[WIP] Allocation implementation for mpibind modifier #646
Conversation
| @@ -24,6 +25,7 @@ class Amg2023( | |||
| WeakScaling, | |||
| ThroughputScaling, | |||
| Caliper, | |||
| Mpibind, | |||
| name="greedy:0", | ||
| description="Run mpibind in very greedy mode", | ||
| ) | ||
| depends_on = ["mpibind"] |
There was a problem hiding this comment.
I don't think these should be passed as alternate modes for the allocation modifier: I'd prefer this was passed as something like an mpibind variable (with all these possible values)
| return f"{base_string} {mpi_string}{mpi_end}" | ||
| else: | ||
| mpi_end = self._usage_mode | ||
| return f"{base_string} {mpi_string}{mpi_end}" |
There was a problem hiding this comment.
I think this is just adding an extra option onto the mpi command invocation.
I'd recommend modifying _init_batch_and_cmd_opts to include this extra option, vs doing string manipulation on the command.
| def mpibind(self, executable_name, executable, app_inst=None): | ||
| pre_exec = [] | ||
| post_exec = [] | ||
| output_file = "{experiment_run_dir}/{experiment_name}.out" |
There was a problem hiding this comment.
(question) is .out the output captured by Ramble from running the experiment? And then parse_mpibind_output.py is searching for output from mpibind that is generated as part of running that experiment? I don't know how mpibind works: are these values determined by mpibind itself (i.e. not known before running the experiment)?
| } | ||
|
|
||
| output_file = "mpibind_log_file.json" | ||
| with open(output_file, "w") as json_file: |
There was a problem hiding this comment.
This looks relative: are you sure it runs with the basedir you want? You might want to specify an abspath (e.g. as another argument to parse_mpibind_output.
Description
Alternative implementation for mpibind(#608), moving all mpi-command editing to the allocation modifier, where this already takes place by default. Functionality remains the same.
Example command:
benchpark experiment init --dest=saxpy saxpy mpibind=v +openmpSo far, modes include: off, on, v, vv, greedy:0
Adding/modifying a benchmark (docs: Adding a Benchmark)
experiments/benchmark_name/experiment.pyto define a single node and multi-node experiments.github/workflows/run.yml