I've been running some tests on a couple clusters, and it looks like the
srun --cpu-freq=<kHz> arg is ignored when used with --cpu_bind.  Has anyone
else seen this?  Seems like it might be a bug, but I haven't looked at the
source... maybe there's a good explanation.

A simple example:
$ srun --cpu-freq=2101000 --cpu_bind=rank grep -m1 MHz /proc/cpuinfo
cpu MHz : 1200.000
$ srun --cpu-freq=2101000 grep -m1 MHz /proc/cpuinfo
cpu MHz : 2101.000

Looking at the cpufreq tree, you can see that it's not even switching to
the userspace governor:
$ srun --cpu-freq=2101000 --cpu_bind=rank cat
/sys/devices/system/cpu/cpu0/cpufreq/scaling_governor
ondemand
$ srun --cpu-freq=2101000 cat
/sys/devices/system/cpu/cpu0/cpufreq/scaling_governor
userspace

Both clusters are running slurm 14.11.6.  One runs RH6.5 computes, the
other RH7.2.

Thanks,
Kris

Reply via email to