[mpich-discuss] MPICH2 uses network to communicate instead of share memory on this NUMA architecture

[Rajeev Thakur]

Did you specify placement of processes using a hostfile as described here: http://wiki.mcs.anl.gov/mpich2/index.php/Using_the_Hydra_Process_Manager

On Oct 31, 2012, at 3:14 AM, CLARK wrote:

> the architecture of CPU is：
> <2CF9F3FD at 30170568.CADD9050>
> this CPU has 8 cores，2 nodes，node 0 and node 1，each node has 4 cores，only one node‘s Memory Controller connected memory. I think it is an NUMA architecture. the node in the right accessing memory is slower than the left one.
> 
> In this architecture，i config mpich2 1.5：./configure --prefix=XXX --with-device=ch3:nemesis --with-shared-memory=mmap, then make and make install.
> 
> I found that，when running 2 processes on this CPU，network card lo has sended and received data，but in other NUMA architecture，network card lo has no sended and received data.
> 
> I donot know why，maybe mpich cannot identity this architecture，or is there anything wrong with my OS？
> And，which codes and conditions are nemesis used to choose the way to communicate？
> 
> thank you.
> _______________________________________________
> mpich-discuss mailing list     mpich-discuss at mcs.anl.gov
> To manage subscription options or unsubscribe:
> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss