AMBER Archive (2005)

Subject: Re: AMBER: maximum number of processors

From: Michael Crowley (crowley_at_scripps.edu)
Date: Thu May 12 2005 - 12:00:09 CDT

Dear Murat,

As already mentioned, the parallel efficiency always decreases with more
processors so that eventually you will cross over the point were adding
processors does not speed up the simulation and even slows it down. There
are many sources of this slowdown, mostly in the communications between
processors which increases the more processors you have. Even when the
increase in communication time is very small, the work load per processor
becomes so small that the communications time can swamp your calculation.

On the other hand, it is possible that you are not actually getting 32
processors, you should verify that you get the number you ask for.

Check your timings in the output. You should see the "Direct Ewald" time
or Nonbond force time go down for more processors. If so, you can get an
idea of where the time is going by examining the other timings, in
particular, the communications times.

I hope that helps a little more,
Sincerely,
Mike

-----------------------------------------------------------------
Physical mail: Dr. Michael F. Crowley
                 Department of Molecular Biology, TPC6
                 The Scripps Research Institute
                 10550 North Torrey Pines Road
                 La Jolla, California 92037

Electronic mail: crowley_at_scripps.edu
Telephone: 858/784-9290
Fax: 858/784-8688
-----------------------------------------------------------------

-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber_at_scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo_at_scripps.edu