Updated and alternative profiling of parflow.

[ian-bertolacci]

Its been a while since we did any good performance profiling, and we'd like to know performance profile of existing master before doing performance comparison of the new shared memory parallel versions.

Additionally, I watched this stunning talk by Emery Berger on performance profiling (Performance Matters) where he talks about a few of his profiling tools that take the statistical approach we need to be using in our performance analysis.

Discussion:

• What profilers?
  • Suggest all of:
    • grpof
      • We (at least I have) been using this from the beginning, leverages our existing experiences with parflow performance.
    • coz (plasma-umass/coz)
      • Causal profiler.
    • stabilizer (plasma-umass/stabilizer)
      • Profiler that eliminates effect so address space layout (which can apparently have incredible effects on performance, and can be sourced from benign artifacts, such as long/short usernames)
• What test cases?
  • Suggest all of:
    • ClayL
    • RU-Conus
    • TFG-Conus
    • Big Sinusoidal

Deliverables:

• Profiling results from parflow/parflow/master
• Profiling results from hydroframe/ParFlow_PerfTeam/pf_cuda
  • With neither CUDA or OpenMP enabled
  • With CUDA enabled
  • With OpenMP enabled
  • (If OpenMP and CUDA be used together) With both CUDA and OpenMP enabled.

Done When:

• Profiling results for the implementations and test cases are uploaded here.