Embed Size (px)
Citation preview
From Clusters to GridsMaster 2 Research Lecture: Parallel Systems
Vincent Danjean, MCF UJF, LIG/INRIA/MoaisDerick Kondo, CR INRIA, LIG/INRIA/Mescal
Arnaud Legrand, CR CNRS, LIG/INRIA/MescalJean-Francois Mehaut, PR UJF, LIG/INRIA/Mescal
Bruno Raffin, CR INRIA, LIG/INRIA/MoaisJean-Louis Roch, MCF ENSIMAG, LIG/INRIA/Moais
Alexandre Termier, MCF UJF, LIG/Hadas
LIG laboratory, [email protected]
October 13, 2008
A. Legrand (CNRS-LIG) INRIA-MESCAL From Clusters to Grids 1 / 23
Outline
1 Clusters
2 What next?
A. Legrand (CNRS-LIG) INRIA-MESCAL From Clusters to Grids 2 / 23
Outline
1 Clusters
2 What next?
A. Legrand (CNRS-LIG) INRIA-MESCAL From Clusters to Grids Clusters 3 / 23
Motivation
Parallel Machines
I Parallel machines are expensive.I The development tools for workstations are more mature
than the contrasting proprietary solutions for parallel com-puters - mainly due to the non-standard nature of manyparallel systems.
Workstation evolution
I Surveys show utilization of CPU cycles of desktop worksta-tions is typically < 10%.
I Performance of workstations and PCs is rapidly improvingI The communications bandwidth between workstations is in-
creasing as new networking technologies and protocols areimplemented in LANs and WANs.
I As performance grows, percent utilization will decrease evenfurther! Organizations are reluctant to buy large supercom-puters, due to the large expense and short useful life span.
A. Legrand (CNRS-LIG) INRIA-MESCAL From Clusters to Grids Clusters 4 / 23
Towards clusters of workstations
I Workstation clusters are easier to integrate into existing net-works than special parallel computers.
I Workstation clusters are a cheap and readily available alterna-tive to specialized High Performance Computing (HPC) plat-forms.
I Use of clusters of workstations as a distributed compute re-source is very cost effective - incremental growth of system!!!
Definition.
A cluster is a type of parallel or distributed processing system(MIMD), which consists of a collection of interconnected stand-alone/complete computers cooperatively working together as a sin-gle, integrated computing resource.
A. Legrand (CNRS-LIG) INRIA-MESCAL From Clusters to Grids Clusters 5 / 23
Definition
A typical clusterI A cluster is mainly homogeneous and is made of high per-
formance and generally rather low cost components (PCs,Workstations, SMPs).
I Composed of a few to hundreds of machines.I Network: Faster, closer connection than a typical LAN net-
work; often a high speed low latency network (e.g. Myrinet,InfiniBand, Quadrix, etc.); low latency communication pro-tocols; looser connection than SMP.
Typical usageI Dedicated computation (rack, no screen and mouse).I Non dedicated computation: Classical usage during the day
(word, latex, mail, gcc) / HPC applications usage during thenight and week-end.
Biggest clusters can be split in several parts:I computing nodes;
I I/O nodes;
I front (interactive) node.
A. Legrand (CNRS-LIG) INRIA-MESCAL From Clusters to Grids Clusters 6 / 23
A few examples
Berkeley NOW (1997)
I 100 SUN UltraSPARCs.
I Myrinet 160MB/s.
I Fast Ethernet.
A. Legrand (CNRS-LIG) INRIA-MESCAL From Clusters to Grids Clusters 7 / 23
A few examples
Icluster (2000)
I 225 HP iVectra PIII 733Mhz.
I Fast Ethernet.
I 81.6 Gflops (216 nodes).
I top 500 (385) June 2001.
A. Legrand (CNRS-LIG) INRIA-MESCAL From Clusters to Grids Clusters 7 / 23
A few examples
Digitalis (2008)
I 34 nodes (2 xeon quad cores); 272 cores with 2× 8Gb ofRAM and 2 × 160Gb of HDeach.
I Infiniband.
I Giga Ethernet.
A. Legrand (CNRS-LIG) INRIA-MESCAL From Clusters to Grids Clusters 7 / 23
Outline
1 Clusters
2 What next?
A. Legrand (CNRS-LIG) INRIA-MESCAL From Clusters to Grids What next? 8 / 23
Clusters of clusters (HyperClusters)
DAS3: ASCI (Advanced School for Computing andImaging),Netherlands.
I Five Linux supercomputerclusters with 550 AMDOpteron processors.
I 1TB of memory and 100TBof storage.
I Myricom Myri-10G networkinside clusters.
I Clusters are interconnectedby a SURFnet’s multi-coloroptical backbone.
A. Legrand (CNRS-LIG) INRIA-MESCAL From Clusters to Grids What next? 9 / 23
The concept of Grid. . .
The Grid: Blueprint for a New Com-puting Infrastructure (1998); Ian Fos-ter, Carl Kesselman, Jack Dongarra, FranBerman, . . . .Analogy with the electric supply:
I You don’t know where the energycomes from when you turn on yourcoffee machine.
I You don’t need to know where yourcomputations are done.
A. Legrand (CNRS-LIG) INRIA-MESCAL From Clusters to Grids What next? 10 / 23
The concept of Grid (cont’d)
A grid is an infrastructure that couples:
I Computers (PCs, workstations, clusters, traditional supercom-puters, and even laptops, notebooks, mobile computers, PDA,and so on);
I Software Databases (e.g., transparent access to human genomedatabase);
I Special Instruments (e.g., radio telescope–SETI@Home Search-ing for Life in galaxy, Astrophysics@Swinburne for pulsars, acave);
I People (maybe even animals who knows ?;-)
across the local/wide-area networks (enterprise, organizations, or In-ternet) and presents them as an unified integrated (single) resource.
A. Legrand (CNRS-LIG) INRIA-MESCAL From Clusters to Grids What next? 11 / 23
What does a Grid look like?
It is very big and very heterogeneous!
A. Legrand (CNRS-LIG) INRIA-MESCAL From Clusters to Grids What next? 12 / 23
Various versions of “Grid ”
You have probably heard of many buzzwords.
I Super-computing;
I Global Computing;
I Internet Computing;
I Grid Computing;
I Meta-computing;
I Web Services;
I Cloud Computing;
I Ambient computing;
I Peer-to-peer;
I Web;
Large Scale Distributed Systems
“A distributed system is a collection of independentcomputers that appear to the users of the system as asingle computer”Distributed Operating System. A. Tannenbaum, Pren-tice Hall, 1994
A. Legrand (CNRS-LIG) INRIA-MESCAL From Clusters to Grids What next? 13 / 23
Tentative taxonomy
PurposeI Information: share knowledge.I Data: large-scale data storage.I Computation: aggregate computing power.
Deployment modelI Not necessarily fully centralized.
I Use of caches and proxys to reduce con-gestion.
I Hierarchical structure is often used.
I Centralized information
Client/server
Congestion Area
Server Server
Client
Client ProxyCache
Client
ClientProxyCache
ProxyCache
Client
Client
Client
I Each peer acts both as a client and aserver.
I The load is distributed over the wholenetwork.
I Distributed information.
Peer-to-peer
ServerClient
ServerClient
ServerClient
ServerClient
ServerClient
ServerClient
ServerClient
ServerClientServer
Client
ServerClient
ServerClient
ServerClient
A. Legrand (CNRS-LIG) INRIA-MESCAL From Clusters to Grids What next? 14 / 23
Example: Web sitesClient/server; information grid
Context
I Probably the first “grid”.I Information is accessed through a URL or more often through
a search engine.I Information access is fully transparent: you generally don’t
know where the informations comes from (mirrors, RSS feeds,...).
Challenges Going peer-to-peer ? Web 2.0: users also contribute.
I Social networks (Facebook).I Recommendations (google and amazon.com).I Crowdsourcing (wikipedia, marmiton).I Video and photo sharing (youtube).I Media improvement (e.g., linking picassa and google maps).I Ease of finding relevant information and ability to tag data.
A. Legrand (CNRS-LIG) INRIA-MESCAL From Clusters to Grids What next? 15 / 23
Example: NapsterClient/server; data grid
Context
I The first massively popular “peer-to-peer” file (MP3 only) sharing system(1999).
I Central servers maintain indexes of con-nected peers and the files they provide.
I Actual transactions are conducted di-rectly between peers.
Drawbacks
I More client/server than truly peer-to-peer.
I Hence, servers have been attacked (bycourts and by others to track peers of-fering copyrighted materials).
A. Legrand (CNRS-LIG) INRIA-MESCAL From Clusters to Grids What next? 16 / 23
Example: Gnutella, Kazaa, Freenet, ChordP2P; data grid
ContextI Removal of servers: searching can be done by flooding in
unstructured overlays.I Use of supernodes/ultrapeers (nodes with a good CPU and
high bandwidth) for searching.I Structured (hypercubes, torus, . . . ) overlay networks.I Downloading from multiple sources using hash blocks and
redundancy.
33
96
10.0
34
10.0
32
95
10.0
31
10.0
90
16
10.0
70
10.0
63
1
10.0
82
10.0
21
65
10.0
6710.0
71
26
10.0
56
10.0
7
64
10.0
23
10.0
80
76
10.0
30
10.0
61
10.0
94
10.0
99
10.0
10.0
18
10.0
75
10.0
72
41
10.0
88
10.0
58
10.0
59 10.0
44
10.0
10.0
55
10.0
10.0
84
57
10.0
10.0
74
10.0
10.0
27
68
10.0
10.0
10.0
10.0
2
10.0
10.0
60
10.0
62
10.0
20
77
10.0
13
10.0
92
87
10.0
10.0
89
39
10.0
10.0
10
73
10.0
4
10.0
10.0
10.0
35
10.0
10.0
11
10.0
83
10.0
91
10.0
14
10.0
7810.0
10.0
48
93
10.0
10.0
10.0
10.0
10.0
85
10.0
10.0
10.0
10.0
10.0
29
10.0
10.0
50
10.0
24
10.0
10.0
81
10.0
10.0
10.0
97
10.0
28
10.0
10.0
10.0
10.0
46
10.0
12
10.0
10.0
15
10.0
3
10.0
10.0
10.0
52
10.0
10.0
10.0
10.0
10.0
36
10.0
10.0
10.0
66
10.0
10.0
45
10.0
54
10.0
86
10.0
10.0
10.0
10.0
19
79
10.0
10.0
10.0
10.0
10.0
10.0
10.0
10.0
10.0
10.0
10.0
10.0
10.0
10.0
17
10.0
10.0
10.0
10.0
10.0
10.0
10.0
10.0
10.0
10.0
25
10.0
10.0
10.0
10.0
10.0
10.0
10.0
6
10.0
40
10.0
10.0
10.0
10.0
69
0
10.0
10.0
10.0
10.0
49
10.0
51
10.0
10.0
10.0
53
10.0
10.0
10.0
10.0
22
10.0
10.0
42
10.0
10.0
10.0
10.0
10.0
10.0
10.0
10.0
10.0
10.0
10.0
10.0
10.0
10.0
10.0
10.0
10.0
10.0
10.0
10.0
10.0
10.0
9
10.0
10.0
10.0
98
10.0
47
10.0
10.0
8
10.0
10.0
38
10.0
10.0
10.0
10.0
10.0
10.0
10.0
10.0
37
10.0
10.0
43
10.0
10.0
5
10.0
10.0
A. Legrand (CNRS-LIG) INRIA-MESCAL From Clusters to Grids What next? 17 / 23
Example: Gnutella, Kazaa, Freenet, ChordP2P; data grid
ContextI Removal of servers: searching can be done by flooding in
unstructured overlays.I Use of supernodes/ultrapeers (nodes with a good CPU and
high bandwidth) for searching.I Structured (hypercubes, torus, . . . ) overlay networks.I Downloading from multiple sources using hash blocks and
redundancy.
Challenges
I Ensuring anonymity.I Ensuring good throughput and efficient multi-cast (network
coding, redundancy).I Avoiding polluted data.I Publish-subscribe overlays for fuzzy or complex queries.I Free-riders.
A. Legrand (CNRS-LIG) INRIA-MESCAL From Clusters to Grids What next? 17 / 23
Example: Internet Computing (SETI@home)Client/server; computation grid
Context
I Search for possible evidence of radio transmissions from ex-traterrestrial intelligence using data from a telescope.
I The client is generally embedded into a screensaver.I The server distributes the work-units to volunteer clients.I Attracting volunteers with hall of fame and teams.I Need to cross-check the results to detect false positives.I 5.2 million participants worldwide, over two million years of
aggregate computing time since its launch in 1999. 528TeraFLOPS (Blue Gene peaks at just over 596 TFLOPSwith sustained rate of 478 TFLOPS).
I Evolved into BOINC: Berkeley Open Infrastructure for Net-work Computing (climate prediction, protein folding, primenumber factorizing, fight cancer, Africa@home, . . . ).
A. Legrand (CNRS-LIG) INRIA-MESCAL From Clusters to Grids What next? 18 / 23
Example: Internet Computing (SETI@home)Client/server; computation grid
Challenges
I Attract more volunteers: credits, ribbons and medals, con-nect with facebook.
I Volunteer thinking: use people’s brains (intelligence, knowl-edge, cognition) to locate’ solar dust, fossils, fold proteins.
I Works well for computation intensive embarrassingly parallelapplications.
I Really parallel applications.I Data intensive applications.I Soft real-time applications.
I Security.I Would you let anyone execute anything on your PC?I Use sandboxing and virtual machines.
I Need to go peer-to-peer (CGP2P, OurGrid).
A. Legrand (CNRS-LIG) INRIA-MESCAL From Clusters to Grids What next? 19 / 23
Example: Meta-computingClient/server; computation grid
Context
I Principle: buy computing ser-vices (pre-installed applications+ computers) on the Internet.
I Examples: Netsolve (UTK),NINF (Tsukuba), DIET andScilab // (ENS Lyon/INRIA),
Challenges
I Data storage and distribution: avoid multiple transfers be-tween clients and servers when executing a sequence of op-erations.
I Efficient data redistribution.I Security for file transfersI Peer-to-peer deployment.
A. Legrand (CNRS-LIG) INRIA-MESCAL From Clusters to Grids What next? 20 / 23
Example: grid computingClient/server; computation grid
ContextI Principle: use a virtual supercomputer and execute applications on
remote resources.“I need 200 64 bits machines with 1Tb of storage from 10:20 amto 10:40 pm.”
I Need to match and locate resources, schedule applications, handlereservations, authentication, . . .
I Examples: Globus, Legion, Unicore, Condor, . . .
ChallengesI Obtaining good performances while deploying parallel codes on
multiple domains.I Communication and computation overlap. High-performance com-
munications on heterogeneous networks.I Need for new parallel algorithms that handle heterogeneity, hierar-
chy, dynamic resources,I Complex applications ; code coupling (message passing ; dis-
tributed objects, components).A. Legrand (CNRS-LIG) INRIA-MESCAL From Clusters to Grids What next? 21 / 23
Summary
XXXXXXXXXXXUsageDeployment
Client/Server Peer-to-peer
Data Napster Gnutella, Kazaa,Chord, Freenet. . .
Information Web 1.0 and 1.5Search Engines
Web 2.0
Computing Internet Computing;Meta-computing;Grid Computing
OurGrid
A few other challengesI Security, Authentication, Trust, Error management.I Middleware vs. Operating System.I Algorithms for Grid Computing.I Software engineering.I Social aspects (fairness, selfishness, cooperation).I Energy saving!
A. Legrand (CNRS-LIG) INRIA-MESCAL From Clusters to Grids What next? 22 / 23
Conclusion
I No real new theme but rather a combination of already existingtechnologies for parallel and distributed computing.
I Such combinations and ambitious goals are very hard to achieve.
I This clearly requires a pluri-disciplinary approach with a goodunderstanding of all aspects (OS, network, middleware, security,storage, algorithms, applications, . . . ).
I It would be a mistake to restrict only to computing. Researchon all these aspects should be encouraged.
I It is very important to identify and discriminate new conceptsfrom technology and fad.
I A crucial question is:
“Should we hide the complexity or expose it?”
A. Legrand (CNRS-LIG) INRIA-MESCAL From Clusters to Grids What next? 23 / 23
