[Beowulf] wall clock time for mpi_allreduce?

[James Cownie]

On 11 Sep 2010, at 03:46, xingqiu yuan wrote:

> Hi
> 
> I found that use of mpi_allreduce to calculate the global maximum and
> minimum takes very long time, any better alternatives to calculate the
> global maximum/minimum values?

Before pinning the blame on allreduce, are you sure that you're not seeing the effects
of load imbalance?

How are you measuring the time for the reduction?
Are you measuring the time at a single node, or at every node? (The reduction can't complete
until all the nodes "check in"...)
Have you looked at the allreduce time if you insert a barrier before the reduction?
(That won't help your overall performance, but may make it clear where the problem
really is...)

--
-- Jim
--
James Cownie <jcownie at cantab.net>


-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.scyld.com/pipermail/beowulf/attachments/20100912/f954d893/attachment.html