• Aucun résultat trouvé

The stable configuration in acyclic preference-based systems

N/A
N/A
Protected

Academic year: 2021

Partager "The stable configuration in acyclic preference-based systems"

Copied!
27
0
0

Texte intégral

(1)

HAL Id: inria-00318621

https://hal.inria.fr/inria-00318621

Submitted on 4 Sep 2008

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.

systems

Fabien Mathieu, Gheorghe Postelnicu, Julien Reynier

To cite this version:

Fabien Mathieu, Gheorghe Postelnicu, Julien Reynier. The stable configuration in acyclic preference-

based systems. [Research Report] RR-6628, INRIA. 2008, pp.26. �inria-00318621�

(2)

a p p o r t

d e r e c h e r c h e

N0249-6399ISRNINRIA/RR--6628--FR+ENG

Thème NUM

The stable configuration in acyclic preference-based systems

Fabien Mathieu — Gheorghe Postelnicu — Julien Reynier

N° 6628

September 2008

(3)
(4)

Unité de recherche INRIA Rocquencourt

Domaine de Voluceau, Rocquencourt, BP 105, 78153 Le Chesnay Cedex (France)

Fabien Mathieu

, Gheorghe Postelniu

, Julien Reynier

ThèmeNUMSystèmesnumériques

ProjetsGANG

Rapportdereherhe 6628September200823pages

Abstrat: Aylipreferenesreentlyappearedasanelegantwayto modelmany distributedsystems. An

ayli instane admits a unique stable onguration, whih an reveal the performane of the system. In

this paper,wegivethestatistial propertiesof thestableongurationforthreelassesofaylipreferenes:

node-basedpreferenes,distane-basedpreferenes,andrandomaylisystems. Usingrandomoverlaygraphs,

we prove using mean-eld and uid-limit tehniques that these systems have an asymptotially ontinuous

independent rank distribution for a proper saling, and the analytial solution is ompared to simulations.

These results provide a theoretial ground for validating the performane of bandwidth-based or proximity-

basedunstruturedsystems.

Key-words: Ayliity,rankdistribution,uidlimit,mean-eld,small-worlds,PDE

SupportedbyCRCMARDIII

SupportedbyANRProjetShaman

OrangeLabs

Google

ReehAiMResearh

(5)

Résumé : Les systèmes à préférenes ayliques sont réemment apparus omme une méthode élégante

de modélisation de ertainssystèmes ditribués de typepair-à-pair. Uneinstane ayliqueadmet une unique

onguration stable, auto-stabilisante, qui donne une bonne indiation du omportement du système. Dans

e rapport, nousdonnons ladistributionstatistique dela ongurationstable pourtrois typesdepréférenes

ayliques: lespréférenesglobales(baséessurunordretotaldesn÷uds), lespréférenesdedistane (leplus

proheestpréféré),etlespréférenesayliquesaléatoires. Sousl'hypothèsed'ungraphedeompatibilitéErdös-

Rényi,nousmontronsàl'aidedetehniquesdelimitesuideset dehampmoyenl'existened'unedistribution

limite ontinue. Lapertinenedesrésultatsest vériéeàl'aidedesimulations.

Mots-lés : Systèmesayliques,distribution,limite uid,hampmoyen,petit-mondes,EDP

(6)

Contents

1 Introdution 4

1.1 Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.2 Roadmap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Model and notation 5 3 Ayli formulas 5 3.1 Generiformula. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

3.2 Mean-eldapproximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

4 Node-basedpreferenes 6 4.1 Fluidlimit. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

4.2 Constantdegreesaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

4.3 Convergenetheorem. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

4.4 Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

4.5 Exatresolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

5 Ayli and distane-based preferenes 9 5.1 Completerankdistribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

5.1.1 Fluidlimit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

5.1.2 Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

5.2 Distanedistribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

5.3 Aeptablerankdistribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

6 b-mathinggeneralization 13 6.1 MeanFieldformulas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

6.2 Fluidlimits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

6.3 Disussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

6.3.1 Stratiationtrade-o . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

6.3.2 Small-Worldeet ingeometripreferenes . . . . . . . . . . . . . . . . . . . . . . . . . . 15

7 Conlusion 16 A Proofof Theorem1 18 A.1 Uniformonvergene . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

A.2 PDE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

A.3 Resolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

B Exat resolutionofthe node-based stableonguration 21 B.1 Reursiveformula. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

B.2 Uniformonvergene . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

B.3 PDE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

(7)

1 Introdution

Mathing problems with preferenes haveappliations in a variety of real-worldsituations, inluding dating

agenies, ollege admissions, roommate attributions, assignment of graduatingmedial students to theirrst

hospitalappointment,orkidneyexhangesprograms[8,9,10,19,20℄.

Reently,mathingproblemsalsoappearedasanelegantwaytomodelmanydistributedsystems,inluding

ad-hoandpeer-to-peernetworks[12,6,16,15,7℄. Indistributedsystems,thepreferenesgenerallyomefrom

diret measurements. Thosemeasurementsanbenode-related(CPU,upload/downloadbandwidths,storage,

battery, uptime), oredge-related (Round-Trip Time, physial/virtual distanes, link apaity, o-uptime). In

mostases,theresultingpreferenesareayli: thereannotexistayleofmorethantwonodessuhthateah

nodeprefersitssuessortoitspredeessor. Asaonsequene,therealwaysexistsauniquestableonguration,

whihisself-stabilizing[6,1℄. Thismakesthingsmuheasierthaninothermathing problems,wherending,

ountingand omparingthestable ongurationsaresomeofthemainissues[8,21,17,20℄.

Modeling distributedsystemswith aylipreferenes allowsus to predittheeetiveollaborationsthat

will our,whih, inturn, allowsusto infertheperformaneof agivensystem. Foronveniene, thestudyof

anaylidistributedsystemisoftensplitintotwomain problems:

ˆ Howfastisthestabilizationproess? Beausedistributedsystemsareoftenhighlydynami,withonstant

hurnandpreferenealteration,thespeedofonvergeneanbeusedtodeterminehowfartheeetive

ongurationsarefromthetime-evolvingstableonguration.

ˆ Whatarethepropertiesofthe stableonguration? Ifthestabilizationproessisfastenough,theeetive

and stable ongurationswill be lose. Analyzing the latteran then givevaluable information on the

former.

Inapreviouswork,Mathieuinvestigatedtherstquestion[16,15℄. Heprovedthat eveniftheonvergene

an be prohibitive under an adversary sheduler, it is fast for realisti senarios. The seond question has

been answeredfor spei aylipreferenes: for real-world lateny-based preferenes, the stable ongura-

tion shows, for b-mathing(several mates per nodes allowed), small-worldproperties (low diameter and high lusteringoeient)[6℄;fornode-relatedpreferenes,thestableongurationtendstopairnodeswithsimilar

values[7℄: thisisthestratiationeet,whihallows,forinstane,tounderstandupload/downloadorrelations

in inentivenetworkslikeBitTorrent[4℄.

1.1 Contribution

Thestudiesproposedin[7℄and[6℄gaveonlypartial,mostlyempirial,answersaboutthelinkdistributioninthe

stableonguration,andproposedsomeonjetures. Thegoalofthispaperistoompleteandgivetheoretial

proofsontheshapeofthestableonguration.

Weextendtheseminalresultsthatweregivenin[7℄fornode-basedpreferenes: forb= 1(simplemathing

ase),weprovetheexisteneofalimitontinuousdistribution andsolvetheorrespondingPartialDierential

Equation(PDE).Thenweapplyasimilarmethodfordistane-basedandrandom-aylipreferenes,andalso

givetheexpliitsolutionoftheorrespondingPDE.

Lastly,weextendtheresultsforb >1(multiple mathings). Inthatase,thereisnosimpleexpressionthat

givestheexatsolutionofthePDEs system,but disreteequationsareusedtoobserveasymptotialbehavior

ofthedistribution. Fornode-basedpreferenes,theexponentialbehaviorvalidatesthestratiationeet (the

probabilitytobemathedwithadistantpeer dereasesexponentiallywiththedistane),while thepowerlaw

obtainedforthetwootherasesindiatesthatthesmallworldeetobservedin[6℄forlatenyisinfatommon

to alldistane-basedpreferenes 1

.

1.2 Roadmap

InSetion 2wedenethe modeland notationforpreferene-basedsystems. Setion 3givesthegenerimean

eld methodusedin thispapertosolvethesimplemathingase. Theaseofnode-basedpreferenesissolved

in Setion 4,then the resultsare adapted to thedistane-based andrandom-ayli preferenesin Setion 5.

Setion 6 extend the formulas to multiple mathings, and asymptotial properties of the distributions are

desribed. Lastly,Setion7onludes.

1

Lateniesannotbeonsideredasrealdistanes,mainlybeausethetriangularinequalityisnotalwaysveried. However,they

formaninframetri,whihisnotoofarfromarealmetri[13℄.

(8)

2 Model and notation

A preferene-based systemisaset V ofN nodes, whosepossibleinterations aredesribedbyan aeptane graphG,amarkmatrixmandaquotavetorb.

Thequotavetorb limitstheollaborations: apeeriannothavemoreb(i)simultaneous mates.

The aeptane graphG = (V, E) is an undireted, non-reexivegraph. It desribes allowed mathings:

a node i and anode j an be mated (we say that i is aeptable for j, and vie versa) if, and only if (i) {i, j} ∈E. Forinstane,in peer-to-peernetworks,anodeannotbediretlyonnetedtoallotherpeersofthe system, beauseof salability, andpeers that are notdiretly onneted annot be mated. Inthis paper,we

onsider Erdös-RényigraphsG(N, p)(eahpossibleedgeexists withprobabilitypindependentlyoftheothers;

henetheexpeteddegreeis d=p(N−1)).

Themarkmatrixmisusedto onstrutthepeers'preferenes: giventwonodesj and kaeptablefori,i

ranksj betterthankimi,j < mi,k (thesignisarbitrary). Thefollowingmarksareonsidered inthispaper:

Node-based m(., i)isonstant(nodeshaveintrinsivalues). These preferenesaresuitedtomodelingpeer-

relatedperformane,likeaess bandwidth,storage,CPU,uptime...

Geometri the nodes are assoiated to N points piked uniformly at random on a n-dimensional torus (n≥1). Themarksare thedistanes betweenthose points. These preferenes allowatheoretial analysisof

proximity-basedperformane.

Meridian latenies we onsidered random subsets of N nodes taken from the 2500 nodes dataset of the

Meridian Projet [18℄. The marks are the (symmetri) latenies between those nodes. We do not perform

analysis forthosemarks,but usetheminŸ6.3.2forvalidating thegeometriapproah.

Randomayli eahedgereeivesarandomuniformlydistributed value. Thenameisjustiedbeauseall

aylipreferenesanbedesribedbymarksontheedges(whihisequivalenttoassumethatmissymmetri).

Hene uniformly distributed (symmetri) randommarksare aonvenientwayto perform auniformsampling

oftheaylipreferenes[6,1℄.

Alltheonsideredmarksareayli,andthereforea(G, m, b)systemadmitsauniquestable onguration C ∈E, whih is self-stabilizing[12, 16℄. The neighbors of i in C arethe stable mates of i, and the notation i↔j isusedtoexpressthat iandj arestablemates.

WeassumeforsimpliitythatmisompleteandnotlimitedtotheedgesofG. Forallonsideredpreferenes

butrandomayli,theompletionisstraightforward.Forrandomaylipreferenes,weassumethatdummy

randomvaluesareassignedtonon-aeptableedges.

The preferenes are denoted like follows: ifj is aeptable for i, ri(j) denotes therank of j in i's list (1

beingthe best). ri isalled the aeptable rankingof i. If i hasmorethan k aeptableneighbors, ri−1(k)is

thekthnode ini'saeptableranking. Similarly,forj 6=i, Ri(j)denotestherankofj in theomplete graph

(the aeptabilityonditionisomitted). Ri isalledtheomplete rankingofi ForK < N,R−1i (K)istheKth

nodeini'sompleteranking.

AllstablematingprobabilitiesthataredisussedinthisartilearedesignedbyD. Subsriptsandarguments

areusedto preisethemeaningofD wheneverneeded. Forinstane:

ˆ DRi(K)istheprobabilitythati hasastablematewithompleterankK.

ˆ DN,d(i, j) is theprobabilitythat i↔ j, knowing there isN nodes and that theexpeted degree of the

aeptanegraphisd.

ˆ forc≤b(i,),Dri,c(k)istheprobabilitythat thecthstablemate ofihasrelativerankk.

ˆ ...

The omplementary umulativedistributionfuntion (CCDF) ofD is denoted S, andthesaled versionofD

and Sare denotedDandS.

3 Ayli formulas

We rst onsider the ase b = 1 (simple mathing) (the results will be extended to multiple mathings in

Setion 6). Wegiveageneriformulathat desribestheompleterankofthemate C(i)ofapeeri.

(9)

3.1 Generi formula

Let DRi(K) be the probability that Ri(C(i)) = K (the probability that the mate of i, if any, has rank K).

The CCDF of D is SRi(K) := 1−PK−1

L=1, whih is the probability that i'smate hasa rankgreater than K

(Ri(C(i))≥K)orhasno mate (short notation: Ri(C(i)) ≮K). Following theapproah proposed in [7℄, we

rstgiveageneriexatformulathat desribesDRi,thenweproposeasimplied mean-eldapproximation.

InordertosolveDRi(K),onean observethatiismatedwithitsKthpeerj=R−1i (K)i:

ˆ {i, j}isanedgeoftheaeptanegraph;thishappenswithprobabilitypasGissupposedtobeaG(N, p)

graph.

ˆ iisnotmatedwithanodebetterthanj (Ri(C(i))≮K);

ˆ j isnotmatedwithanodebetterthani(Rj(C(j))≮Rj(i)).

Thisleadsto thefollowingexatformula:

DRi(K) =pP(Ri(C(i))≮K)×

×P(Rj(C(j))≮Rj(i)|Ri(C(i))≮K)

=pSRi(K)P(Rj(C(j))≮Rj(i)|Ri(C(i))≮K)

(1)

3.2 Mean-eld approximation

Solving (1) is diult to handle, mainly beause of possible orrelations between Rj(C(j)) ≮ Rj(i) and Ri(C(i))≮K. Thesolutionistoadoptameaneld assumption:

Assumption 1 Theevents nodeiisnotwithanodebetterthanj and nodej isnotwithanodebetterthan i areindependent.

Thisassumptionhasbeenproposedin[7℄tosolve(1)intheaseofnode-basedpreferenes. Itisreasonable

whenN islargeandpissmall. Then(1)anbeapproximatedby

DRi(K) =pSRi(K)SRj(Rj(i)). (2)

Now,inthenexttwosetions,weproposeto solveEquation2forspeipreferenes.

4 Node-based preferenes

We assumeherethat the preferenesomes from markson nodes. This is equivalentto assume atotalorder

among thenodes. Thereforewedonotneed to expliitthemarkmatrix m, andweanusean orderednode

labelinginstead. Wearbitraryhoose1, . . . , N aslabels,1 beenthebest(if 1is rankedrstforallnodesthat

aept1,andsoon...).

Beausethenodes'labelexpresstheirompleteranks,weandiretlyonsiderD(i, j),theprobabilitythat nodeiismatedwithnodej. Nodej hasrankj foriifj < i,andj−1 ifj > i,beausearankdoesnotrank

itself. ThisgivestherelationbetweenD andDR:

D(i, j) =

DRi(j)ifj < i,

0ifj=i(matingisnotreexive), DRi(j−1)ifj > i.

(3)

UsingtheCCDFS(i, j) := 1−Pj−1

k=1D(i, k),wegetthenode-basedversionofEquation2:

D(i, j) =

0ifi=j,

pS(i, j)S(j, i)otherwise. (4)

Thisequation,whihwasoriginallyproposedin [7℄,whihalsoshowthatitgivesaverygoodapproximationof

empirialdistribution. It anbenumeriallysolvedbyusingadoubleiteration.

Références

Documents relatifs

Dynamis of preferene systems We have onsidered xed aeptane graph and preferene lists.

Abstract This paper introduces a design approach based on system analysis and game theory for the identification of architectural equilibrium which guarantees the

The VIRGO interferometer is basically a Michelson interferometer (MITF), with two 3 km long, perpendicular arms. The large length of 3 km is chosen mainly to decrease

E : non je n’ai rien remarqué de particulier c’est plus une tendance j’ai envie de dire élèves ayant de la facilité par rapport aux autres la séparation serait plutôt à

Dans la suite de ce travail, nous avons réalisé une transposition didactique du plan d’études pour créer un plan de cours d’informatique destinés à tous les élèves en

A second treatment option aimed at early, individual optimization of blood coagulation after major trauma is to assess each trauma patient’s coagulation status on admission in

We directly prove that on the one hand the operator is smooth with respect to all variables including the parameters and on the other hand it is a contraction mapping with respect

A fragmentation process is a Markov process which describes how an object with given total mass evolves as it breaks into several fragments randomly as time passes. Notice there may