Kein Folientitel

1
Gesellschaft für Parallele Anwendungen und
Systeme mbH
WP1 Contribution
Karl Solchenbach, Pallas GmbH Apart Meeting,
June 25, 1999
Pallas GmbHHermülheimer Straße 10D-50321 Brühl,
Germanyinfo_at_pallas.dehttp//www.pallas.com
2
Performance Evaluation

• Intelligent use of Vampir and Dimemas
• To qualify the performance of a parallel code
• (Speed-up of 9.3 on 16 processors, good or bad?)
• To understand reasons of poor performance
• (Communication, latency, IO, imbalance, ...)
• To develop strategies for code tuning

3
Total time Useful time Overhead
Overhead
times efficiencies ?
4
Other efficiencies

• Single node efficiency
• sustained vs. peak rate (Enode r / r? )
• often lt 10
• Total efficiency
• sustained vs. parallel peak rate (Enode ?
  Epar)
• Algorithmical efficiency
• best sequential algorithm vs. parallel algorithm

5
Example Sequential chain

• Etrans 1
• Elb 1
• Edep 1/P
• Econc 1/P

6
Example One process busy

• Etrans 1
• Elb 1/P
• Edep 1
• Econc 1/P

7
Example Ring

• Elat 85
• Ebw 59
• Elb 1
• Edep 1
• Econc 50

8
Example NAS LU/ (Class W) on T3E (8 PEs)

• Elat 91.9
• Ebw 99.6
• Etrans 91.5
• Elb 91.1
• Edep 93.7
• Eblock 85.4
• Econc 78.1

9
Example NAS LU (detail)
dependency
10
Comparison on T3E and SX-4
Cray T3E NEC SX-4 Elat 91.9 96.7 Ebw
99.6 100.0 Etrans 91.5 96.7 Elb 91.1
95.0 Edep 93.7 90.0 Eblock 85.4
85.5 Econc 78.1 82.7
11
FE-Code Scalability considerations
12
Increasing floating-point speed
13
Conclusion

• Vampir and Dimemas can quantify and separate the
  main sources for efficiency loss(without source
  analysis or code know-how)
• Strategy for code tuning can be developed
• replace parallel algorithm ?
• improve implementation of existing algorithm ?
• use/buy other HPC platform ?
• Powerful tools and their intelligent use improve
  productivity in performance tuning