curs 1 - babeș-bolyai universityvniculescu/didactic/ppd/c1.pdf · curs 1 - ppd - 10 limite ale...

Curs1

ProgramareParalelasiDistribuita

Curs 1 - PPD - 2

Continutulcursului(realizatsipebazapehttp://grid.cs.gsu.edu/~tcpp/curriculum/?q=home)

Teoretic•  Notiuniintroductive:arhitecturi,

concurenta,paralelism•  Etapeindezvoltareaprogramelorparalele•  Evaluareaperformanteiprogramelor

paralele•  Modeledeprogramareparalela

–  Diferentaintrecelebazatepememoriepartajatasimemoriedistribuita

•  Patterns–  Ptprogramareparalela–  Ptprogramaredistribuita

Practic•  Javathreads(lowlevelAPI)•  C++(>=C++11)threads•  High-levelAPI:pacheteJava->

java.util.concurrentpackages.•  Javastreams•  OpenMP(C++)•  CUDA(C++)•  MPI–MessagePassingInterface

–  exemplificariC,C++

Curs 1 - PPD - 3

ProblemaConcurs

Alternativalaexamenulscris•  Link:

https://drive.google.com/drive/folders/1qU77HrC-wG96x6EXtP2NeDs20hcpmecW?usp=sharing

•  Nuesteobligatoriu!•  Sepoateincercadarpentruvalidaretrebuierespectate

conditiile!

Curs 1 - PPD - 4

Bibliografie•  IanFoster.DesigningandBuildingParallelPrograms,Addison-Wesley1995.•  BernaL.Massingill,TimothyG.Mattson,andBeverlyA.Sanders,AddisonAPatternLanguagefor

ParallelProgramming.WesleySoftwarePatternsSeries,2004.•  MichaelMcCool,ArchRobinson,JamesReinders,StructuredParallelProgramming:Patternsfor

EfficientComputation,”MorganKaufmann,,2012.•  D.Culler,J.PalSingh,A.Gupta.ParallelComputerArchitecture:AHardware/SoftwareApproach.

MorganKaufmann.1998.•  Grama,A.Gupta,G.Karypis,V.Kumar.IntroductiontoParallelComputing,AddisonWesley,2003.•  D.Grigoras.CalcululParalel.Delasistemelaprogramareaaplicatiilor.ComputerLibrisAgora,2000.•  V.Niculescu.CalculParalel.Proiectaresidezvoltareformalaaprogramelorparalele.PresaUniv.

Clujana,2006.•  B.Wilkinson,M.Allen,ParallelProgrammingTechniquesandApplicationsUsingNetworked

WorkstationsandParallelComputers,PrenticeHall,2002•  A.Williams.C++ConcurrencyinActionPRACTICALMULTITHREADING.ManningPublisher.2012.•  TutorialeJava:http://docs.oracle.com/javase/tutorial/essential/concurrency/further.html•  C++11http://en.cppreference.com/w/•  OpenMP:http://openmp.org/•  MPI:http://www.mpi-forum.org/

Curs 1 - PPD - 5

Evaluare

•  Laborator(30%)>=5–  Exercitiilaborator–  ProgramefolosindJava/C++threads,OpenMP,MPI

•  Seminar(10%)–  Prezenta(7puncte<–7prezente)–  Participareactiva[-1(insuficient),0(suficient),1(bine),2(foartebine)]

•  Testpractic(multithreading)– 05%•  ProiectCUDA-10%•  Examen>=5

–  Scris– sesiune40%

•  Informatiicurs–  http://www.cs.ubbcluj.ro/~vniculescu/didactic/PPD/CursPPD.html

6Curs 1 - PPD -

ProcesareParalela

•  Uncalculatorparalelesteuncalculator(sistem)carefolosestemultipleelementedeprocesaresimultanaintr-omanieracooperativapentruarezolvaoproblemacomputationala.

•  ProcesareaParalelaincludetehnicisitehnologiicarefacposibilcalcululinparalel–  Hardware,retele,SO,biblioteci,limbaje,compilatoare,algoritmi…

•  Paralelismulestenatural.

•  PERFORMANTA–  Parallelismisverymuchaboutperformance!

7Curs 1 - PPD -

Curs 1 - PPD - 8

CalculSerialvs.Paralel(imagesfromIntroductiontoParallelComputingBlaiseBarney)

“Itwouldappearthatwehavereachedthelimitsofwhatitispossibletoachievewithcomputertechnology,althoughoneshouldbecarefulwithsuchstatements,astheytendtosoundprettysillyin5years.“(JohnvonNeumann,1949)

Curs 1 - PPD - 9

Curs 1 - PPD - 10

Limitealeprogramariiseriale•  Vitezadetransmisie–

•  Vitezaluminii(30cm/nanosecond),limitadetransmisiepefirdecupru(9cm/nanosecond).

•  Limitareaminiaturizarii–numardetrazistoripechip.–  LegealuiMoore:

număruldetranzistoricarepotfiplasatipeunsingurcircuitintegrat(persquareinchchip)sedubleazalafiecare2ani.

–  Darimpunecosturimari.

•  Limitarieconomice

Istoric

•  CrestereaperformanteiprocesorprincrestereafrecventeiceasuluiCPU(CPUclockfrequency)–  RidingMoore’slaw

•  Probleme:incalzireaputernicaachipurilor!–  Frecventaceasmaimare⇒consumelectricmaimare(Pentium4heatsink¦FryinganeggonaPentium4)

•  Solutie–adugaremaimultorcore-uriptaajungelaperformantadorita–  Sepastreazafrecventadeceaslafelsauchiarmicsorare–  nucresteconsumul.

Curs 1 - PPD - 11

Nivelurideparalelism

1.paralelismlaniveldejob:-intrejoburi;-intrefazealejoburilor;2.paralelismlaniveldeprogram:-intrepărţialeprogramului;-inanumitecicluri;3.paralelismlaniveldeinstrucţiune:

-intrediferitefazedeexecuţiealeuneiinstrucţiuni;4.paralelismlanivelaritmeticşilaniveldebit:-  intreelementealeuneioperaţiivectoriale;-  intrecircuitelelogiciiaritmetice.

Curs 1 - PPD - 12

•  Arhitecturilecurentesebazeazatotmaimultpeparalelismlanivelhardwarepentruaimbunatatiperformanta:

–  Multipleexecutionunits–  Pipelinedinstructions–  Multi-core

Curs 1 - PPD - 13

Paralelism<->Concurenta

•  Considerammaimultetaskuricaretrebuieexecutatepeuncalculator•  Taskurileseconsideraafipurparaleledaca:

–  Sepotexecutainacelasitimp(parallelexecution)•  Dependente->executieconcurenta:

–  Untaskarenevoiederezultatelealtora;–  Untasktrebuiesaseexecutedupaceoanumitaconditieeindeplinita–  Maimultetaskuriincearcasafoloseascaaceeasiresursa

=>Formedesincronizaretrebuiefolositepentruasatisfaceconditiile/dependentele

•  Concurentaestefundamentalaincomputerscience–  Sistemedeoperare,bazededate,networking,…

14Curs 1 - PPD -

Paralelismvs.Concurenta

15 Curs 1 - PPD -

Obs:–  Sepotfolosithreadurisauproceseinambelecazuri–  Dacauncalculparalelnecesitaacceslaresursecomuneatunci

estenevoiesasegestionezecorectconcurenta=>Paralelismulpoateimplicaconcurenta

Paralelism:Sefolosescmaimulteresursepentruarezolvaoproblemamairapid

resources

Concurenta:Gestiuneacorectasieficientaaaccesuluilaresursecomune

requestswork

resource

ConcurentasiParalelism

•  Concurentvs.paralel•  ExecutieParalela:

–  Taskurileseexecutaefectivinacelasitimp;

–  Estenecesaraexistentademultipleresursedecalcul

•  Paralelism=concurenta+hardware“paralel”

16Curs 1 - PPD -

Paralelism

•  Existamaimultenivelurideparalelism:–  Procese,threads,routine,instructiuni,…

•  Trebuiesafiesuportatederesurselehardware–  Procesoare,nuclee(cores),…(executiainstructiunilor)–  Memorii,DMA,retele,…(operatiiasociate)

17Curs 1 - PPD -

oabordaresimplista

Curs 1 - PPD - 18

http://www.java-programming.info/tutorial/pdf/java/11-Java-Multithreaded-Programming.pdf

Decesafolosimprogramareparalela?

•  Motiveprimare:–  Timpdecalculmairapid(responsetime)–  Rezolvareaproblemelor‘mari’decalcul(intimprezonabildecalcul)

•  Motivesecundare:–  Folosireaefectivaaresurselordecalcul–  Costurireduse–  Reducereaconstrangerilorasociatememoriei–  Limitarilemasinilorseriale

•  Paralelism=concurenta+hardware‘paralel’+performanta

19Curs 1 - PPD -

Curs 1 - PPD - 20

•  Rezolvareaproblemelordificile,mari:–  "GrandChallenge"(en.wikipedia.org/wiki/Grand_Challenge)problemsrequiring

PetaFLOPSandPetaBytesofcomputingresources.–  Websearchengines/databasesprocessingmillionsoftransactionspersecond

•  Folosirearesurselornon-locale:

–  SETI@home(setiathome.berkeley.edu)usesover330,000computersforacomputepowerover528TeraFLOPS(asofAugust04,2008)

–  Folding@home(folding.stanford.edu)usesover340,000computersforacomputepowerof4.2PetaFLOPS(asofNovember4,2008)

Directiiinprocesareaparalela

•  Arhitecturiparalele–  NecesitatiHardware–  Computersystemdesign

•  Sistemedeoperare(Paralelism/concurenta)•  Gestionareaaspectelorsistempentruuncalculatorparalel•  Programareparalela

–  Biblioteci(low-level,high-level)–  Limbaje–  Mediidedezvoltare–  Software

•  AlgoritmiParaleli•  Evaluareaperformanteiprogramelorparalele•  Testareavs.asigurareacorectitudinii•  Paralleltools:

–  Performanta,analize,vizualizare,…

21Curs 1 - PPD -

Decesastudiemprogramareparalela?

•  Arhitecturidecalcul–  Inovatiileconduclanoimodeledeprogramare

•  Convergentatehnologica–  “killermicro”estepestetot–  Laptop-urilesisupercomputeresuntfundamentalsimilare–  Trend-urileactualeconduclaconvergentaabordarilordiverse

•  Trenduriletehnologicefaccalcululparalelinevitabil–  Multi-coreprocessors!–  Acumoricesistemdecalculesteparalel

•  Intelegereaprincipiilorfundamentale!!!–  Programare,comunicatii,memorie,…–  Performanta

•  “Parallelismisthefutureofcomputing”-BlaiseBarney–  M.Andrews,J.S.Walicki.“Concurrencyandparallelism—futureofcomputing”in

ProceedingofACM'85Proceedingsofthe1985ACMannualconferenceonTherangeofcomputing:mid-80'sperspective.pp.224-231.

22Curs 1 - PPD -

InevitabilitateaProcesariiParalele

•  Cerinteleptaplicatii–  Necesitateauriasadecicluridecalcul

•  Trenduritehnologice–  Procesaresimemorie

•  TrenduriArchitecturale•  Factorieconomici•  Treduriactuale:

–  Today’smicroprocessorshavemultiprocessorsupport–  Serversandworkstationsavailableasmultiprocessors–  Tomorrow’smicroprocessorsaremultiprocessors–  Multi-coreisheretostayand#cores/processorisgrowing–  Accelerators(GPUs,gamingsystems)

23Curs 1 - PPD -

exemple…

•  ProcesorAMDRyzenThreadripper1950X~1300USD(2017)–  MemorieCache40MB–  Frecventaprocesor(MHz)3500–  TurboBoostpanala 4000MHz–  Numarnuclee 16Nuclee=>32Threads

•  Intel®Xeon®ProcessorE7v4Family~8000USD(2017)

–  #ofCores=24–  #ofThreads=48–  ProcessorBaseFrequency=2.40GHz–  MaxTurboFrequency=3.40GHz–  Cache=60MB

Curs 1 - PPD - 24

UBBCLUSTER–IBMIntelligentCluster•  Hybridarchitecture

–  HPCsystem+–  privatecloud

PPD-curs2 25

HPC – IBM NextScale

•  Rpeak 62 Tflops, Rmax 40 Tflops •  68 noduri NX360 M5, din care

–  12 nodes with 2 GPU Nvidia K40X, –  6 nodes with Intel Phi

•  2 processors E5-2660 v3 with 10Cores per node •  128 GB RAM per node, 2 HDD SATA de 500 Gb / node •  Subscription rate 1:1 between nodes based on Switch: IB Melanox SX6512

with 216 ports •  Storage NetApp E5660, 120 HDD SAS cu 600 Gb/Hdd => total 72Tb

–  IBM GPFS 4.x -parallel file system

•  IBM TS3100 Tape library for data archivation •  Operating systems on each node : RedHat Linux 6 with subscription •  Management Software: IBM Platform HPC 4.2

PPD-curs2 26

Private Cloud – IBM Flex System

•  10 virtualization servers Flex System x240 –  128 Gb RAM / server –  Procesoare 2 x Intel Xeon E5-2640 v2 / server –  2 x SSD SATA 240 Gb / server

•  1 management server •  Software for private cloud: IBM cloud manager with OpenStack 4.2 •  Software for monitorizing and management: IBM Flex System Manager

software stack •  Virtualization software: Vmware vSphere Enterprise 5.1

PPD-curs2 27

Mini cluster – IBM nestscale

~90cores

•  4nodesx2processorsX10cores•  1Managementnode•  IP:193.226.40.133

•  NOTA–  Rmax-MaximalLINPACK(benchmark)performanceachieved–  Rpeak-Theoreticalpeakperformance.

•  NodeperformanceinGFlops=(CPUspeedinGHz)x(numberofCPUcores)x(CPUinstructionpercycle)x(numberofCPUspernode)

28PPD-curs2

Aplicatiiinformaticeperformante

•  Performantaaplicatiilorimpunehardwareperformant(rapid,resursemultiple,etc)

•  Hardware-ulavansatgenereazanoiaplicatii•  Noileaplicatiiaucererideperformantamaimari

–  Crestereexponentialaaperformanteimicroprocesoarelor–  Inovatiiinarhitecturileparalelesiinintegrare

•  Cerintedeperformanta=>–  Performantasistemelortrebuiesaseimbunatateascainansamblu–  Schimbari/abordari/reevaluariinSoftwareengineering–  Costuri-technologieavansata

applicationsperformance

hardware

29Curs 1 - PPD -

Curs 1 - PPD - 30

Programareparalelavs.Programaredistribuita

INTERCONNECTION NETWORK

P2

P3

P1

P4 P5

Pn . . . .

Curs 1 - PPD - 31

TiPURI DE MULTIPROCESARE PARALLEL DISTRIBUTED

ASPECTE TEHNICE • PARALLEL COMPUTERS (- IN MOD UZUAL ) LUCREAZA BAZAT PE

•  CUPLARE STRANSA, •  in general bazate pe SINCRONICITATE, •  CU UN SISTEM DE COMUNICATIE FOARTE RAPID SI FIABIL •  Spatiu unic de adresare (intr-o masura mare)

•  DISTRIBUTED COMPUTERS •  MAI INDEPENDENTE, •  COMUNICATIE MAI PUTIN FRECVENTA SI mai putin RAPIDA (ASINCRONA) •  COOPERARE LIMITATA •  NU EXISTA CEAS GLOBAL • “Independent failures”

SCOPURI

•  PARALLEL COMPUTERS COOPEREAZA PENTRU A REZOLVA MAI EFICIENT PROBLEME DIFICILE

•  DISTRIBUTED COMPUTERS AU SCOPURI INDIVIDUALE SI ACTIVITATI PRIVATE. DOAR UNEORI INTERCOMUNICAREA ESTE NECESARA

PARALLEL COMPUTERS: COOPERARE IN SENS “POSITIV”

DISTRIBUTED COMPUTERS: COOPERARE IN SENS “NEGATIV” -- DOAR ATUNCI CAND ESTE NECESARA

Curs 1 - PPD - 32

Aplicatii paralele Suntem interesati sa rezolvam problemele mai rapid in paralel

Aplicatii distribuite

Suntem interesati sa rezolvam anumite probleme specifice :

• COMMUNICATION SERVICES ROUTING BROADCASTING

• MAINTENANCE OF CONTROL STUCTURE TOPOLOGY UPDATE LEADER ELECTION • RESOURCE CONTROL ACTIVITIES LOAD BALANCING MANAGING GLOBAL DIRECTORIES

Ingeneral…

Curs 1 - PPD - 33

Parallelv.s.DistributedSystems(fromM.FUKUDACSS434SystemModels)

Parallel Systems Distributed Systems

Memory Tightly coupled shared memory UMA, NUMA

Distributed memory Message passing, RPC, and/or used of distributed shared memory

Control Global clock control SIMD, MIMD

No global clock control Synchronization algorithms needed

Processor interconnection

Order of Tbps Bus, mesh, tree, mesh of tree, and hypercube (-related) network

Order of Gbps Ethernet(bus), token ring and SCI (ring), myrinet(switching network)

Main focus Performance Ex. - Scientific computing

Performance(cost and scalability) Reliability/availability Information/resource sharing

Unpunctdevedere…

Curs 1 - PPD - 34

SistemeleDistribuite

-potfifolositepentru-

•  Aplicatiidistribuiteimplicit–  BDDistribuite,rezervaribilieteavion/etc.sistembancar

•  Informatiipartajateintreuseri•  Partajareresurse•  Raportcost/performantamaibunptaplicatiiparalele

–  Potfifolositeeficientpt.aplicatiicugranularitatemare(coarse-grained)si/sauptaplicatiiparaleledetipembarrassinglyparallelapplications

•  Fiabilitate(Reliability).•  Scalabilitate

–  Cuplareslaba(Looselycoupledconnection);hotplug-in

•  Flexibilitate–  Reconfiguraresistemptaintrunicerintele

Curs 1 - PPD - 35

Performanta/Scalabilitate

Spredeosebiredesistemeleparaleleceledistribuiteimplica:-  mediumaiputinrapiddetransferaldatelor(reteamaiputinrapida)-  HeterogenitateSolutii:-Procesarebatchamesajelor:

•  SeevitainterventiaSOptfiecaretransferdemesaj.–  Cachedata

•  Seevitarepetareatransferuluiaceleiasidate–  Evitateaentitatilorsiaalgoritmilorcentralizati

•  Evitareasaturariiretelei–  Realizareoperatii“post”lanivelulclientului

•  Evitareatraficuluiintensintreclientisiservere

–  ….

Curs 1 - PPD - 36

Securitate

•  Nuexistadoarunsingurpunctdecontrol

•  Probleme:–  Mesaje,furate,modificate,copiate,…

•  Solutie:folosireCriptografie–  Failures

•  FaultTolerancesolutions

Tipuridesistemedistribuite(Munehiro Fukuda – PDP Fundamentals)

•  Modele–  Minicomputer–  Workstation–  Workstation-server–  Processor-pool–  Cluster–  Gridcomputing

Curs 1 - PPD - 37

38

ModelulMinicomputer

•  ExtensieasistemuluideTimesharing–  Useriitrebuiesaseloghezepepropriulminicomputer.–  Apoiselogheazalaaltamasina(remotemachine)prinprogramede

tiptelnet.•  Partajareresurse

–  DB–  High-performancedevices

Mini-computer

Mini-computer

Mini-computer

net

Curs 1 - PPD -

Curs 1 - PPD - 39

ModelulWorkstation

•  MigrareProcese–  Useriiselogheazamaiintaipestatiadelucrupersonala;–  Dacaexistastatii“inasteptare”–unjob“mare”poatemigrala

unadintreele.•  Probleme:

–  Cumseidentificastatiile“inasteptare”(idle)?–  Cummigreazaunjob?–  Ceseintampladacaunaltuserselogheazapemasinafolosita?

100MbpsLAN

Workstation

Workstation Workstation

WorkstationWorkstation

Curs 1 - PPD - 40

ModelulWorkstation-Server

•  StatiiClient–  AplicatiileGrafice/interactiveseproceseazalocal–  Ptaltecereridecalculsetrimitcererilaservere.

•  Servere(minicomputers)–  Fiecareserverestededicatunuiasaumaimultor

tipurideservicii.•  Modeldecomunicare:Client-Servermodel

–  RPC(RemoteProcedureCall)–  RMI(RemoteMethodInvocation)

•  Unprocesclientcheamaofunctieaprocesuluiserver.

•  Nusefacemigraredeprocese•  Examplu:NFS

100GbpsLAN

Workstation

Workstation Workstation

Mini-Computerfileserver

Mini-Computerhttpserver

Mini-Computercycleserver

Curs 1 - PPD - 41

ModelulProcessor-Pool

•  Clientii:–  Selogheazalaunterminal–  Toateserviciilesuntgestionatede

catreservere.•  Servere:

–  Ptfiecareusersealocanrnecesardeprocesoaredinpool.

•  Utilizarebunadarinteractivitateslaba.

Server1

100MbpsLAN

ServerN

Curs 1 - PPD - 42

ClusterModel

•  ConstainmaimultePC/workstationsconectatelaoreteadetiphigh-speed.

•  Focuspeperformanta.100Mbps

LAN

Workstation

Workstation Workstation

Masternode

Slave1

SlaveN

Slave2

1GbpsSAN

httpserver1

httpserver2

httpserverN

Curs 1 - PPD - 43

High-speedInformationhighway

GridComputing

•  Scop–  Colectareaputeridecalculamaimultor

supercomputeresiclusteredispersategeografic

•  DistributedSupercomputing–  Ptproblemwfoartemari.Dificile.

(CPUintensive,memoryintensive).•  High-ThroughputComputing

–  Folosireamultorresursecarenusuntfolosite

•  On-DemandComputing–  Resurseladistantaintegrateincalculullocal

•  Data-intensiveComputing–  distributeddata

•  CollaborativeComputing–  Suportptcomunicareintremaimulteparti

Super-computer

Cluster

Super-computer Cluster

Mini-computer

Workstation

Workstation Workstation