Agenda. ProActive Parallel Suite. 1. Background: INRIA, ActiveEon 2. ProActive Open Source: Programming, Scheduling, Resourcing

Transcription

1 Des Multi-Cores aux Clouds avec ProActive Parallel Suite Agenda D. Caromel, et al. 1. Background: INRIA, ActiveEon 2. ProActive Open Source: Programming, Scheduling, Resourcing 3. Cloud Seeding with GPU 4. ProActive PACA GRID: Cloud Portal with GPUs Parallelism is Complex: Hiding Infra Details and Seamless Integration with Multi-Cores and GPU

2 1. Background 1. Background 2 2

3 OASIS Team & INRIA A joint team, Now about 35 persons 2004: First ProActive User Group 2009, April: ProActive 4.1, Distributed & Parallel: From Multi-cores to Enterprise GRIDs 3 3

4 OASIS Team Composition (35) Researchers (5): D. Caromel (UNSA, Det. INRIA) E. Madelaine (INRIA) F. Baude (UNSA) F. Huet (UNSA) L. Henrio (CNRS) PhDs (11): PostDoc (1): Regis Gascon (INRIA) Engineers (10): Elaine Isnard (AGOS) Fabien Viale (ANR OMD2, Renault ) Franca Perrina (AGOS) Germain Sigety (INRIA) Yu Feng (ETSI, FP6 EchoGrid) Bastien Sauvan (ADT Galaxy) Florin-Alexandru.Bratu (INRIA CPER) Igor Smirnov (Microsoft) Fabrice Fontenoy (AGOS) Antonio Cansado (INRIA, Conicyt) Brian Amedro (SCS-Agos) Cristian Ruz (INRIA, Conicyt) Elton Mathias (INRIA-Cordi) Imen Filali (SCS-Agos / FP7 SOA4All) Open position (Thales) Marcela Rivera (INRIA, Conicyt) Trainee (2): Muhammad Khan (STIC-Asia) Paul Naoumenko (INRIA/Région PACA) Viet Dung Doan (FP6 Bionets) Virginie Contes (SOA4ALL) Assistants (2): Guilherme Pezzi (AGOS, CIFRE SCP) Patricia Maleyran (INRIA) Located in Sophia Sandra Devauchelle Antipolis, (I3S) between Nice and Cannes, Visitors Welcome! + Visitors + Interns Etienne Vallette d Osia (Master 2 ISI) Laurent Vanni (Master 2 ISI) 4 4

5 Startup Company Born of INRIA Some Customers: Some Partners: Co-developing, Support for ProActive Parallel Suite Worldwide Customers: Fr, UK, Boston USA 5 5

6 2. ProActive Parallel Suite 6 6

7 Product: ProActive Parallel Suite Java Parallel Multi-Platform Resource Toolkit Job Scheduler Manager Used in Production Today: 50 Cores 300 Cores early 2010 Strong Differentiation: Java Parallel Programming + Integration + Portability: Linux, Windows, Mac + Versatility: Desktops, Cluster, Grid, Clouds = Perfect Flexibility 7

8 ProActive Programming: Active Objects 8 8

9 ProActive Programming View GPU nodes 9 9

10 ProActive Programming View GPU nodes 10 10

11 TYPED ASYNCHRONOUS GROUPS 11 11

12 Broadcast and Scatter Broadcast is the default behavior Use a group as parameter, Scattered depends on rankings cg ag JVM c1 s c2 c3 c1 c2 c3 JVM JVM ag.bar(cg); // broadcast cg ProActive.setScatterGroup(cg); ag.bar(cg); g) // scatter cg JVM

13 Dynamic Dispatch Group cg ag Slowest c0 c1 c2 c3 c4 c5 c6 c7 c8 c9 JVM Fastest JVM c0 c2 c4 c6 c8 c1 c3 c5 c7 c9 JVM ag.bar(cg); JVM

14 Abstractions for Parallelism The right Tool to do the Task right

15 ProActive Parallel Suite Workflows in Java Master/Workers SPMD Components 15 15

16 Object-Oriented SPMD 16 16

17 Key Point MPI and programming languages from the 60 s will not make it Jack Dongarra, 2/13/2009, Wake Forest University talk It is time to get ride of MPI Alan Edelman, MIT, 06/16/2010, Ter@tec ScilabTec 10 Users Day (30 mn. ago)

18 OO SPMD: Object-Oriented SPMD A ag = newspmdgroup ( A, [ ], VirtualNode) // In each member mygroup.barrier ( 2D ); // Global Barrier mygroup.barrier ( vertical ); // Any Barrier mygroup.barrier ( north, south, east, west ); A

19 IC2D: Optimizing 19 19

20 IC2D 20 20

21 IC2D 21 21

22 ChartIt 22 22

23 Pies for Analysis and Optimization 23 23

24 Video 1: IC2D Optimizing Monitoring, Debugging, Optimizing 24

25 Scheduling & Resourcing 25 25

26 ProActive Scheduling

27 ProActive Scheduling Big Picture Multi-platform Graphical Client (RCP) ProActive Scheduler ProActive Resource Manager File-based or LDAP authentication Static Workflow Job Scheduling, Native and Java tasks, Retry on Error, Priority Policy, Configuration Scripts, Dynamic and Static node sources, Resource Selection by script, Monitoring and Control GUI, RESOURCES ProActive Deployment capabilities: Desktops, Clusters, Clouds, 27

28 Workflow Example : Picture Denoising Split Denoise Denoise Denoise Denoise Merge Job with selection on native executable availability (ImageMagik, GREYstoration) Multi-platform selection and command generation with file transfer in pre/post p scripts 28

29 ProActive Resourcing

30 RESOURCING User Interface 30 30

31 Clusters to Grids to Clouds: e.g. on Amazon EC

32 Node source Usecase : Configuration for external cloud with EC2 ProActive Scheduler ProActive Resource Manager Static ti Policy Timing Policy Dynamic 12/24 Workload Policy LSF Desktops EC2 Dedicated resources Desktops Amazon EC2 32

33 Video 2: Scheduler, Resource Manager 33

34 3. Cloud Seeding with GPU 34

35 Cloud Seeding with ProActive Amazon EC2 Execution Cloud Seeding strategy to mix heterogeneous computing resources : External GPU resources 35

36 Cloud Seeding with ProActive CPU nodes Web Interface ProActive Scheduler + Resource Manager User Amazon EC2 Noised video file GPU nodes 36

37 Cloud Seeding with ProActive CPU nodes Web Interface ProActive Scheduler + Resource Manager User Amazon EC2 User submit its noised video to the web interface GPU nodes 37

38 Cloud Seeding with ProActive CPU nodes Web Interface ProActive Scheduler + Resource Manager User Amazon EC2 Web Server submit a denoising job the ProActive Scheduler GPU nodes 38

39 Cloud Seeding with ProActive CPU nodes Web Interface ProActive Scheduler + Resource Manager User Amazon EC2 CPU nodes are used to split the video into smaller ones GPU nodes 39

40 Cloud Seeding with ProActive CPU nodes Web Interface ProActive Scheduler + Resource Manager User Amazon EC2 CPU nodes are used to split the video into smaller ones GPU nodes 40

41 Cloud Seeding with ProActive CPU nodes Web Interface ProActive Scheduler + Resource Manager User Amazon EC2 GPU nodes are responsible to denoise these small videos GPU nodes 41

42 Cloud Seeding with ProActive CPU nodes Web Interface ProActive Scheduler + Resource Manager User Amazon EC2 GPU nodes are responsible to denoise these small videos GPU nodes 42

43 Cloud Seeding with ProActive CPU nodes Web Interface ProActive Scheduler + Resource Manager User Amazon EC2 CPU nodes merge the denoised video parts GPU nodes 43

44 Cloud Seeding with ProActive CPU nodes Web Interface ProActive Scheduler + Resource Manager User Amazon EC2 CPU nodes merge the denoised video parts GPU nodes 44

45 Cloud Seeding with ProActive CPU nodes Web Interface ProActive Scheduler + Resource Manager User Amazon EC2 The final denoised video is sent back to the user GPU nodes 45

46 4. ProActive PACA GRID: Cloud Portal with GPUs 46

47

48 The ProActive PACA Grid Platform (4) Total: 816 Cores 480 CUDA Cores 14.8TB Storage Publically Available Today 48

49 ProActive in Cloud Stack ProActive Cloud Portal IaaS 49

50 Conclusion 50 50

51 Conclusion Portability: Windows, Linux, Mac Versatility: Desktops, Grids, Clouds Java Parallel Multi-Platform Resource Toolkit Job Scheduler Manager Free Professional Open Source Software Manycores, Distribution and Cloud Multi-Core: Ready for a strong evolution Cloud: Smooth transition available (Desktop, Server, Cluster) 51

52 52 52

53 ProActive : Active objects JVM A ag = newactive ( A, [ ], VirtualNode) V v1 = ag.foo (param); V v2 = ag.bar (param);... v1.bar(); //Wait-By-Necessity JVM v2 v1 V ag A WBN! A Java Object Active Object Req. Queue Future Object Proxy Request Thread Wait-By-Necessity is a Dataflow Synchronization

54 Use Case 1: Genomics 54

55 Resources set up SOLID machine from 16 nodes Cluster Desktops Nodes can be dynamically added! EC2 Clouds 55 55

56 First Benchmarks The distributed version with ProActive of Mapreads has been tested on the INRIA cluster with two settings: the Reads file is split in either 30 or 10 slices Use Case: Matching 31 millions Sequences with the Human Genome (M=2, L=25) 4 Time FASTER from 20 to 100 Speed Up of 80 / Th. Sequential : 50 h 35 mn EC2 only test: t nearly the same performances as the local SOLiD cluster (+10%) For only $3,2/hour, EC2 has nearly the same perf. as the local SOLiD cluster (16 cores, for 2H30) On going Benchmarks on Windows Desktops and HPCS

57 Benchmark: local vs. hybrid cloud Use case: 3 runs performed in parallel containing a total of 28,5 millions of reads to be matched against the human genome SOLID nodes only Standard configuration using SOLID embedded nodes: 12 SOLID and EC2 nodes 12 SOLiD nodes 12 EC2 machines (type: mlarge, 2 nodes each) Total computation time: 12.5 hours Total computation time: 8 hours Gain: 4,5 hours (36% faster) EC2 costs: $40 57

58 Benchmark: local vs. EC2 cloud Execution time (min) Cost (US$) Standard PBS config 300 NA ProActive Amazon EC US$ For only $3,2/hour, the EC2 setup has nearly the same performances as the local SOLiD cluster 58

59 UC 2: Acceleration of Financial i Valuations 59

60 A High Performance Solution A Collaboration between Pricing Partners and ActiveEon Price-it Excel Accelerated by ProActive Parallel Suite A Global Solution: fully integrated with the same functionalities and interface as Price-it Excel while increasing its computing power High Quality Service: from both companies 60

61 Some Technical Facts Price-It C++ library developed d by Pricing i Partners Pricing solution dedicated to highly complex financial derivatives Specification and Constraints Accelerate Price-It Excel product Built on Price-It library, this product integrates an interface with Excel for input data management and results display Focus on highly parallelizable Greek computation Operating system: Windows 61

62 How Does it Work? Price-it Computing Distribution Price-it Excel Price-it Excel Regular Price-it Excel Interface ProActive Scheduler Automatic execution via job scheduler Pool of shared resources 62

63 Accelerated Price-it Performances Increased Productivity: Reduces Price-it Execution Time by 6 or more! More than 3 times faster with only 4 nodes! Even 6 times faster with 9 nodes! Use Case: Bermuda Vanilla, Model American MC Test conditions: One computation is split in 130 tasks that are distributed Each task uses 300ko Sequential 4 nodes 5 nodes 6 nodes 7 nodes 8 nodes 9 nodes Distributed 63

64 UC 3: SOA Analysis of Web Server Logs 64

65 Parallel Services Separation: BPEL Parallel Serv. Task Flow Standards et Portable High level Business Process Flexibility Domain specific Service Other Basic Service Job Scheduling Parameter Sweeping Parallel Services Scheduling `of Taskflow Jobs Divide & Conquer Other Operational Service Operational Services Scheduling Parameter Sweeping Service Resource Man. Enterprise Grid 65 65

66 Cas d études AMADEUS : Démonstration BPEL Scheduler RM Grille Log Parsing Log Parsing Log Parsing Log Parsing Log Parsing WS DB Log Transfer Application ID Log Transfer Application Log Parsing ID Log Parsing Logs stockés Agos Web Client Admin BPEL Console Admin WS Console Scheduler/ RM Interface MySQL Console BPEL/WS Web Client 66 66

67 Grid Monitoring integration Jobs Scheduler Job X Job Y Scheduler DB Business Availability Center JMX interface Resource Manager JDBC interface Discover Jobs, Tasks & Resources Discovery DDM Universal Create CMDB Configuration Management System information VM1 NATIVE 3 VM2 NATIVE 1 VM3 Pool of nodes in the grid VM4 VMx NATIVE 2 Collect Grid Jobs Type status indicators Collect Grid Components statistics indicators Indicators SiS 67

68 AGOS Platform Management HP- Business Availability Center (HP-BAC) Monitoring of the entire platform Cover all layers in the scope Provide monitoring dashboard and reports Tasks scheduler & Resources manager Integration with grid components Grid insights through indicator collection and running jobs on grid resources 68

69 AGOS: Grid Architecture for SOA Building a Platform for Agile SOA with Grid AGOS Solutions In Open Source with Professional Support 69 69

70 AGOS Infrastructure Management HP Systems Insight Manager (HP- SIM) Citrix XenCenter Monitoring of entire infrastructure Communicates with upper layer Hypervisor and VM management management software (HP BAC) Communicates with upper layer management software (HP BAC) 70

71 AGOS and HP Management tools Integration Services Processes and Jobs Discovery Grid Components Scheduler, Resource Manager Hypervisors/Virtual Machines Xen and Vmware Hosts and Guests Monitoring Hardware Infrastructure Servers, Storage, Network Components 71

72 4. Use Case: Wings Optimization 72

73 Wing performance (GridSystems, Spain) 73 73

74 Wing performance (GridSystems, Spain) 74 74

75 (2) ASP: Asynchronous Sequential Processes ASP Confluence and Determinacy Future updates can occur at any time Execution characterized by the order of request senders Determinacy of programs communicating over trees, A strong guide for implementation, Fault-Tolerance l and checkpointing, Model-Checking, 75 75

76 Key Point MPI and programming languages from the 60 s will not make it Jack Dongarra, 2/13/2009, Wake Forest University talk

77 NAS Parallel Benchmarks Experimented on 3D ElectroMagnetism, and Nasa Benchmarks Designed by NASA to evaluate benefits of high performance systems Strongly based on CFD 5 benchmarks (kernels) to test different aspects of a system 2 categories or focus variations: communication intensive and computation intensive 77 77

78 Communication Intensive CG Kernel (Conjugate Gradient) Floating point operations Eigen value computation High number of unstructured communications calls/node 570 MB sent/node 1 min % comms/wt Message density distribution Data density distribution 78 78

79 Communication Intensive CG Kernel (Conjugate Gradient) Comparable Performances 79 79

80 3. Parallel Scilab Simulations 80 80

81 Seamless Parallel & Distributed Scilab Static ti Policy Timing Policy Dynamic 12/24 Workload Policy LSF Desktops EC2 Dedicated resources Desktops Amazon EC2 81

82 Interface ProActive Scilab 82

83 Interface ProActive Matlab 83

84 Interface ProActive Scilab / Matlab Primitives de haut niveau Scilab/Matlab pour distribuer les calculs : f = PAeval(@foo,p1) // evalue f(p1) en asynchrone // f est un futur L = PAsolve(@foo, {p1,p2, pk}) // evalue foo(p1), foo(p2).. foo(pk) en asynchrone // L = f1, fk est une liste de futurs 84

85 Synchronization des calculs Primitives d attente : r = PAwaitForAny(L) // bloque et attend jusqu a l obtention du // premier des fk et retourne le resultat r lr = PAgetAny(L) // non-blocant, retourne une liste des resultats // deja obtenus 85

86 4. Use Case: OMD

87 OMD2 Interfaces Open Sources pour faciliter les Optimisations i Multi-Disciplinaires i li i Distribuées ib 87

88 OMD2 : Interfaces Open Sources pour faciliter les Optimisations Multi-Disciplinaires Distribuées 88

89 Plateforme OMD2, vue générale : Input (Xi,obj fct, constr) User / Renault DIGITEO / SCILA AB f(x1) f(xn) f(x2) ENSM-SE, SE ECP TAO, UTC, ENS, INRIA Optimizer (surrogate based, CMA ES...) test case 1 locall test cases 2,3,4 webservice Middleware Proactive end Sirehna + CD-adapco Output (Analysis) Cluster 1 Cluster 2... Cluster N adapco UTC INRIA 89

90 Distribution des calculs Acquisition des maillages, etc WS Catia Stratégie de résolution Boucle optimisation Soumission du PO N1 N2 Ressources Ni1 Scheduler PO Visualisation stats, etc Nik Calculs f(x) BD Mailleur Solveur 90

91 Les cas tests Conduit d'admission 350 P2 P min CPU P4 P6 P3 P5 P7 P8 P10 P9 <1min CPU Conduit de climatisation 3D P12 P Conduit de climatisation 2D 100h CPU 1000h CPU Culasse multi-disciplinaire Aéro externe 91

92 Interface ProActive Scilab 92

93 Exemple d utilisation : Boucle d Optimisation futurelist = {}; PAconnect( //machine where the scheduler is running ) do { if size(futurelist) < MAX_SENT { x=opt.ask(); // asks a new vector future = PAeval (@f, {x}); // asynchronously send f(x) to grid futurelist = append(futurelist, future); } else { resultlist = PAgetAny(futureList); // Get Available results If resultlist!= void { // Updates the optimizer with new results for i=1:size(resultlist) { [y,x] = resultlist(i); opt.tell(y,x); } } else Sleep (t) } } while ~(opt.stop()) 93