Petaflops Applicatios Group Meeting

1
Petaflops Applicatios Group Meeting

• January 16, 2004

2
Agenda

• Status report
• Quick summary of goals
• Mark on pNeo

3
Status report

• New web page (in progress)
• http//www-unix.mcs.anl.gov/siegela/peta-apps.htm
  
• Mail me any info you would like to add to web
  page
• PAPI temporarily not working on recorder after
  security breach. Rick Bradshaw is working on
  this.
• Status of PAPI/Tau on Jazz?
• Rick looked into installing VTune. Said the
  system requirements may make it impossible
• icc mpi builds on jazz a big mess. Ill send
  email about this.

4
PetaApps Goals

• Understand requirements of prominent HPC
  applications in terms of machine
  hardware/software
• to guide application developers in optimization
• Performance modeling to guide next generation
  (peta-scale) architecture/tools development
• CPU architecture
• how well do application algorithm needs map onto
  current proc hardware
• appropriate metrics -- what fraction of peak can
  be obtained? What are bottlenecks (e.g. cache
  size, branch mispredictions with long pipelines,
  wrong instruction mix).
• How well can compilers automate these
  optimizations?

5
Goals, cont.

• Parallel issues
• what do algorithms need out parallel environment
  to maintain efficiency MPI bandwidth and
  latency?
• How will these algorithms scale to 100K procs,
  etc?
• Can a methodology or science be established for
  carrying out this process?
• How adequate are tools used to analyze
  performance?

6
Goals, cont.

• Methodology
• Form collaboration with application developers.
• Improve software process to facilitate
  collaboration.
• Create repository of applications and carry out
  analyses. Characterize applications according to
  principle bottlenecks and hardware/software
  needs.
• Model performance on new systems.
• How does this relate to PERC peformance summary,
  Jeffrey Vetter, etc.

7
Offshoots

• Tools
• HPC regression testing tool
• Automated cataloging and searching of performance
  data