[Mckernel-users 38] McKernel on multi-sockets

[FINIEL, JEREMIE]

Hello Balazs,

We found something strange with McKernel when used multiple sockets.
When executing an HPCCG benchmark with openMP threads, it take much more time when threads are in different sockets.
We used numactl to set a cpu mask. (OMP_NUM_THREADS=2 ./mcexec numactl -C 0,2 ~/HPCCG 104 104 104)
For example, using HPCCG with 2 openMP threads, it take around 4.66s with Linux and McKernel using both thread on one socket, and it take 5.42s in McKernel using different sockets when Linux take always 4.66s.
Are you aware of this issue?
Is there a mistake in our execution environment?

We tried to use the -c parameter of mcexec to set a cpu mask in McKernel application, but it look like it does not have any effect on placement. Is it a deprecated feature or is it not its purpose?

Thank you in advance for your help.

Jérémie Finiel


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.pccluster.org/pipermail/mckernel-users/attachments/20170601/83e455bb/attachment.html>