sandialabs / LoQS / 33094065631

89%

push

github

Performance-profiling tools for ParallelStrategy configs (#127)

* Add profile_strategies performance-profiling harness to paralleltools

Adds a caller-driven tool for measuring real wall-clock time and
resource usage across a chosen set of ParallelStrategy configurations
against a chosen working example.
- Each dispatched chunk self-reports its own worker's peak
  memory/CPU usage via new ParallelStrategy fields, correct even for
  a genuinely multi-node strategy.
- profile_strategies/ProfileResult/format_profile_table/
  plot_profile_results report timing, resource usage, a warmup
  option, and speedup vs. a fully-serial baseline strategy.
- program_executor now also accepts an ExecutorSpec, resolved lazily
  so several differently-sized strategies can coexist safely despite
  loky's process-wide singleton pool.
- reuse_slurm_allocation lets a submitit-based sweep dispatch against
  one already-held SLURM allocation instead of each queueing
  independently.
- Fixes CPU-percent measurement, which previously measured usage
  since each process started rather than since the last sample.
- Adds psutil to the parallel/mpi extras and updates the
  parallel-execution tutorial.

* Fix cross-strategy executor reuse in profile_strategies

Ensures profile_strategies gives honest, independent timing measurements when comparing ParallelStrategy configs with different executor sizes.
- Shut down and reset ExecutorSpec-resolved program executors after each strategy's measurement, so loky's process-wide singleton can't let one strategy's warm workers leak into another's timing.
- Warn when a strategy's timed repeats regress sharply from its fastest measurement, pointing to loky's default 10-second idle timeout and how to override it.

* Rewrite the parallel tutorial's profiling section around a real workload

Combines "Comparing worker shapes" into "Performance profiling" and replaces the trivial toy-program measurement with a real fault-tolerant workload, since th... (continued)

2981 of 3546 branches covered (84.07%)

Branch coverage included in aggregate %.

9568 of 10532 relevant lines covered (90.85%)

0.91 hits per line