Re: AMBER: sander on multiprocessors

From: Stef <steletch.free.fr>
Date: Tue, 06 Jul 2004 10:42:36 +0200

Le mar, 06/07/2004 à 09:38 +0200, cailliez a écrit :
> Dear AMBER users,
>
> I am running AMBER7 on a cluster of SGI Origin 3800 with mpi.
> I did a test by changing the number of processors to use, and I observed
> differences
> in the results.
> My test ran only during 3 ps but I already saw differences after 500
> steps (1 ps).
> I do not see clearly why there should be differences when I change the
> number
> of processors since all the processors are identical as far as I know.
> Does anyone ever experienced this ? Or does anyone know if these
> differences are normal
> and where they come from ?
>
> Thank you in advance for your answers,
> Sincerely yours,
> Fabien

As explained by Robert Duke and others about a year ago when i asked
something similar, it is a normal behaviour.

In fact except running all calculations on the same machine, you'll
always get differences. The network introduces some incoherencies in
data (some little data corruption), the load balancing itself can lead
to slight differences, the seed is another point, ... Of course
different architectures will also lead to small discrepancies : 64/32
bits, AMD/Intel, SGI/Sun/IBM/HP, ... The calcultations should stay the
same for little time but will always diverge after a while. But as was
told, the algoritm contains more approximations than this hardware
differencies, so finally you must now it exists but given that, you can
trust any trajectory, as far as it seems 'reasonable' (no blowup for
instance).

If you want more info, search for his posts on the archive.

Best wishes,

Stef

-- 
http://www.steletch.org
Fingerprint: 1E83 26EE A2F1 87E5 5865 4DBF F2B4 9083 F44C AE88
GPG key . http://www.steletch.org/article.php3?id_article=13
Mandrakelinux release 10.1 (Cooker) . http://www.mandrakesoft.com/
10:35:37 up 13:38, 7 users, load average: 0.44, 0.38, 0.30
-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber.scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
Received on Tue Jul 06 2004 - 09:53:01 PDT
Custom Search