b er group NA CK w ait time is of GR TT and
maxim um link loss rate is Net w ork loss rate is appro ximately
0
10
20
30
40
50
0 10 20 30 40 50 60 70 80 90 100
# packets
sequence #
ACK & NACKS
ACKS Sent
NACKS Sent
a mem b er group
0
10
20
30
40
50
0 10 20 30 40 50 60 70 80 90 100
# packets
sequence #
ACK & NACKS
ACKS Sent
NACKS Sent
b mem b er group
0
10
20
30
40
50
0 10 20 30 40 50 60 70 80 90 100
# packets
sequence #
ACK & NACKS
ACKS Sent
NACKS Sent
c mem b er group
Figure Eects of group size on amoun t of feedbac k W e use and mem b er groups NA CK w ait time is of GR TT maxim um link loss rate is
and n um b er of represen tativ es is The net w ork loss rate is appro ximately
F rom Figure w e observ e that the size of the group has relativ ely little
impact on the o v erall feedbac k receiv ed This conrms that a few represen tativ es
are quite eectiv ein k eeping feedbac k amountlo w A larger represen tativeset
represen tativ es instead of for the mem b er group Figure c results
in less v ariation in the amoun t of feedbac k As w e p oin ted out in Section this is due to the fact that more represen tativ es pro vide a b etter co v erage of
pac k et loss
Similarly to the amoun t of feedbac k generated Figure sho ws that feedbac k
timeliness do es not degrade as the group size increase On a v erage the time for
the source to receiv e feedbackis k ept under GR TT for all groups sizes sho wn
Congestion
W e can see from Figure that when congested links are mo deled in a m ulticast
tree the impact of represen tativ es on the amoun t and v ariabilit y of feedbac k
is ob vious With no represen tativ es there is a large v ariabilit y in the amoun t
of feedbac k As represen tativ es are added the v ariabilit y in the amoun t of
feedbac k is greatly reduced Spik es do o ccasionally o ccur whic h is undesirable
The spik es o ccur when congestion o ccurs far up the m ulticast tree causing a
large n um ber of receiv ers to sc hedule NA CKS The suppression mec hanism is
then not able to suppress as man y replies as wew ould lik e The spik es could b e
reduced b y increasing the timers v alues but this w ould also reduce the o v erall
timeliness of the resp onses More w ork is required to address this problem
In Figure the timeliness of the NA CKs receiv ed in the represen tativecase
compares fa v orably with the nonrepresen tativ e case
F uture W ork
Weha veev aluated the utilit y of represen tativ es in purely lossy net w orks W e
exp ect that p erformance will be ev en b etter in real net w orks in whic h losses
result primarily from congestion With this in mind future sim ulations need to
takein to accoun t net w ork congestion and b ottlenec k links In the short term
w e can extend the curren t simple sim ulator to create esp ecially lossy links to
sim ulate congestion and v ary these loss rates o v er time to sim ulate transien t
congestion in the net w ork
An ob vious optimization to reduce feedbac k is to only allo w represen tativ es
to send A CKs This will greatly reduce A CKs while still pro viding protection
in the case of congestion collapse NA CK feedbac k is not aected
In addition to the random top ologies weha v e examined w e will in v estigate
star net w orks that one migh t nd in satellite systems as w ell as net w orks con
taining mobile hosts W e will also in v estigate the p ossibilit y of automatically
determining the a go o d represen tativ e set size
0
0.5
1
1.5
2
2.5
3
0 10 20 30 40 50 60 70 80 90 100
# RTTs
sequence #
NACK Response Times
Min NACK Response
a mem b er group
0
0.5
1
1.5
2
2.5
3
0 10 20 30 40 50 60 70 80 90 100
# RTTs
sequence #
NACK Response Times
Min NACK Response
b mem b er group
0
0.5
1
1.5
2
2.5
3
0 10 20 30 40 50 60 70 80 90 100
# RTTs
sequence #
NACK Response Times
Min NACK Response
c mem b er group
Figure Eects of group size on timeliness of feedbac k W e use and mem ber groups NA CK w ait time is of GR TT maxim um link
loss rate is and n um ber of represen tativ es is The net w ork loss rate is
appro ximately
0
10
20
30
40
50
0 50 100 150 200250 300 350 400 450 500
# packets
sequence #
ACK & NACKS
ACKS Sent
NACKS Sent
a No represen tativ es
0
10
20
30
40
50
0 50 100 150 200250 300 350 400 450 500
# packets
sequence #
ACK & NACKS
ACKS Sent
NACKS Sent
b represen tativ es
0
10
20
30
40
50
0 50 100 150 200250 300 350 400 450 500
# packets
sequence #
ACK & NACKS
ACKS Sent
NACKS Sent
c represen tativ es
Figure Eects of congestion on the amountof NA CK feedbackfor the
mem b er group Aggregate net w ork loss rate is appro ximately
0
0.5
1
1.5
2
2.5
3
0 50 100 150 200250 300 350 400 450 500
# RTTs
sequence #
NACK Response Times
Min NACK Response
a No represen tativ es
0
0.5
1
1.5
2
2.5
3
0 50 100 150 200250 300 350 400 450 500
# RTTs
sequence #
NACK Response Times
Min NACK Response
b represen tativ es
0
0.5
1
1.5
2
2.5
3
0 50 100 150 200250 300 350 400 450 500
# RTTs
sequence #
NACK Response Times
Min NACK Response
c represen tativ es
Figure Eects of congestion on the timeliness of NA CK feedbac k for the mem b er group Aggregate net w ork loss rate is appro ximately
Our goal is to build a full o w con trol mec hanism suitable for bulk data
distribution in a m ulticast en vironmen t Since an y o w con trol sc heme m ust
detect congestion in the net w ork w e need to in v estigate congestion detection
sc hemes In addition to ev aluating the use of pac k et loss as a congestion metric
as is done in traditional TCP w e w ould also lik e to in v estigate dela ybased
metrics lik e that used in TCPV egas Conclusion
This pap er describ ed a m ulticast feedbac k con trol sc heme that is part of our
eorts to build a scalable reliable o wcon trolled m ulticast transp ort proto col
suitable for bulk data transfer applications
The ma jor b enet of our algorithm deriv es from the fact that wedo not need
to compute round trip time from receiv ers to the source and w e do not require
kno wledge of group mem b ership or net w ork top ology W e use a small set of represen tativ es in com bination with probabilistic sup
pression to limit feedbac k y et not signican tly degrading the timeliness of the
feedbac k with resp ect to the w orst round trip time for the group
Wein v estigated the p erformance of our feedbac k con trol mec hanism in purely
lossy net w orks with no capacit y mo deling Represen tativ es sho w a mark ed im
pro v emento v er a purely suppression orien ted algorithm
References
JS Ahn P B Danzig Z Liu and L Y an Ev aluation of tcp v egas Em u
lation and exp erimen t A CM SIGCOMM Confer enc e pages
Octob er AJ Ballardi P F F rancis and J Cro w croft Core based trees cbt Pr o c
of the A CM SIGCOMM San F r ancisc o CA August LS Brakmo SW OMalley and LL P eterson TCP V egas New tec h
niques for congestion detection and a v oidance A CM SIGCOMM
Confer enc e pages Ma y DD Clark ML Lam b ert and L Zhang Netblt A high throughput
transp ort proto col A CM SIGCOMM Confer enc e pages
August P B Danzig Optimal ly Sele cting the Par ameters of A daptive Backo A lgo
rithms for Computer Networks and Multipr o c essors PhD thesis Univ ersit y
of California Berk eleyDecem b er
S Deering D Estrin D F arinacci V Jacobson C Liu and L W ei
An arc hitecture for widearea m ulticast routing A CM SIGCOMM
Conference F ebruary S Flo yd V Jacobson S McCanne CG Liu and L Zhang A reliable
m ulticast framew ork for ligh tw eigh t sessions and applicationlev el framing
A CM Sigc omm Confer enc e pages Octob er M Grossglauser Optimal deterministic timeouts for reliable scalable m ulti
cast Pr o c of the IEEE Info c omm San F r ancisc o CA pages Marc h S Hotz and R Nagamati Net w ork top ology generator NTG A to ol for
generating net w ork top ology and p olicy for proto col sim ulation purp oses
T ec hnical Rep ort Computer Science Departmen t Univ ersit y of Southern
California Spring SK Singhal HW Holbro ok and DR Cheriton Logbased receiv erreliable
m ulticast for in teractivesim ulation A CM SIGCOMM Confer enc e pages Septem b er V Jacobson Berk eley tcp ev olution from taho e to reno Pro ceedings
of the British Colum bia In ternet Engineering T ask F orce July V an Jacobson Congestion a v oidance and con trol A CM SIGCOMM pages T T urletti JC Bolot and I W ak eman Scalable feedbac k con trol for m ul
ticast video distribution in the in ternet Pr o c of the Confer enc e on Com
munic ations A r chite ctur es Pr oto c ols and Applic ations A CM SIGCOMM
August JB Nagle On pac k et switc hes with innite storage IEEE T r ansactions
on Communic ations April S Pingali D T o wsley and J F Kurose A comparison of senderinitiated
and receiv erinitiated reliable m ulticast proto cols A CM SIGMET
RICS Confer enc eMa y KK Ramakrishnan and Ra j Jain A binary feedbac ksc heme for congestion
a v oidance in computer net w orks A CM T r ansactions on Computer Systems Ma y H Sc h ulzrinne A transp ort proto col for realtime applications In ter
net Draft In ternet Engineering T ask F orce AudioVideo T ransp ort W G
Marc h T T urletti H soft w are co dec for video conferencing o v er the in ternet
INRIA Researc h Rep ort Jan uary
Abstract (if available)
Linked assets
Computer Science Technical Report Archive
Conceptually similar
PDF
USC Computer Science Technical Reports, no. 651 (1997)
PDF
USC Computer Science Technical Reports, no. 658 (1997)
PDF
USC Computer Science Technical Reports, no. 680 (1998)
PDF
USC Computer Science Technical Reports, no. 735 (2000)
PDF
USC Computer Science Technical Reports, no. 707 (1999)
PDF
USC Computer Science Technical Reports, no. 595 (1994)
PDF
USC Computer Science Technical Reports, no. 662 (1997)
PDF
USC Computer Science Technical Reports, no. 652 (1997)
PDF
USC Computer Science Technical Reports, no. 713 (1999)
PDF
USC Computer Science Technical Reports, no. 660 (1997)
PDF
USC Computer Science Technical Reports, no. 714 (1999)
PDF
USC Computer Science Technical Reports, no. 738 (2000)
PDF
USC Computer Science Technical Reports, no. 698 (1999)
PDF
USC Computer Science Technical Reports, no. 737 (2000)
PDF
USC Computer Science Technical Reports, no. 689 (1998)
PDF
USC Computer Science Technical Reports, no. 854 (2005)
PDF
USC Computer Science Technical Reports, no. 682 (1998)
PDF
USC Computer Science Technical Reports, no. 747 (2001)
PDF
USC Computer Science Technical Reports, no. 755 (2002)
PDF
USC Computer Science Technical Reports, no. 657 (1997)
Description
Dante DeLucia and Katia Obraczka. "Multicast feedback suppression using representatives." Computer Science Technical Reports (Los Angeles, California, USA: University of Southern California. Department of Computer Science) no. 638 (1996).
Asset Metadata
Creator
DeLucia, Dante
(author),
Obraczka, Katia
(author)
Core Title
USC Computer Science Technical Reports, no. 638 (1996)
Alternative Title
Multicast feedback suppression using representatives (
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
25 pages
(extent),
technical reports
(aat)
Language
English
Unique identifier
UC16270398
Identifier
96-638 Multicast Feedback Suppression Using Representatives (filename)
Legacy Identifier
usc-cstr-96-638
Format
25 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
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/