[slurm-users] TaskPlugin task/affinity and task/cgroup
Max Quast
max at quast.de
Tue Oct 20 13:14:32 UTC 2020
Dear Slurm users,
i would like to use the taskplugin cgroups on my system. I have configured
cgroups as recommended by slurm:
cgroup.conf:
CgroupAutomount=yes
CgroupMountpoint=/sys/fs/cgroup
TaskAffinity=no
ConstrainCores=yes
slurm.conf:
...
TaskPlugin=task/affinity,task/cgroup
#TaskPluginParam=Sched
...
# COMPUTE NODES
GresTypes=gpu
NodeName=lsm[216-217] Gres=gpu:tesla:1 CPUs=64 RealMemory=192073 Sockets=2
CoresPerSocket=16 ThreadsPerCore=2 State=UNKNOWN
PartitionName=admin Nodes=lsm[216-217] Default=YES MaxTime=INFINITE State=UP
It works, but the performance is much worse than before with
TaskPlugin=task/none.
What I noticed:
With TaskPlugin=task/none the tasks often change the execution unit
(thread). With affinity/cgroups the tasks remain on their assigned thread
from start to finish.
As an example:
With srun -n9 -N2 ./prog and TaskPlugin=task/affinity,task/cgroup, the tasks
are divided among the threads:
node 1: 1,17,33,49,34
Node 2: 1,17,33,49
Execution time: ~210 sec
With srun -n9 -N2 ./prog and TaskPlugin=task/none the taks change the
threads during execution.
execution time: ~140 sec (like manual with mpirun)
What are proper parameters for task affinity and cgroups?
Thanks for any help :)
-max
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.schedmd.com/pipermail/slurm-users/attachments/20201020/539f4d8b/attachment.htm>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/pkcs7-signature
Size: 5014 bytes
Desc: not available
URL: <http://lists.schedmd.com/pipermail/slurm-users/attachments/20201020/539f4d8b/attachment.bin>
More information about the slurm-users
mailing list