Greetings -

We have a shared memory system configured as a Slurm partition and an
OpenMP job that needs a large amount of memory but is sensitive
(performance wise) to memory latency.

The machine consists of "blades" connected by an interconnect in a enhanced
3D hypercube. Each blade contains two numa nodes (Intel processor). Each
blade shows up as a board in slurmd.

We are trying to figure out how to reliably pin a job to stay within the
blade numa nodes, and if we need to go beyond a single blade get an
adjacent blade allocated.

We are using cgroups with :

TaskPlugin              = task/cgroup,task/affinity
TaskPluginParam         = cores

SelectType              = select/cons_res
SelectTypeParameters    = CR_CPU_MEMORY

# slurmd -C
ClusterName=(null) NodeName=machinename CPUs=256 Boards=16
SocketsPerBoard=32 CoresPerSocket=8 ThreadsPerCore=1 RealMemory=4006379
TmpDisk=234080
UpTime=5-23:39:30

# slurmd -V
slurm 14.03.11


This bug gets at what we're trying to figure out, but not exactly:

http://bugs.schedmd.com/show_bug.cgi?id=2109


CR_Socket or CR_Board isn't necessarily what we are after - we do have a
fair amount of 1 core high throughput jobs we want to handle also.

Any pointers would be greatly appreciated!

Jeff

Reply via email to