[mvapich-discuss] MV2_USE_BLOCKING=1 with ch3:psm

[Adam T. Moody]

Hi David and Leonard,
We have many QLogic clusters at LLNL.

The problem lies with Intel's PSM library itself.  PSM does not provide 
a notification mechanism to wake MPI when a message comes in, so MPI's 
only option is to poll.  Right now, MVAPICH repeatedly calls 
psm_mq_ipeek() to test for a received message.

The best you can do right now is to yeild the CPU on occassion from 
within MPI.  However, this just releases the CPU if another process is 
ready to run, so your CPU still churns at 100%.

If you value blocking behavior more than bandwidth and latency, one 
solution may be to configure MVAPICH to use verbs instead of PSM.
-Adam


David Winslow wrote:

>Leonard,
>I have the same issue. Hari suggested MV2_ON_DEMAND_THRESHOLD=1 but it
>didn't work for me. CPU usage still at 100% when just waiting. You could
>try this and see if it works for you. Let me know if it does.
>
>Thanks,
>David
>
>
>On Thu, Aug 21, 2014 at 11:29 AM, Leonardo Piga <leonardo.piga at gmail.com>
>wrote:
>
>  
>
>>Hi all,
>>
>>I am a new user of MVAPICH2.
>>
>>I am trying to run mvapich2-2.0 using MV2_USE_BLOCKING=1. However, the
>>CPU usage is still high during wait (I tested for MPI_Allreduce).
>>
>>I don't want to keep CPU usage high when a process is waiting. Is
>>there a way to do this with CH3 PSM?
>>
>>Thank you.
>>--
>>Leonardo
>>_______________________________________________
>>mvapich-discuss mailing list
>>mvapich-discuss at cse.ohio-state.edu
>>http://mailman.cse.ohio-state.edu/mailman/listinfo/mvapich-discuss
>>
>>    
>>
>
>  
>
>------------------------------------------------------------------------
>
>_______________________________________________
>mvapich-discuss mailing list
>mvapich-discuss at cse.ohio-state.edu
>http://mailman.cse.ohio-state.edu/mailman/listinfo/mvapich-discuss
>  
>