[slurm-users] cpu allocation on a NUMA node

[Lincy]

 Hi all,
          I have a cpu allocation problem on the 2 socket 18-core machine,
with CPUSpecList=17,35 reserved.
Ideally, if I submit a job with two 17-core tasks in a node, it allocates
each task on a single NUMA node.

# *srun -N1 -n2 --cpus-per-task=**17**  numactl  -s *physcpubind: *0 1 2 3
4 5 6 7 8 9 10 11 12 13 14 15 16* cpubind: 0 membind: 0 1 physcpubind: *18
19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34* cpubind: 1 membind: 0 1
However, a job with two 16-core tasks show a bad allocation (cover two numa
nodes) even if I enforce it by --cpu-bind=mask_cpu:0xffff,0x3fffc0000

# *srun -N1 -n2 --cpus-per-task=**16**  numactl  -s*, physcpubind: *0 1 2 3
4 5 6 7 8 9 10 11 12 13 14 15* cpubind: 0 membind: 0 1 physcpubind: *16 18
19 20 21 22 23 24 25 26 27 28 29 30 31 32* cpubind: 0 1 membind: 0 1
How should I allocate it on 0-15, 18-33 for example?

Many thanks,
Chris
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.schedmd.com/pipermail/slurm-users/attachments/20190419/ae52b3ae/attachment.html>