I have for the first time tried to get affinity to work on my cluster, but seem to misunderstand something.
Version 16.05.4. Proctracktype=proctrack/cgroup TaskPlugin=task/cgroup cgroups.conf on compute node says: CgroupAutomount=yes CgroupReleaseAgentDir="/etc/slurm/cgroup" ConstrainCores=yes TaskAffinity=yes ConstrainRAMSpace=yes AllowedRAMSpace=100 ConstrainSwapSpace=yes AllowedSwapSpace=5 ConstrainDevices=no hwloc and hwloc-devel are installed, and rpmbuild is run with "-with hwloc". "scontrol show job" indicates that affinity is configured. A/ Can you in your sbatch command ask that all requested cores (like in -n 8) will be placed within a single socket? How? (It seems not, and in a change log for version 14.03.8 I find that "--cpu_bind" is deprecated from sbatch and salloc.) B/ In you jobscript, can you ask that all requested cores in a srun call (like in -n 8) will be placed within a single socket? How? Cheers, -- Lennart Karlsson, UPPMAX, Uppsala University, Sweden
