On Fri, May 29, 2015 at 3:56 PM, Jed Brown <[email protected]> wrote:

Mark Adams <[email protected]> writes:

...

Yea, I realized that VecAssembly should see this load imbalance unless it had a barrier before its timer. So I'm not sure what is going on.

Since you don't have any outgoing entries, I think the huge MPI_Allreduce is expensive. No surprise.

This is on Edison and this was not a problem a few months ago. Edison could have changed however.