Re: amber-developers: Testing Parallel

From: David A. Case <case.scripps.edu>
Date: Mon, 14 Apr 2008 17:09:00 -0700

On Mon, Apr 14, 2008, Lachele Foley wrote:

> Another good idea. I tried that, too (first, actually...). The result is:

There comes a point at which you have to give up/in. Test the code (serial
and parallel) on a machine you have control over. The main thing
is that you are using the same compiler as on your big parallel machine that
someone else controls. (Granted, that might mean you are not
using the same MPI installation.) There is no need to have more than a single
CPU involved--you can easily test 2, 4 or 8 threads on a single CPU.

Then, run a couple of tests by hand, submitting them to the same batch
environment you plan to run on.

There is the small danger that scali MPI has some bug that (a) is not obvious;
and (b) only shows up in some weird configuration that you didn't get to test.
You probably have to live with this. A bug of this nature might well escape
the test suite anyway.

Periodically, make sure that some job you are really running gives the same
results (for a short run) in serial and parallel. If possible, compare the
outputs of pmemd and sander.

If you do all of this, you are being far more careful than most people.

...dac
Received on Fri Apr 18 2008 - 21:19:23 PDT
Custom Search