Hmm. There are modules of NWChem that have datasets capable of scaling from 1s to 100's of thousands of cores. You take a module that doesn't scale by its nature, say you prove that you can scale it with an incompatible model, and allowing NWChem to use 75% of compute power (try seeing your utilized flop rate) make a statement about it ... and this doesn't have anything to do with anything? The particular hardware did not have good support for one-sided communication. It was a thin card that relied on host a lot and in this case, using the host gave some performance. This in my opinion is a bad way of making a point about comparing two different models. Bills point about implementations hits its sweet spot here. Please excuse my understanding if I am missing your point here.

Date: Fri, 24 Apr 2009 17:18:27 -0700 From: [email protected] To: [email protected] Subject: Re: [Mpi-forum] MPI One-Sided Communication

On Fri, Apr 24, 2009 at 05:50:23PM -0400, Vinod Tipparaju wrote:

...

The scaling in the chart discussed below is immaterial because the comparision point is flawed. I tried to convey this to some of the developers on the phone a couple of years ago but invain. The fact this claim of performance went no where beyond the white paper should convey something. Did you know dft (siosi6) doesn't scale (for various reasons) beyond 1k?

Pretty much every code and dataset have a maximum scalability point. I don't see what that has to do with anything.

The point of the chart is that a pure MPI-1 implementation scaled great and had great absolute performance on particular hardware/MPI, even though the developers were sure that specialized hardware & non-MPI underlying messaging software was needed.

-- greg

_______________________________________________ mpi-forum mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-forum