Petaflops Application Meeting

1
Petaflops Application Meeting

• Dec. 12 2004

2
Agenda

• Updates
• Repositories
• Tools
• Projects
• Katherine on Steve Piepers Code
• Status of other codes

3
Updates

• nek5 repository on cvs server
• Will add permissions for anybody interested
• Other codes should go there also if possible
• Tisha can do this quickly
• Started to put together simple generic regression
  testing tools to go with above
• mysql database setup on database server
• Will be used collect performance information in
  standardized way
• Need to nail down specs on what standard contains
• Will email connection information
• Who would like an account?
• Next for nek5
• Most interesting stuff serial
• Understand why in particular flop count low for
  various arhitectures

4
Tools

• Performs (Performance on Real Memory Systems)
• Read/write latencies for memory subsystem
• Not installed locally
• Valgrind
• Linux x86 debugger/profiler
• On Jazz?
• PAPI (Performance API) University of TN
• Portable hardware counters
• Requires kernel module, not installed on Jazz
• On recorder.mcs.anl.gov
• TAU (Tuning and Analysis Utilities) University of
  Oregon
• Very portable, broad set of tools for full
  spectrum of profiling shared and distributed
  memory, multi-language support, layered on
  Dyninst, Opari, Epilogue, Vampir, Paraver, etc.
  Multiple outputs, views.
• On Jazz?

5
Tools

• Vampir (Intel)
• Paradyn / Dyninst (University of Wisconsin)
• Pablo Toolkit (University of Illinois)
• Paraver/Dimemas (Barcelona)
• Jumpshot (Argonne)
• Complex views of MPI messaging patterns for large
  application
• Installed on Jazz and many other places
• FPMPI (Argonne)
• Simple lightweight text-based MPI messaging
  diagnostics
• Installed on Jazz and many other places

6
Tools

• HPM
• IBMs
• Speedshop
• SGIs
• Gprof
• Linux-x86 call graphs/timings

7
Related projects

• PERC (Performance Evaluation Research Center)
• Scidac ISIC for Science of performance
• Paul and Boyana co-PIs here
• IPACS
• Development of realistic scientific benchmarks