Ticket 2236

Hi there,

We've have had a chance to do our upgrade to 15.08.4 now (hooray!), and I've found an odd behaviour which I suspect is a Slurm bug.

Here's an interactive job which works as expected:

[samuel@barcoo001 ~]$ cat /proc/$$/cpuset
/slurm/uid_500/job_4993544/step_0

[samuel@barcoo001 ~]$ cat /proc/$$/cgroup
4:cpuacct:/slurm/uid_500/job_4993544/step_0/task_0
3:memory:/slurm/uid_500/job_4993544/step_0
2:cpuset:/slurm/uid_500/job_4993544/step_0
1:freezer:/slurm/uid_500/job_4993544/step_0

[samuel@barcoo001 ~]$ lstopo | fgrep Core | wc -l
1


All good.

When I SSH into that node I get *almost* the same thanks to pam_slurm_adopt, except my CPUs are not constrained..

[samuel@barcoo001 ~]$ cat /proc/$$/cpuset
/

[samuel@barcoo001 ~]$ cat /proc/$$/cgroup
4:cpuacct:/slurm/uid_500/job_4993544/step_extern/task_0
3:memory:/slurm/uid_500/job_4993544/step_extern/task_0
2:cpuset:/
1:freezer:/slurm/uid_500/job_4993544/step_extern

[samuel@barcoo001 ~]$ lstopo | fgrep Core | wc -l
16


Looking at the code for pam_slurm_adopt it appears that it just tells Slurm to adopt the job through stepd_add_extern_pid() and the rest is up to Slurm, is that correct?

This is RHEL 6.7 x86-64 with latest updates (except for libpng which came out today).

cheers!
Chris