Skip to content

hpc.md

Latest commit

 

History

History
43 lines (31 loc) · 1.58 KB

hpc.md

File metadata and controls

43 lines (31 loc) · 1.58 KB
Raw

HPC usage

runledger is scheduler-agnostic. It does not submit jobs; it runs commands and records them. You typically invoke runledger sweep inside an allocation (the node(s) you already hold) and let the anytime cutoff respect the wall limit.

Launchers and binaries

# Launch each run through a templated launcher; {ranks}, {omp}, {bin} expand
runledger sweep configs.tsv \
  --launcher "mpirun -n {ranks}" \
  --bin-template "build/{bin}" \
  --budget 10 --elapse 1800 --objective min-elapsed

Command assembly precedence for each config:

  1. --launcher template (if any), then
  2. --bin-template → else the trailing -- <cmd> → else --bindir/bin → else bin, then
  3. the config's args.

omp is exported as OMP_NUM_THREADS; ranks/omp/bin are available to the templates.

Anytime under a wall limit

Set --elapse to the remaining wall time of your allocation. Before each config runledger checks spent + budget*rep + margin and stops early rather than starting work it cannot finish — so a preempted job still leaves a valid results.csv and state/incumbent.json.

Scheduler submit (not in core)

PJM/SLURM/PBS submission is intentionally out of scope for v0.1. runledger runs where it is started. Submit wrappers may arrive later as examples/adapters; they are not part of the core tool.

Resource capture

resource.json is best-effort: a getrusage(RUSAGE_CHILDREN) delta on Unix (user_sec, sys_sec, maxrss), and {} on platforms without the resource module. Treat maxrss as a high-water mark, not a precise per-run figure.