Документ взят из кэша поисковой машины. Адрес оригинального документа : http://www.parallel.ru/sites/default/files/ftp/computers/scali/tokyo_may00.pdf Дата изменения: Wed Nov 2 11:53:59 2011 Дата индексирования: Tue Oct 2 03:46:48 2012 Кодировка:

Scalable Linux Systems
Einar Rustad Scali AS einar@scali.com http://www.scali.com Junzo Tamada Northern Lights Computers KЕre LЬchsen Dolphin Interconnect Solutions


Scalable System Requirements · Balanced Hardware Resources
­ ­ ­ ­ High Processor Speed Scalable Memory and Storage High Bandwidth Interconnect Low latency Communication

· Efficient Middleware, Standard APIs · Ease of Use · Easy and Flexible System Administration
Slide 2 - 10/12/00

Scalable Linux Systems


The Value Chain
Interconnect Hardware Interconnect Software Complete(MPI, SCI-SAN) Integrated Systems System Management- Support Tools

Slide 3 - 10/12/00

Scalable Linux Systems


Dolphin Technology
· Based on SCI, the only bus extension technology · Proven over years in applications like
­ ­ ­ ­ ­ Clustering of SUN's high availability servers Fujitsu Siemens large scale IO system Data General AViiON ccNUMA servers Mirage and Rafale flight computers Scali and Nothern Lights ISP, ASP and HPC Servers

· Solutions for Serversand Embedded Computing · 2µs ° Application Latency, 500 Mbytes/s Link Speed · Dolphin sells Chips, Cards, Switches and licenses technology · Dolphin makes the Scali and WulfkKit card assembly and sells the WulfKit to system buidlers
Slide 4 - 10/12/00

Scalable Linux Systems


Scalable Linux Systems Advantages
· Industry Standard Programming Model - MPI
­ Porting = Recompilation

· Scalability
­ Scalable to thousands of Processors

· Lower Cost
­ COTS based Hardware = lower system price ­ Lower Total Cost of Ownership

· Redundancy · Single System Image to users and administrator · Choice of Front-End OS
­ Linux ­ Solaris ­ Windows NT

· Better Performance
­ Always "Latest & Greatest" Processors ­ Superior Standard Interconnect - SCI
Slide 5 - 10/12/00

Scalable Linux Systems


Torus Topology - Distributed Switching
Distributed Switching:
PCI-bus PSB B-Link LC-2 LC-2

Horizontal SCI Ring

Vertical SCI Ring

Slide 6 - 10/12/00

Scalable Linux Systems


Theoretical Scalability
100

10 GByte/s
Ringlet 2D-Torus 3D-Torus PCI

1

0,1 1 10 Number of Nodes
Slide 7 - 10/12/00

100

1000

Scalable Linux Systems


Versus Myrinet (1)
Barrier synchronization
200 180 160 140 120 100 80 60 40 20 0 2
Slide 8 - 10/12/00

MPICH/Myrinet GM barrier Scali MPI/SCI barrier

4

8 Number of nodes

9

16

Scalable Linux Systems


Versus Myrinet (2)
All-to-all performance
90 80 70 60 50 40 30 MPICH/Myrinet GM all-to-all 20 10 0 2 4 8 Number of nodes
Slide 9 - 10/12/00

Scali MPI/SCI all-to-all

9

16

Scalable Linux Systems


Versus Origin 2000 (1)
All-to-all Bandwidth per Node
120
Origin2k

100 80 60 40 20 0 2
Slide 10 - 10/12/00

ScaMPI/SCI

4

8

9

16

25

32

36

49

64

Number of Nodes

Scalable Linux Systems


Versus Origin 2000 (2)
Barrier Synchronization
500 450 400 350 300 250 200 150 100 50 0 2
Slide 11 - 10/12/00

Origin2k ScaMPI/SCI

4

8

9

16

25

32

36

49

64

Number of Nodes

Scalable Linux Systems


Fault Tolerance
· Automatic Rerouting · Scali advanced routing algorithm:
­ From the Turn Model family of routing algorithms

43 42 41 44

13 12 11 14

23 22 21 24

33 32 31 34

· All nodes but the failing ones can be utilised as one big partition
Slide 12 - 10/12/00

Scalable Linux Systems


The Scali Universe

Slide 13 - 10/12/00

Scalable Linux Systems


Software Configuration Management

Slide 14 - 10/12/00

Scalable Linux Systems


System Monitoring

Slide 15 - 10/12/00

Scalable Linux Systems


Platform Attraction

PGI

Totalview

DQS

ScaMPI

GUI System Monitoring
Vampir

ICM

Config. Mngmnt
TimeScan

Slide 16 - 10/12/00

Scalable Linux Systems