job-output shows the scheduler spool files crewster pointed -o / -e at. Some MPI runtimes route rank stdout elsewhere: on one Fujitsu TCS deployment, mpirun under PJM was observed writing per-rank output to <workdir>/output.<jobid>/<rank>/1/stdout.1.0 while leaving the -o spool empty (the exact layout likely varies by site configuration), so a COMPLETED job that produced output looks silent in crewster job-output.
This routing is site/runtime policy rather than part of crewster's contract, so hardcoding the Fujitsu layout would be wrong. Two mechanism-only options, composable:
- a config such as
[pjm].extra_output_globs (with {job_id} substitution): when the spool is empty, job-output falls back to files matching the globs, keeping the site-specific layout in user config and only the fallback mechanism in crewster
- minimal variant: when the job is terminal and the spool is empty, print a one-line hint that MPI runtimes may route rank output to job-local files
job-outputshows the scheduler spool files crewster pointed-o/-eat. Some MPI runtimes route rank stdout elsewhere: on one Fujitsu TCS deployment, mpirun under PJM was observed writing per-rank output to<workdir>/output.<jobid>/<rank>/1/stdout.1.0while leaving the-ospool empty (the exact layout likely varies by site configuration), so a COMPLETED job that produced output looks silent increwster job-output.This routing is site/runtime policy rather than part of crewster's contract, so hardcoding the Fujitsu layout would be wrong. Two mechanism-only options, composable:
[pjm].extra_output_globs(with{job_id}substitution): when the spool is empty,job-outputfalls back to files matching the globs, keeping the site-specific layout in user config and only the fallback mechanism in crewster