I tried the following since I posted last. 1. Removed all cgroup/cpuset configuration within slurm.conf 2. Allocated 24 (out of 48) CPUs for slurm in slurm.conf 2. Created an external cpuset with the constraints I wanted (cpu list) and moved munged/slurmd/slurmctld into this cpuset. 3. Started jobs (Now, all jobs will be scheduled on the CPUs I want slurm jobs to use)
Now, I have what I need, but.. I guess slurm still thinks it can schedule jobs on any 24 of the 48 CPUs, while all jobs actually get scheduled on the CPUs that are allowed for these jobs (via cpusets). Can this cause any inconsistencies in Slurm? Thanks, Pavan On Wed, Apr 5, 2017 at 11:34 AM, pavan tc <[email protected]> wrote: > Hi, > > [I had sent this mail without subscribing to the group. So, wasn't sure if > it made it. Subscribed now. This is a resend]. > > I have a node with 48 CPUs. I need to set aside 24 of these CPUs for Slurm > and the remaining for non-Slurm workloads. > > For example, I want slurm to use only the even CPUs - 0,2,4,..46 > > I have configured the number of CPUs available for Slurm as 24 in the > "Nodes" config, but I guess this only means use *any* 24 of the available > 48. That is not my requirement. I need slurm to use only specific CPUs. > > I spent a good deal of time looking for this in the group archives and > going through the slurm.conf and cgroup.conf manpages but have been unable > to find the right configuration parameter to bring this about. > > How do I go about it? > > Thanks, > Pavan > >
