• Aucun résultat trouvé

From asynchronous to synchronous specification for distributed program synthesis

N/A
N/A
Protected

Academic year: 2021

Partager "From asynchronous to synchronous specification for distributed program synthesis"

Copied!
13
0
0

Texte intégral

(1)

HAL Id: hal-00339740

https://hal.archives-ouvertes.fr/hal-00339740

Submitted on 25 Nov 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

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

From asynchronous to synchronous specification for distributed program synthesis

Julien Bernet, David Janin

To cite this version:

Julien Bernet, David Janin. From asynchronous to synchronous specification for distributed program

synthesis. SOFSEM, Jan 2008, Slovakia. pp.161-173. �hal-00339740�

(2)

Speiations for Distributed Program Synthesis

JulienBernetandDavid Janin

LaBRI,UniversitédeBordeauxI

351,oursdelaLibération

33405TaleneedexFRANCE

{bernet|janin}labri.fr

Abstrat. Distributed games[7℄ allowthe speiation ofvariousdis-

tributed program synthesis problems. In suh games, a team of Pro-

essplayersompeteagainstauniqueEnvironmentplayer,eahProess

playingonitsownarena,withoutexpliitommuniationswithitsteam-

mates.

Thestandarddenitionofdistributedgamesallowssomedegreeofasyn-

hrony :theEnvironmentanplayonlyonpartofthearenas,therefore

onealingtosomeProessplayersthattheplayisgoingon.Whilethis

is onvenient for modeling distributed problems(espeially those that

themselvesmakeuseofasynhrony),thesegamesare bynomeanseasy

tomanipulate,andtheexistingonstrutionsaremuhmoretedious.

In this paper, we provide a uniform redution of any nite state dis-

tributedgame,synhronousorasynhronous,toasynhronous onewith

thesamenumberofplayers.Thisredutionisshowntobeorretinthe

sensethat it preservestheexistene of nitestate distributed winning

strategiesfor theteamofProesses. Itisuniforminthesensethatitis

designedfor arbitraryinputdistributed gamewithout any priorknowl-

edge aboutitssatisability.Thesizeoftheresultingsynhronousgame

isalsolinearinthesizeoftheoriginalasynhronousoneand,moreover,

thereisnoblowup ofthe sizeof thewinningstrategies. Additionalex-

petedpreservationproperties(e.g.informationows)are studied.Sur-

prisingly, it seems that nosuh redution exists for arbitrary (innite

state)strategies.

Introdution

Asynhronyin adistributed environmentanbedened in anabstrat way as

theabilityofsomeproessesoragentstoperformsomeationinthedistributed

systemswhileotherproessesoragentsnotonlyperformnoationbutareeven

not aware that they are staying idle. In other words, in a model where every

event in a proess (e.g. a loal ation or an inomingor outgoing message)is

triggeredbythetikofaloallok,asynhronyourswhenthereisnoglobal

lokonwhihloalloksaresynhronized.

However,if thememory of adistributed systemhas nitely many possible

states,itisommonlyunderstoodthatsynhronousandasynhronousbehaviors

(3)

eventsbetweenanytwosynhronousevents anbe bounded bythe numberof

globalstates.Morepreisely,the(presumablyasynhronous)distributedsystem

anbemodeled(atamoreabstratlevel)byanequivalentsynhronousone.

For instane, within a distributed system with an asynhronous message

passing mehanism, if the number of messages that have been sent but not

yet reeived is bounded, there annot be an unbounded number of (relevant)

asynhronouseventsbetweenanytwosynhronizations.

Of ourse asynhrony makes a dierene at the implementation level. For

instaneommuniationprotoolsdohavetoopewithasynhronytobeorret.

However,atsomemoreabstrat level,provided theglobalmemoryofasystem

is nite, it seems that,in some sense,asynhronous and synhronous systems

havethesameomputationalpower.

Themain objetiveofthispaperisto substantiatethisintuition.Morepre-

isely,inthesettingofdistributedprogramsynthesisproblem,weaimatproving

that(1)whenthedistributedenvironmentisnitestateand(2)whenoneon-

sidersonly nitestateprograms, anysynhronousorasynhronousdistributed

programsynthesisproblem is equivalentto asynhronousdistributed program

synthesisproblem.

Animmediate diultyforthis taskishowevertondtheappropriatefor-

malism.

Infat,there exists plentyof models of distributed systemsand behaviors;

onsider for instane the diversity of models of ommuniations between pro-

essesoragents:fromommuniationviaglobalsharedmemorymehanismsto

manytypesofmessagepassingsystems.Fordistributedprogramsynthesisspe-

iation,whereonenotonlyaimsatmodelingasinglebehaviorbut afulllass

of behaviors that may, or may not, fulll the design objetives, there may be

evenmoreformalisms.SeeforinstanePnueliandRosner'snotionofdistributed

arhiteture[10,4℄orWonham etal.'sdistributedontrol theory[5,6℄.

Inthispaper,wehoosetostudytherelationshipbetweensynhronousand

asynhronousdistributed synthesisproblemsinafairlyabstratthoughgeneral

model: themodelofdistributedgames[7℄.

Thesegames,designedafterPetersonandReif'spartialinformationgames[9℄,

relyon anexpliit modeling ofthe loal information- projetionofthe global

state-available toeveryagent(orproess)forguiding itsownbehaviors.Asa

matteroffat,noexpliitommuniationismodeled,heneforthnoommitment

has to be done in favor of suh or suh ommuniationmehanism.Moreover,

mostdistributedsynthesisproblemanbeenodedandsolvedwithindistributed

gamesproblem[2,7℄.

Inadistributed game, ateam of Proessplayersompete againstaunique

Environmentplayer,eahProessplayingonitsownarena-withitsownproje-

tionoftheglobalstateofthesystem-whileEnvironmentplayerhasaomplete

knowledgeoftheglobalstate.

Communiations are impliitly modeled as follows. In a distributed game

denition, one may restrit Environment player. From a given global state -

(4)

to reah someglobalstateswith moreinformativeloal projetionsthan other

globalstates.

Asynhrony,that indues anextra layerof partial information -absene of

globallok-ismodeledasfollows:inadistributedgame,theEnvironmentmay

playonlyonpartofthearenas,thereforeonealingtosomeProessplayersthat

theplayisgoingonelsewhere.

Inthispaperweprovethat, asfar asnitesystemsand programsare on-

erned,asynhronydoesnotyieldextraexpressivepower.

Morepreisely,weprovidearedutionof anyasynhronousnite statedis-

tributed game to a synhronous one with the same number of players. This

redutionisshowntobeorretinthesensethatitpreservestheexisteneof-

nitestatedistributedwinningstrategiesfortheteamofProesses.Thesizeofthe

resultingsynhronousgameisalsolinearinthesizeoftheoriginalasynhronous

oneand,moreover,thereisnoblowupofthesizeofthewinningstrategies.Addi-

tionalexpetedpreservationproperties(e.g.informationows)arealsostudied.

For strategies with unbounded memory, it is onjetured that there is no

suh a redution. Still; onemay ask whether suh a result ould be extended

to moregenerallassesofprogramswithinnitememorysuh as,forinstane,

pushdownstrategies.Thisremainsanopenproblem.However,evendeidability

questionsare yetnotsettledforthesemoregeneralstrategies.

1 Notations

ForanynitealphabetA,aword overAisafuntionw:N→Awhosedomain

isaninitialsegmentofN.Theonlywordsuhwithemptydomainisalledthe emptyword,anddenotedbyǫ.

Thefollowingnotations are used : A is theset of nite words(i.e. whose

domainis nite)overA, Aω istheset ofinnitewordsoverA, A=A∪Aω

is the set of niteand innite words overA, andA+ =A− {ǫ} is the set of

nitenon-emptywordsoverA.

Foranynite wordw, itslength |w| is theardinal ofits domain. Forany

wordsu∈ A and v ∈Aω, thewordu·v is theonatenation ofwordsuand v.Foranyintegerk,theworduk isbuiltbyonatenatingkopiesofu.When u6=ǫ,theinniteworduωistheinniteonatenationofuwithitself.

Theseoperationsareextendedtothelanguages.GivenX ⊆AandY ⊆A,

weusethenotationsX·Y ={u·v : u∈X, v∈Y},X0={ǫ},Xk ={w·u : w∈Xk−1, u∈X}foranyk >0,X=S

k≥0Xk and,whenǫ /∈X,Xω forthe

set of all wordsthat an be dened asaninnite produt ofwordsof X. The

followingadditionalnotationsarealso used :X+Y =X∪Y, X−Y =X\Y

andX?={ǫ} ∪X.

Givennsets X1, . . . , Xn,onsider theirdiretprodutX =X1× · · · ×Xn.

For every set I = {i1, . . . , ik} ⊆ {1, . . . , n}, onsider the natural projetion πI : X → Xi1 × · · · ×Xik dened by πI((x1, . . . , xk) = (xi1, . . . , xik). For any x∈X,wealsousethemoreonvenientnotationx[I] =πI(x).

(5)

Followingthesameidea,for anyset Y ⊆X, denotebyY[I] theset πI(Y).

If R ⊆ Xm is a m-ary relation over X, we should also write use R[I] = {(x1[I], . . . , xm[I]) : (x1, . . . , xm)∈R}.WhenIinanintervalofintegersofthe

form[i, j]weusethenotationsx[i, j], Y[i, j], R[i, j].WheneverIisreduedtoa

singleintegeri,these notationssimplify tox[i], Y[i], R[i].

Moreover, these notations are extended to words and languages : for any

wordw=x0·x1· · · ∈X,thenw[I] =x0[I]·x1[I]· · ·.

2 Distributed games

An-proessdistributedarenaisagamearenabuiltfromsomeprodutofnloal

standardgamearena.

Denition. GivennarenasGi =hPi, Ei, TP,i, TE,iifori∈[1, n]with disjoint

sets of Proess position Pi, sets of Environment position Ei, sets of Proess

moves TP,i ⊆ Pi×Ei and sets of Environment moves TE,i ⊆ Ei ×Pi, an n-

Proess distributed game arena builtfrom theloal game arenas {Gi}i∈[1,n] is

anygamearenaG=hP, E, TP, TEisuhthat:

Environmentpositions:E=Q

i∈[1,n]Ei,

Proesses positions :P =Q

i∈[1,n](Ei∪Pi)−Q

i∈[1,n]Ei,

Proesses moves : TP is the set ofall pairs(p, e)∈ (P×E)suh that,for i∈[1, n]:

eitherp[i]∈Pi and(p[i], e[i])∈TP,i (Proessi isativeinp),

orp[i]∈Ei andp[i] =e[i](Proessiisinativein p),

andEnvironmentmoves:TE issome subsetoftheset ofallpairs(e, p)∈ (E×P)suhthat,fori∈[1, n]:

eitherp[i]∈Pi and(e[i], p[i])∈TP,i(EnvironmentativatesProessi),

orp[i]∈Ei andp[i] =e[i](EnvironmentkeepsProessiinative).

WhenthesetTEofEnvironmentmovesismaximal,weallsuhanarenathefree asynhronousprodutofarenas{Gi}i∈[1,n]anditiswrittenG1⊗G2⊗ · · · ⊗Gn.

Remark. The essential idea behind this denition is to get a denition of

a multiplayer game in whih a team of Proesses ompete against a unique

Environmentto ahievesomeinnitary goal. Thefollowing point isimportant

: thisdenition allowstheEnvironmentto playonlyonasubsetofthearenas,

therefore hidingthat the play is going on to theProesses on whih arenas it

doesnotplay.Thiswill bereferredin thefollowingasasynhronousmove,and

allowstoenodeneatlymanydistributedsynthesisproblemsfromtheliterature,

suhas[10,4℄or[5℄.

Indistributed games, asynhronyours when Environment playerdeides

to keeponeormoreProess playersinative.A synhronousdistributed arena

anthusbedenedasfollows.

Denition. Ann-proessdistributedarenaG=hP, E, TP, TEiisasynhronous

distributedarena whenTE⊆E×Q

i∈[1,n]P[i].

Sineadistributedarenaisbuiltuponnsimplearenas,weneedadenition

tospeakaboutitsloalomponents:

Denition. Given a distributed arena G = hP, E, TP, TEi as above, given a

non empty set I⊆[1, . . . , n] wedene theanonial projetion G[I] of GonI

(6)

asthearenaG[I] =hP, E, TP, TEigivenby:P =P[I]−E[I](possiblysmaller

thanP[I]!),E=E[I],TP =TP[I]∩(P×E),andTE =TE[I]∩(E×P).

A distributed game arena is, at rst sight, a partiular ase of standard

disrete and turn base two player game arena. Standard notions of plays and

strategiesarestilldened.However,in ordertoavoidonfusionwithwhat may

happenintheloalarenaadistributed gameisbuild upon,weshallspeaknow

ofaglobalplayandglobalandloalstrategy.Partialinformationisthenaptured

bymeansofthenotionofloalview of play anddistributedstrategy.

Denition. Given an n-proess distributed arena G, a global play from an

initial position e ∈ E is dened as a path in G (seen as a bipartite graph)

emanatingfrompositionethatisbuiltalternativelybytheEnvironmentplayer andtheProessteam.

Morepreisely, from aurrentposition x∈ P ∪E, either x ∈ E and it is

Environment player turn to play by hoosing some position y ∈ P suh that (x, y) ∈ TE or x ∈ P and it is Proess team turn to play by hoosing some

positiony∈E suhthat(x, y)∈TP.

Aordingly,aglobal strategy fortheProessteam isapartial funtion σ : (E.P)+ → E suh that for every play of the form w.p ∈ dom(σ) with w ∈ E.(P.E) andp∈P onehas(p, σ(w.p))∈TP.

A play w is said ompatible with strategy σ when, for everyintegern ≥0

suhthat w[n]∈P onehasw[n+ 1] =σ(w[0, n]) wherew[0, n] isthe prexof woflengthn+ 1.

Denition. A gameG=hP, E, TE, TP, e0,Wiis agame arenahP, E, TE, TPi

equipped with an extra initial position e0 ∈ E and a distinguished set W ⊆ (E+P)ωalled innitaryonditionfortheProessteam.

Wesaythatglobalstrategyσisawinningstrategy fortheProessteamfrom

positione0∈E withonditionW wheneverymaximalplaysstartingine0 and

ompatible withstrategyσiseither niteandendsin anenvironmentposition or isinniteandbelongstoW.

Astrategywith nitememory fortheProessteamisgivenasatupleM= hM, m0, µ: M ×(P∪E) → M, h: M ×P → Ei, where M is a nite set of

memory states, m0 is the initial memory, µ is the update funtion, and h is

the hint funtion.The indued strategy σM : (E·P)+ → E is then dened,

foranyplayw·p∈(E·P)+,byσM=h(µ(m0, w), p)(whereµ is denedby µ(m, ǫ) =m,andµ(m, w·x) =µ(µ(m, w), x)foreverym∈M,w∈(E∪P), x∈(E∪P)).

In distributed games, it is intended that, within the Proess team, every

proesshasonlyapartialviewofaglobalplay.Notonlyeveryproessonlysees

itsownprojetionofeveryglobalpositions,but,whenidle,aproessisevennot

awarethat theplayisgoingon.Thisintentionisformallydenedasfollows.

Denition. Theloalview Proessihasofaglobalplayinadistributedgame

G is givenby the map viewi : (E·P)·E? → (Ei·Pi)·Ei? dened in the

followingway:

viewi(ǫ) =ǫ

viewi(x) =x[i]

(7)

viewi(w·x·y) =

viewi(w·x)ifx[i] =y[i]

viewi(w·x)·y[i]otherwise.

A play w ∈ (E·P)+ is said to be ative for Proess i when w ends in a

positionp∈P suhthat p[i]∈P[i].

Remark. Observe that in a synhronousdistributed arena, asexpeted, for

every play w ∈ (E.P).E? one has viewi(w) = w[i], i.e. the loal view of a

globalplayis justtheprojetionofthisplay.

Denition. A globalstrategy σfor theProess teamis adistributedstrategy when,foreveryi∈[1, n],thereisaproessstrategyσi : (E[i].P[i])+→E[i]in

the loalgame G[i], from nowonalled loal strategy forProess i, suhthat,

foranyplayoftheform w·p∈(E·P)+, giventhesetI ⊆ {1, . . . , n}of ative

proessesintheglobalProessespositionp,σ(w·p) =eifandonlyif

e[i] =σi(viewi(w)·p[i])fori∈I

e[i] =p[i]fori∈ {1, . . . , n} −I

Inthisase,wewriteσ1⊗σ2⊗ · · · ⊗σn forthedistributedstrategyσ.

Remark. Observethat when Gis synhronous, thedistributed strategy σ= σ1⊗ · · · ⊗σn ansimplybedenedforanyglobalplayw∈(E·P) by:

σ(w) = (σ1(w[1]), . . . , σn(w[n]))

Remark. Globalstrategiesarenotalwaysdistributed.Inpartiular,thereare

distributed games where Proess team has a winning global strategy, but no

winning distributedstrategy.Formoredetails,thereaderanreferto[7℄.

3 From asynhronous game to synhronous game

Weproveherethatevery(asynhronous)distributedgameisequivalentinsome

senseto asynhronousdistributed game.Morepreisely:

Theorem1. There exists a mapping that maps every distributed game G to

existsadistributedgameGe suhthattheProessteamhaveadistributedwinning strategy with nite memory in G if and only if they have one in Ge. Moreover,

game Ge has the same numberof Proess players asgame Gwith only alinear

inreaseofnumberof positions.Moreover,thismappingisdeneduniformlyon

distributedgames, bethemwinning for theproessteamor not.

Theremainingofthissetionisdevotedtotheproofofthisresult.

Let G = hP, E, T, e0,Wi be a n-proesses distributed game. First, we are goingtodesribethesynhronousgameGe,thenwewillshowthatitisequivalent

toGin termsofdistributed strategieswithnitememory.

For any set Ei, dene cEi as an equipotent set, suh that Ei ∩cEi = ∅;

for any e ∈ Ei, denote by ebi its image in cEi. Let Eb = Q

i∈{1,...,n}cEi and let

(8)

Pb=Q

i∈{1,...,n}Pbi(i.e.werestrittorelevantproesspositionsinasynhronous

game:proesspositionswhereeveryproessisative).

ConsiderGe=hP ,e E,e T , ee 0,Wif,whosepositionsare:

Pei=Pi ∪ Eci ; Efi=Ei (foralli∈ {1, . . . n})

For any position x ∈ P ∪E, denote by xb the position of Pe obtained by

replaingin eeahomponentfrom Ei withtheirimageinEci ,i.e.:

b x[i] =

dx[i]ifx[i]∈Ei

x[i]ifx[i]∈Pi

The funtion that maps any x ∈ P ∪E to its image xb in Pe is triviallya

bijetion.ThemovesofGearedened asfollows:

TfiP =TiP∪ {(be, e) :e∈Ei}

TfE={(e,p)b ∈Ee×Pe|(e, p)∈TE}

∪ {(e,be) :e∈E}

Thefuntionanel(Ee·P)e →(E·P)erasesanyasynhronousmovefroma globalplay:anel(ǫ) =ǫ,anel(w·e·p) =anel(w)·e·p,andanel(w·e·be) =

anel(w)(wherep∈P,e∈E).

This funtion is generalized to innite words by: anel(x0 ·x1 · · · ·) = limi→∞anel(x0· · ·xi)(it is aonverging sequene, in the sense of the prex

topologyoverwords,asdenedforinstanein [8℄).

Thewinning onditionofGe isthendened asfollows:

Wf=anel−1(W) ∪ (Ee·Pe)·(E·E)b ω

Remark. Theunderlying bipartitegraphofGisembedded intotheonefrom Ge. The graph of the arena of Ge is atually nothing more than the subgraph

induedbythisembedding,whereoneahpositionoftheenvironmentaloopof

size2hasbeenadded,orrespondingtoatotallyasynhronousmove.Moreover,

thewinning onditionWfisnotmuh moreompliated thanW :amongstthe

usualinnitary winningonditions(reahability,safety,parity,Muller,et....),

onlythesafetyonditioninnotpreservedbythis onstrution.

Lemma1 (From asynhronousto synhronous). Forany distributedand

winningstrategyfortheProessteaminG,thereisawinningdistributedstrategy for the Proess teamin Ge.

TheideaisatuallytoopythisstrategyonGe,ensuringintheproessthat

theProessplayersdonottaketheasynhronousmovesplayedontotheirarena

intoaount.

Foranyi∈ {1, . . . , n},letusdeneafuntionaneli: (fEi·Pei)→(Ei·Pi)

that erases the loal asynhronous moves: aneli(ǫ) = ǫ, aneli(w·e·p) =

Références

Documents relatifs

Human astrocytes expressing F3–T3 were also compared to human astrocytes that expressed kinase-dead F3–T3 (F3–T3(K508M)) or were transduced with an empty vector (Extended Data

We then evaluated the effects of IGF1 (Insulin-like growth factor 1), IGF2 (Insulin-like growth factor 2), EGF (Epidermal Growth Factor) and glargine (an insulin analogue used in

We looked for G ×G interactions by analyzing the fungal fitness (total number of spores on three plants from the same pot) and three traits used to estimate the success of

greatest efficiency during the whole cycle, without warm-up strategy. Schematic view of the distributed profile temperature inside a catalyst jointly with thermal exchanges...

Moreover, we apply this equation to the population dynamics of the spin-GORM model (a two-level subsystem interacting in a non-diagonal way with an environment described by

Dans certains cas, une amélioration clinique peut être observée sous analogues nucléos(t)idiques après une période de 3 à 6 mois, la transplantation peut

Although peer to peer metadata management based on Bloom filter can scale easily due to the loose coupling among MDSs, such strategies are generally probabilistic in terms of

X This work presents a systematic numerical method for modeling cellular materials in acoustics: (i) the local geometrical properties of ordered foams is idealized a PUC based on