[mpich-discuss] Node vs CPU control

[Dave Goodell]

If you drop the ":4" from the end of each line of your hostfile then  
processes will be assigned in round-robin order (nodeA, nodeB,  
nodeC, ..., nodeA, nodeB, etc).  This will spread your processes out  
in the 16-process case, but in the 64-process case your ranks will not  
be assigned sequentially to each node.  That may or may not be a  
problem for your applications.

If you want a sequential mapping sometimes and a round-robin mapping  
other times, you will need to use two different host files and specify  
the non-default one via "mpiexec -machinefile your_file OTHER_ARGS".

-Dave

On Mar 19, 2010, at 11:22 AM, Robertson, Andrew wrote:

> Folks,
> I have a 16 node cluster with 4 cpus per node. I would like to just  
> leave an mpd up and running that was booted on all 16 nodes with  
> four cpus per node.
> Ie
> NodeName:4
> On each line of my hosts file
>
> Is there a way to force jobs which use less than the full cluster  
> to  spread out and use an equal number of CPUs on each node? I find  
> that  if for instance I ask for 16 cpus I get four nodes with four  
> cpus each, I would like to get 16 nodes with on cpu each?
>
> Still getting the hang of all this
> Any help is appreciated
>
> Thanks
>
>
> -Andy
>
>
> _______________________________________________
> mpich-discuss mailing list
> mpich-discuss at mcs.anl.gov
> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss