NPACI Panel on Clusters David E. Culler Computer Science Division University of California, Berkeley...

19
NPACI Panel on Clusters David E. Culler Computer Science Division University of California, Berkeley http://www.cs.berkeley.edu/~culler
  • date post

    19-Dec-2015
  • Category

    Documents

  • view

    219
  • download

    0

Transcript of NPACI Panel on Clusters David E. Culler Computer Science Division University of California, Berkeley...

Page 1: NPACI Panel on Clusters David E. Culler Computer Science Division University of California, Berkeley culler.

NPACI Panel on Clusters

David E. Culler

Computer Science Division

University of California, Berkeleyhttp://www.cs.berkeley.edu/~culler

Page 2: NPACI Panel on Clusters David E. Culler Computer Science Division University of California, Berkeley culler.

1/29/99 NAPCI Clusters 2

Clusters have Happened

• IBM

• SGI

• SUN

• HP

• Microsoft

• ...

Page 3: NPACI Panel on Clusters David E. Culler Computer Science Division University of California, Berkeley culler.

1/29/99 NAPCI Clusters 3

Performance + Cost-Performance

• April 1998 T3E vs Intel Pentium II on NAS Benchmarkshttp://www.nersc.gov/research/FTG/pcp

T3E-900 PII-400

BT 63.5 62.2

LU 70.2 73.6

SP 44.7 45.5

LIN

PA

CK

(M

FL

OP

S)

1

10

100

1,000

10,000

1975 1980 1985 1990 1995 2000

CRAY n = 100 CRAY n = 1,000

Micro n = 100 Micro n = 1,000

CRAY 1s

Xmp/14se

Xmp/416Ymp

C90

T94

DEC 8200

IBM Power2/990MIPS R4400

HP9000/735DEC Alpha

DEC Alpha AXPHP 9000/750

IBM RS6000/540

MIPS M/2000

MIPS M/120

Sun 4/260

Page 4: NPACI Panel on Clusters David E. Culler Computer Science Division University of California, Berkeley culler.

1/29/99 NAPCI Clusters 4

Berkeley (NPACI) NOW

• 100 Sun Ultra2 workstations

• Inteligent network interface

– proc + mem

• Myrinet Network

– 160 MB/s per link

– 300 ns per hop

http://now.cs.berkeley.edu

Page 5: NPACI Panel on Clusters David E. Culler Computer Science Division University of California, Berkeley culler.

1/29/99 NAPCI Clusters 5

Beowolf Consortium

• LINUX PCs

• Fast ethernet

• Basic stand-alone and batch

• http://www.beowulf.org/consortium.html

• http://www.beowulf.org/

• http://ns.baltzer.nl/cluster/cluster.html

• http://www.cs.orst.edu/swarm/about.html

• cluster cookbook

Page 6: NPACI Panel on Clusters David E. Culler Computer Science Division University of California, Berkeley culler.

1/29/99 NAPCI Clusters 6

HPVM (NPACE)

• NT PCs

• Myricom Network

• Fast Messages

• LSF start-up

http://www-csag.cs.uiuc.edu/projects/hpvm.html

Page 7: NPACI Panel on Clusters David E. Culler Computer Science Division University of California, Berkeley culler.

1/29/99 NAPCI Clusters 7

Berkeley Millennium

• PC-based Unix and NT Clusters

• Departmental and Campus

• Shared as Computational Economy

Gigabit Ethernet

SIMS

C.S.

E.E.

M.E.

BMRC

N.E.

IEORC. E. MSME

NERSC

Transport

Business

Chemistry

Astro

Physics

Biology

EconomyMath

http://www.millennium.berkeley.edu

Page 8: NPACI Panel on Clusters David E. Culler Computer Science Division University of California, Berkeley culler.

1/29/99 NAPCI Clusters 8

What you get off the shelf

• Go to your favorite web site and order– Dual PII 450 MHz, 1 GB Mem, 36 GB disk

– $9,563 at Dell

• 4400 CPU Hours per Quarter to your self!– $1.20 CPU hours (at 30% over 3 years)

• Buy 5?

Page 9: NPACI Panel on Clusters David E. Culler Computer Science Division University of California, Berkeley culler.

1/29/99 NAPCI Clusters 9

Three Kinds of Clusters

• Throughput Clusters

• Availability Clusters

• High-Performance Parallel Clusters

Page 10: NPACI Panel on Clusters David E. Culler Computer Science Division University of California, Berkeley culler.

1/29/99 NAPCI Clusters 10

Throughput Clusters

• Workstation / PC Farms

• Provide resource pool for large numbers of sequential jobs

• Used widely in industry– toy story on 2000 sparcstations

– ultrasparc on 1000 sparcstations

• Can also include background on desktops– CONDOR

• Application specific front-ends attractive– parametric studies, monte carlo

• Fill cracks in parallel clusters

• Big farms require infrastructure

Page 11: NPACI Panel on Clusters David E. Culler Computer Science Division University of California, Berkeley culler.

1/29/99 NAPCI Clusters 11

Availability Clusters

• Use system redundancy to mask faults– all big databases do it

• VAX Clusters => IBM sysplex => Wolf Pack

ClientsClients

Disk array ADisk array A

Disk array BDisk array B

InterconnectInterconnectServerServer AA

ServerServer BB

Page 12: NPACI Panel on Clusters David E. Culler Computer Science Division University of California, Berkeley culler.

1/29/99 NAPCI Clusters 12

High-Performance Clusters

• Utilize modern system area networks and user-level communication layers to construct general purpose parallel machine from commodity parts

Page 13: NPACI Panel on Clusters David E. Culler Computer Science Division University of California, Berkeley culler.

1/29/99 NAPCI Clusters 13

Emerging System Area Networks

• Gigabit Ethernet– price dropping, widely deployed

• System Area Networks– Myricom

– ServerNet

– Synfinity

• Virtual Interface Architecture– Intel/Microsoft/Compaq std based on univ.

research prototypes

Page 14: NPACI Panel on Clusters David E. Culler Computer Science Division University of California, Berkeley culler.

1/29/99 NAPCI Clusters 14

MPI Performance

0

10

20

30

40

50

60

70

10 100 1000 10000 100000 1000000

Message Size (bytes)

Ban

dwid

th (

MB

/s)

SGI Challenge

Meiko CS2

NOW

IBM SP2

Cray T3D

0

10

20

30

40

50

60

70

80

90

SGI

Chal

leng

e

Meiko NO

W

IBM

SP2

Cray

T3D

mic

rose

cond

s

Page 15: NPACI Panel on Clusters David E. Culler Computer Science Division University of California, Berkeley culler.

1/29/99 NAPCI Clusters 15

0

25

50

75

100

125

0 25 50 75 100 125

Processors

Sp

ee

du

p o

n L

U-A

T3D

SP2

NOW

Example: NAS Parallel Benchmarks

• Better node performance than the Cray T3D

• Better scalability than the IBM SP-2

Page 16: NPACI Panel on Clusters David E. Culler Computer Science Division University of California, Berkeley culler.

1/29/99 NAPCI Clusters 16

Cluster-Wide Parallel I/O

• Sustain 500 MB/s disk bandwidth and 1,000 MB/s network bandwidth by driving all the disks

Minute Sort

SGI Power Challenge

SGI Orgin

0123456789

0 10 20 30 40 50 60 70 80 90 100

Processors

Gig

abyt

es s

orted

Page 17: NPACI Panel on Clusters David E. Culler Computer Science Division University of California, Berkeley culler.

1/29/99 NAPCI Clusters 17

Software Base is Growing

• Technical software moving to Linux and NT– NAG, matlab, petc, ...

• Cluster prototypes being hardened and packaged

• Cookbooks emerging

• Few cluster-integrator companies

Page 18: NPACI Panel on Clusters David E. Culler Computer Science Division University of California, Berkeley culler.

1/29/99 NAPCI Clusters 18

Summary of the State of the Art

• Cluster designs are emerging in many areas– throughput, availability, parallel computing

– technology is advancing

• Still immature software base– strong ties to free software movement

• Many small clusters by spit and baling wire

• Large clusters require engineering– commercial components improving

• Rapid pace of change presents sys. Admin challenge

– not unlike the desktop problem

• Management tools badly needed

Page 19: NPACI Panel on Clusters David E. Culler Computer Science Division University of California, Berkeley culler.

1/29/99 NAPCI Clusters 19

What Does it Mean for NPACI

• Where do clusters fit with computational science

and engineering needs?

• Cycles vs Software vs Administration vs Expertise?

• What role should the center take?

• What role should partner sites have?