• Aucun résultat trouvé

A simple stability condition for RED using TCP mean-field modeling

N/A
N/A
Protected

Academic year: 2021

Partager "A simple stability condition for RED using TCP mean-field modeling"

Copied!
25
0
0

Texte intégral

(1)

HAL Id: inria-00091036

https://hal.inria.fr/inria-00091036

Submitted on 5 Sep 2006

HAL

is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire

HAL, est

destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.

mean-field modeling

Julien Reynier

To cite this version:

Julien Reynier. A simple stability condition for RED using TCP mean-field modeling. [Research

Report] 2006, pp.24. �inria-00091036�

(2)

inria-00091036, version 1 - 5 Sep 2006

A simple stability condition for RED using TCP mean field modeling

Julien Reynier

N° ????

Août 2006

(3)
(4)

Julien Reynier

ThèmeCOM Systèmesommuniants

ProjetsTREC

Rapportdereherhe ???? Août200621pages

Abstrat: Congestion on the Internet is an old problem but still a subjet of intensive researh. The

TCPprotoolwith itsAIMD(AdditiveInreaseandMultipliativeDerease) behaviorhidesveryhallenging

problems;oneofthemistounderstandtheinterationbetweenalargenumberofuserswithdelayedfeedbak.

This artilewill fouson twomodeling issues ofTCP whih appeared to be importantto takleonrete

senarioswhenimplementingthemodelproposedin[7℄;rstlythemodelingofthemaximumTCPwindowsize:

thismaximumanbereahedquiklyinmanypratialases;seondlythedelaystruture: theusualLittle-like

formula behavesreally poorly when queuing delays are variable, and may hange dramatially the evolution

of the predited queue size, whih makes it useless to study drop-tail or RED (Random Early Detetion)

mehanisms.

Within proposed TCPmodeling improvements,weareenabled to lookat aonrete examplewhere RED

should beusedinFIFOroutersinsteadoflettingthedefaultdrop-tailhappen. Westudymathematiallyxed

pointsofthewindowsize distributionandloalstabilityofRED.AninterestingaseiswhenRED operatesat

thelimitwhentheongestionstarts,itavoidsunwantedlossofbandwidthanddelayvariations.

Key-words: TCP,AQM,drop-tail,RED,ongestion

TREC

ÉoleNormaleSupérieure,45,rued'Ulm,75005Paris

(5)

Résumé : Le ontrle de ongestion dans Internet est depuis longtemps le sujet de reherhes poussées.

Le protooleTCPavesonomportementAIMD(pouraroissementslinéaire,déroissanemultipliativeen

anglais)ahedesproblèmesexessivementompliqués. L'und'entreeuxestdeomprendrel'interationentre

denombreuxutilisateursaveundélaideréponsedusystème.

CerapportvasefoalisersurdeuxpointsdanslamodélisationdeTCP.Cespointssontapparusimportant

lorsque nous avons voulu onfronter à des sénarii onrets le modèle proposé dans [7℄; Tout d'abord la

modélisation delafenêtre maximale deTCP: ette valeurpeur être atteinte très failementdansla pratique;

ensuite, lastruture des délais: la formuletype Littlehabituellement employée donnedes résultatsloin de la

réalité quand les délaissont variables. Cette hypothèse de modélisation a un impat important sur la taille

prédite de lale d'attente e qui rend vaineslestentativesde omparaisonentre lesméanismes drop-tailet

RED.

Grâeàesaméliorationsapportéesaumodèle,noussommesapablesdansunaspréisd'étudieromment

paramétrer RED dansdes routeurs FIFOpourqu'ilaméliore lesperformanespar rapport auas pardéfaut

de drop-tail. Nousétudions mathématiquementlespointsxes etladistribution delataille desfenêtres et la

stabilité loaledeRED. Unas intéressantestquand RED setrouvedans sondomaine defontionnementau

début delaongestion,iléviteunemauvaise utilisationdelabandepassanteetdesvariationsdanslesdélais.

Mots-lés : TCP,AQM,drop-tail,RED,ongestion

(6)

Contents

1 Introdution 4

1.1 TCProuterontrolissue. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.2 Ourpreviousworksandmotivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.3 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.4 Newresults . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Model and equations 4 2.1 Evolutionof ongestionwindowsizes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2.2 Delayin thesystem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2.2.1 LimitsofLittle-likeformula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2.2.2 Howtoimprovedelaymodel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2.2.3 Delayequations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2.2.4 Bandwidthevolutionwhenrossingthereeivinguser . . . . . . . . . . . . . . . . . . . . 6

2.2.5 Generationoflosses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

3 Fixed point 7 3.1 Fixedpointequations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

3.2 Fixedpointresolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

3.2.1 Firstiteration. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

3.2.2 Seonditeration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

3.3 Normalization. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

3.3.1 Regularityproperties. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

3.3.2 Computationoftheintegral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

3.4 Equilibriumvalues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

4 Stabilityanalysis 12 4.1 Arstremark. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

4.2 Stabilityequations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

4.3 Dierentialequationswithtimedelay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

5 Simulation results 15 5.1 Resultswithdrop-tail . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

5.1.1 Beforeongestionhappens. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

5.1.2 Theearlyongestionphase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

5.1.3 Strongongestionphase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

5.2 ResultswithRED . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

5.2.1 Conguration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

5.2.2 RED initsworkingregime . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

5.2.3 RED workinglikeadrop-tail . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

5.3 Inreasingthelateny . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

5.4 Mixinglatenies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

6 Conlusion 18 7 Related Works 18 7.1 TCPmodelingarea. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

7.2 ControltheoryappliedtoTCP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

7.3 BuersizingforIProuters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

8 Further Works 19

(7)

1 Introdution

1.1 TCP router ontrol issue

TCPahievesadistributedongestionontroloftheInternet. ThisartileprovesausablelosedformulaRED

stability. RED (RandomEarlyDetetion)wasintroduedbyFloyd in[14℄;it istobedeployedat arouterto

send ongestioninformationtoTCPRenoend users. TheideabehindREDisthat therstsignofongestion

is whentherouterqueuestartstobeusedmorethanto buernormaltrautuations;thenthebuergets

full andthedrop-tailmehanismdestroyspaketsarrivingwithoutanyroomleft inthequeueto tin.

Drop-tailleadstotwoissues: rstly,thequeue sizeosillationsprovokedelayjitters- thishasdetrimental

eets for appliations using TCPfor realtime ontent -seondly, drop-tail synhronizes soures, resultingin

bandwidthunder-utilizationoftheongestedlink (thisideawasrstintrodued in[41℄for TCPTahoe). This

serves as leverage beause at the time bandwidth demand reahes apaity, the goodput diminishes by the

synhronizationeet,worseningthestartingongestion(thisassertionwillbeexplainedlearlylater).

Thesetwomain reasonsexplaintheinterestforRED andotherAQM(AtiveQueueManagement)to deal

with TCP ongestion at therouter level. RED often works in an admirableway, leadingto reduedqueuing

delays,avoidingjittersandreahingoptimalbandwidthutilization... butsometimesREDperformsworsethan

doingnothingatall(drop-tail). Thisis thereasonwhymanysystemadministratorsarerelutanttouseRED

althoughitisdeployedinalmosteveryrouteroftheInternet. ThispaperwillshowhowtotuneRED inaway

itissometimesoptimalandalwaysbetterthandrop-tail.

1.2 Our previous works and motivation

In[7,29℄and[40℄weinvestigatedmeaneldTCPmodelingbyontinuingtheuidTCPmodelintroduedand

studiedin[15,25,31℄. Despitetheinterestingresultsarisingfromthemodels,therewerestillsomediultiesin

understandingtheoriginalproblemoftuningREDandomparingitauratelytodrop-tail. Twopointsneeded

to beaddressed. Firstly,withthedevelopmentof highspeedaess,itbeomesdiulttosupposethat TCP

alwaysworkswithinitsongestionavoidanemodeinaAIMDmanner. Thesizeofpaketsoftheorderof1kB

makesthemaximumTCPwindowsize relativelysmall(most ommonpaketsize isaround1.4kB). Weshall

sayWmax= 64 paketsevenifthereeiverdoesnotimposeanyreeptionwindowlimitation. Thisfat isdue to theodingofthewindowsize on16bitsaddressingwindowbyBytes(216B = 64kB).

Seondly, another limiting modeling assumption is afat notied by Hongin [16℄: when thequeue is not

empty,aknowledgementsarriveobviouslyattheongested routerbandwidth. This remarkisruial beause

TCPdynamiisverysensitivetothedelayedfeedbak.

1.3 Outline

Setion 2explainsanddenes ourmodel;thenwestudy thesteadystatewindowdistributionwith amaximal

windowsizeinsetion3. NextsteponsistsinseekingastabilityregionfortheRED algorithm,whih isdone

in setion4. Wenishwith showingsimulationresultsonaonreteexamplein setion5. Thislast example

showshowtousepreviousresultstoongureRED inarouterinordertoavoidollapseattheearlystagesof

ongestion.

1.4 New results

Whereas modelingWmax and ACK bandwidthare not newideas(see [34℄ and [16℄), adapting them in mean

eld equationsto obtainaurateevolutionequationstogetherwiththewindowdistribution onstitutesastep

forward. Thesteady statesolutionforthewindowdistribution takinginto aounttheWmax phenomenon in

setion3isanextensionof[7℄whihisimportantfromapratialpointofview. Insetion4,thestabilityresult

obtainedforREDwiththeorem4isaverysimplelosedformula. Finallytheexampleinsetion5explainshow

REDshouldbetunedtoinreaseroutereieny,thisisanimportantresultbeause,aswesaid,thesuggested

tuning anbeapplied withoutanyhardwaremodiationinalmosteveryrouterbyenablingRED.

2 Model and equations

A number N, relatively large, of users share a ommon bottlenek router (gure 1 to see the modeled net-

work topology). We an onsider the histogram of users' ongestion windowsizes; in [29℄, we sawthat this

(8)

histogram onverges"gently"to adeterministiwindowdistribution asN tends toinnity. In[7℄, westudied

this asymptoti distributionwhih satises apartial dierentialequation the resultswere appliable evenfor

small numbers (N = 25 for RED, and N = 10 for adrop-tail). Hereinafter weadapt thepartial dierential equationsrstpresentedin [7℄;thisisdoneinawayweouldprovethemeaneld limitaswedidin[29℄,but

weshallnotenter insuhdevelopmentsinthisartile.

Queue, 235 packets User 1, W

1

User 2, W

2

… User N, W

N

1Gb/s, 4 ms

100 Mb/s, 1 ms

TCP sink 1 ms

1 ms

Queue, 235 packets User 1, W

1

User 2, W

2

… User N, W

N

1Gb/s, 4 ms

100 Mb/s, 1 ms

TCP sink 1 ms

1 ms

Figure1: Modeled topologyandsimulatedsenario,N isvariable.

2.1 Evolution of ongestion window sizes

Imagineusershaveanotiationoflossesoftheformκ(t)inproportionoftheinomingaknowledgementow.

Denote byA(t) a funtion indiating theowevolutionof windows(for exampleA(t) = rtt(t)1 in usual TCP

models). ThequestionofdelaysandhowthefuntionsAandκevolveomelaterintheartile;thesequestions

arenotrelevanttostudytheintrinsiuserongestionwindowsizeevolution. Then,thedistributionofwindow

sizes isoftheform

D(t, w) =p(t, w)dw+M(t)δWmax.

Whih leadsto twoequations,thePDE:

1 A(t)

∂p

∂t(t, w) + ∂p

∂w(t, w) = (1)

κ(t) 4wp(t,2w)χw<Wmax/2−wp(t, w)χw<Wmax

Wmax

2 M(t)κ(t)Wmax

and

1 A(t)

dM

dt (t) =p(t, Wmax)−M(t)κ(t)Wmax. (2)

Intuitively, theoeientA(t) isthe inomingbandwidth; when itis small, thewindowsizes haveaslow

reation,when itislarge,theyreatinafasterway. Theoeientsχw<Wmax only indiatethat thewindow

size annotbelargerthanWmax. Whennolossours,theoeient ∂p

∂w(t, w)indiatesthat thewindowsize

inreaseslinearly. Whenlossesarise,theoeientκenables−κw(p(w)),whihmeansthataertainproportion

ofusersthatwereatwindowwhangeto anothervalueofthewindowsize;italsoenables4κwp(t,2w),whih

meansthat usersthat wereatwindow2wand2w+ 1 (or2w−1)moveto windoww.

2.2 Delay in the system

2.2.1 Limitsof Little-likeformula

Asnotiedin [16℄andin[40℄,theLittle-likeapproximationmadeinusualTCPmodels(forexample[7,15,35℄)

laks realism and strongly limits the way models an explain reality. This approximationonsists in saying

that attimet,thebandwidthB(t)ofauserisrelatedto theRTT,R(t)(Round-TripTime)anditsongestion

(9)

windowsizeW(t)andbyB(t) =W(t)/R(t). IftheRTTisalmostonstant(forinstanelosetothepropagation delay),itisaratheraeptablesimpliation,whereaswhenR(t)isvariable,themodelanleadtounaeptable

onsequenes: itis easytounderstand that whenonewantstostudythe stabilityofRED (withanon empty

queue),sayingB(t) =W(t)/R(t)orB(t)isonstantentailsdierentonlusions.

2.2.2 How to improve delaymodel

Theideaomesfrom[8℄whereasimpledelaylineisintroduedtostudythelimitbehavior(whenthebandwidth

tendstoinnity)ofoneuserimplementingMulTCPorsalableTCP([12,20,21℄). Althoughtheuseofadelay

line ompliates equations,the model isstill easy to simulate. Furthermore loal stability ofxed pointsan

bestudiedmathematially. Herewewilladaptdelayline modelingtolargenumberofTCPRenousers.

2.2.3 Delayequations

Delayandqueuesize LetusintrodueQ(t)thequeuesizemesuredinseonds,therouterissupposedFIFO

(in otherwords,Q(t)isthequeuingdelay). DenotebyK(t),thedestrution probabilityforapaket entering the queue; allBi(t)theinoming bandwidth tothe queueand Bo(t) theoutgoing bandwidth,saled by the

numberofusers. C istherouterapaityperuser. Then:

Bo(t) =

min(C, Bi(t))ifQ(t) = 0

C else. (3)

RTT CallR(t) = T+Q(τ(t)), the RTT virtually written bythe queue onpaket arriving the router τ(t).

This paket beomesanACKthat generatesnewpaketswhere thevalueisopied. Bydenition wesaythat

this valueomesbakattherouterrouterattimet,Whihmakest=τ(t) +R(t)leadingtotherelation:

R(t) =T+Q(t−R(t)). (4)

Wedisussedin[7℄thefat thatthisimpliitequationanalsobewritten: R(t+T+Q(t)) =T+Q(t),whih

doesnotraiseanydenitionissuesandiseasierfornumerialomputations.

Advane A(t)ofwindowsizes ThewindowsizeapproximatelyinreasesbyoneeveryW(t)arrivedpakets.

Theinomingbandwidthforagivenuseristheprobabilitythatthepaketisoneofhismultipliedbythetotal

bandwidth:

Wj(t−R(t)) PN

i Wi(t−R(t))Bo(t−T).

Infat wewantto omputetheadvane ofwindowsize,whihmeans thatdestroyednon-arrivingpakets

arryinformation. ThusthemodiedACKbandwidthis:

1 1−K(t−R(t))

Wj(t−R(t)) PN

i Wi(t−R(t))Bo(t−T).

Thefatorofadvaneweshalluseinwindowsizes evolutionequationsis:

A(t) = 1 1−K(t−R(t))

1

F(t−R(t))Bo(t−T), (5)

where F(t) = N1 P

iWi(t) =R

wp(t, w)dw representsthenumberof paketson-the-ight.

Loss rateindiator Itisgivenby:

κ(t) =K(t−R(t)). (6)

2.2.4 Bandwidthevolutionwhen rossing thereeiving user

Togofullirle 1

weneedtosaywhat isthevalueofthebandwidthBi(t)knowingthewindowsize evolutions

andtheACKbandwidthBo(t). TheevolutionofthisnumberonlyomesfromnewpaketsbeingsentorACK

beingreeived(ountingasACKtheindiationofalost paket);thus:

dF

dt (t) =Bi(t)− 1

1−K(t−R(t))Bo(t−T). (7)

1

AKA:givethelastequation

Références

Documents relatifs

In our numerical experiments, we solve high-dimensional instances of the dynamical optimal transport OT problem [12, 79], a prototypical instance of a potential MFG, and a mean

The fact that our solutions are stable for both the physical stability and the eductive stability backs up the mean field game approach to find rel- evant

The present work establishes the mean-field limit of a N-particle system towards a regularized variant of the relativistic Vlasov-Maxwell system, following the work of

The formula for the mean transmission rate is shown to boil down to the classical square root mean value formula for persistent flows when the average file size tends to

2014 A general analytic and very rapid method for the calculation of the mean square vibration amplitude of crystal atoms is given.. This method has been tested on a

More precisely, the local (resp. non-local) asymptotic stability of the origin (resp. global attractivity of a compact set) is ensured by a region-dependent small gain condition..

Bojak &amp; Liley also argue that anesthetic drugs reduce the firing rate of spontaneous action potentials in a relatively smooth dose-dependent manner and as a result, mean

Symmetric ground states for a stationary relativistic mean- field model for nucleons in the non relativistic limit. The concentration-compactness principle in the calculus