[MPICH] any way to ask nemesis to turn-off and turn of active polling ?

Darius Buntinas buntinas at mcs.anl.gov
Mon Dec 17 12:50:12 CST 2007


Try setting the processor affinity for the "average" processes (map each 
one to its own processor).  If you have a kernel with the "broken" 
sched_yield implementations, that may not help.

If you have a "broken" sched_yield implementation, you can try doing 
this as root:
   sysctl kernel.sched_compat_yield=1
or
   echo "1">/proc/sys/kernel/sched_compat_yield

-d


On 12/17/2007 11:35 AM, chong tan wrote:
> Yes, in a very subtle way which has major impact on performance.  I will 
> try to decribe it a litle here:
>  
> system has 32G, total image 35G.  Load is a litle offbalance 
> mathematically, 4X dual core, running 5 processes.
> 4 processes are the same size, each runs on a CPU.  the last process is 
> very small, about10% of others, run
> on a core of one of the CPU.  SO 1 CPU runs 2 procs: average (P1)one and 
> light one (P2).
>  
> All proc do first MPI comm in a fixed algorithmic point.  The 'useful' 
> image is about 29G at that point, and should
> fit into the physical memory.  P2 get there in a heart beat, then 
> others., followed by P1  which took another 60+ minutes
> to get there.  If I combine P1 and P2 into 1 process, then I don;t no 
> see this extra delay.
>  
> tan
> 
> 
>  
> ----- Original Message ----
> From: Darius Buntinas <buntinas at mcs.anl.gov>
> To: chong tan <chong_guan_tan at yahoo.com>
> Cc: mpich-discuss at mcs.anl.gov
> Sent: Monday, December 17, 2007 8:02:17 AM
> Subject: Re: [MPICH] any way to ask nemesis to turn-off and turn of 
> active polling ?
> 
> 
> No, there's no way to do that.  Even MPI_Barrier will do active polling.
> 
> Are you having issues where an MPI process that is waiting in a blocking
> call is taking CPU time away from other processes?
> 
> -d
> 
> On 12/14/2007 04:53 PM, chong tan wrote:
>  > My issue is like this :
>  > 
>  > among all the processess, some will get to the point of first MPI
>  > communication points faster than
>  > than other.  Is there a way that I tell nemesis to start without doing
>  > active polling, and then turn
>  > on active polling with some function ?
>  > 
>  > Or should I just use MPI_Barrier() on that ?
>  > 
>  > thanks
>  > tan
>  > 
>  >
>  > ------------------------------------------------------------------------
>  > Be a better friend, newshound, and know-it-all with Yahoo! Mobile. Try
>  > it now.
>  > 
> <http://us.rd.yahoo.com/evt=51733/*http://mobile.yahoo.com/;_ylt=Ahu06i62sR8HDtDypao8Wcj9tAcJ 
> 
>  >  >
> 
> 
> ------------------------------------------------------------------------
> Be a better friend, newshound, and know-it-all with Yahoo! Mobile. Try 
> it now. 
> <http://us.rd.yahoo.com/evt=51733/*http://mobile.yahoo.com/;_ylt=Ahu06i62sR8HDtDypao8Wcj9tAcJ 
>  >




More information about the mpich-discuss mailing list