[mpich-discuss] Any application using mpi_barrier a lot?

[Darius Buntinas]

I did some barrier optimization work in grad school and had the same
problem of finding applications that use a lot of barriers.  Basically,
if you have an application that uses a lot of barriers it's probably not
going to perform well at large scale, so most "real" apps don't use many
(or any) barriers, sorry.  Another point to consider is that much of the
time processes spend waiting in a barrier may not be due to the latency
of the barrier operation itself, but rather due to waiting for other
processes to reach the barrier.  So even if you reduce the barrier
latency to zero, you may still not see much improvement.

This is not meant to discourage you.  There are many collectives that,
when optimized can show significant application performance improvement.
 But I think barrier isn't one of them.

-d

On 03/10/2009 03:00 PM, Shanyuan Gao wrote:
> My current research is trying to rewrite some collective MPI operations
> to work with our system.  Barrier is my first step, maybe I will have
> bcast and reduce in the future.  I have done some synthetic tests.  Now
> I want a real application to measure the performance improvement versus
> normal MPI_Barrier.  And the improvement can only be measured if the
> barriers are executed many times.  If you know any, please let me know. 
> Any comment are welcomed, thanks!
> 
> 
> Shan