Re: CCL:EWALD PME problem
- From: Michael Crowley <crowley-.at.-scripps.edu>
- Subject: Re: CCL:EWALD PME problem
- Date: Mon, 11 Aug 2003 10:49:11 -0700 (PDT)
Hi, Rick,
The transpose in CHARMM is a problem mainly because of the limitations
in CHARMM's communication library. WIth full freedom to use
simple MPI asynchronous send and receives, we could resolve a
lot of the trouble. That remark is to justify the work I
did to try to fix an abominable communication time during
the FFT transpose (for the same basic algorithm CHARMM took
3-4 times as long as AMBER). Of the several methods I
came up with, what appeared best was a binary-tree type idea
which required a power-of-two number of PEs.
THere is another method I worked out which only requires
even number of PEs so there are always pairs communicating between
each other.
Anyway the bottom line is that the 2**N restriction will work the
best, though an asychronous method is much better.
You should try MPIFFT in your pref.dat and see if that gets you
more of what you want. Yikes, I forgot all about these restrictions
and making sure they got documented.
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
-----------------------------------------------------------------