Replicate RAJAPerf Experiments

[michaelmckinsey1]

Description

Replicate recent RAJAPerf experiments that were built with cmake.

This PR dovetails with LLNL/RAJAPerf#501. @daboehme I couldn't tag you there, but please help us get both RAJAPerf and Benchpark into a better state, and so they can work together.

• Enable passing additional command line arguments to raja-perf binary
• Caliper configs we need to be able to use:
  • time
  • topdown
  • gpu-side calls (name?)
  • ncu
• Replicate Datasets:
  • V100
  • MI250x
  • MI300a
  • A100
  • GH200

Adding/modifying a benchmark (docs: Adding a Benchmark)

• Modify experiments/raja-perf/experiment.py
• Modify repo/raja-perf/application.py

V100 setup reproducer

benchpark system init --dest=lassen-system llnl-sierra
benchpark experiment init --dest=raja-perf-benchmark raja-perf+cuda+strong~single_node caliper=time version=2024.07.0 scaling-factor=4 scaling-iterations=2 total_size=33554432 repfact=5
benchpark setup ./raja-perf-benchmark/ ./lassen-system/ v100-raja-perf-2025/

(text values shown on figure are raw values, not speedup)

[michaelmckinsey1]

I would like to have the option for --variants and --tunings instead of having to manually edit the execute_experiment files. The problem with this code is that when no argument is provided, the arguments are still injected, e.g. raja-perf.exe --repfact 5 ... --variants --tunings, when ideally they should not appear if not provided raja-perf.exe --repfact 5 ...

This can be done by checking in experiment.py, but that code results in the arguments appearing into experiment_run_dir (e.g. ...raja-perf.exe/raja-perf_suite_strong_1_33554432_5 --variants RAJA_CUDA --tunings block_256 ...) since execute is an experiment variable.