We use the --nodelist option to inform slurm it must run a job on a specific hardware node.
Slurm honors the request in most, but not all, cases. This behavior seems like slurm takes the --nodelist as a suggestion rather than a requirement. In heavy use cases, slurm seems do fit the job wherever it can.
From the HelloSameNode test directory, the slurm testing shows just 1 test case where the requirement is not honored. It is repeatable.
The log shows the --nodelist=rzwhippet69 requirement. But the code was run on rzwhippet68.
#ATS:commandLine =srun --comment="nompitype" --label -J t60_23test_file_write_t23_label_072717 --comment=nopartition --exclusive --comment="nounbuffered" --mpi
bind=off --comment=no_distribution --nodes=1-3 --cpus-per-task=1 --nodelist=rzwhippet69 --ntasks=60 --comment=nosargs /usr/WS2/dawson/Git-ATS-GitHub-SameNode/t
est/HelloSameNode/test_file_write.py nodeB_23
Writing a file n /usr/tmp on node rzwhippet68
We will need to work with the Slurm developers to resolve this issue.
For comparison, the equivalent option in Flux always works, flux takes the hardware node as a requirement.
We use the --nodelist option to inform slurm it must run a job on a specific hardware node.
Slurm honors the request in most, but not all, cases. This behavior seems like slurm takes the --nodelist as a suggestion rather than a requirement. In heavy use cases, slurm seems do fit the job wherever it can.
From the HelloSameNode test directory, the slurm testing shows just 1 test case where the requirement is not honored. It is repeatable.
The log shows the --nodelist=rzwhippet69 requirement. But the code was run on rzwhippet68.
#ATS:commandLine =srun --comment="nompitype" --label -J t60_23test_file_write_t23_label_072717 --comment=nopartition --exclusive --comment="nounbuffered" --mpi
bind=off --comment=no_distribution --nodes=1-3 --cpus-per-task=1 --nodelist=rzwhippet69 --ntasks=60 --comment=nosargs /usr/WS2/dawson/Git-ATS-GitHub-SameNode/t
est/HelloSameNode/test_file_write.py nodeB_23
Writing a file n /usr/tmp on node rzwhippet68
We will need to work with the Slurm developers to resolve this issue.
For comparison, the equivalent option in Flux always works, flux takes the hardware node as a requirement.