[mvapich-discuss] Odd process to CPU mapping
Jonathan Perkins
perkinjo at cse.ohio-state.edu
Tue Nov 4 15:36:00 EST 2014
On Tue, Nov 04, 2014 at 02:30:13PM -0600, Vladimir Florinski wrote:
> Hi,
>
> We are seeing a rather odd behavior since upgrading to mvapich2-x version
> 2.1 (RPM install). If one requests, say, 2 nodes with 2 octa-core CPUs per
> node (32 cores in all), the software would place two processes on each core
> of the first CPU in the node, leaving the second CPU idle. Because
> hyperthreading is disabled, this policy does not make any sense at all. Any
> ideas on what could be wrong here?
Thanks for your note. This is not intended behavior. Can you send us
the output of the following:
uname -a
ibstat
The output from your run with MV2_SHOW_CPU_BINDING=1 set in your
environment (you only need to send us the header showing the binding).
The output from your run with MV2_SHOW_ENV_INFO=2 set may also be
useful.
--
Jonathan Perkins
More information about the mvapich-discuss
mailing list