Folks -
I have a new tarball of pmemd (dated 042607) that is available on request to
dev team members. This one supports both extra points (as long as the ep
frame does not span a residue - so tip4p, tip5p are known to work) as well
as the Jarzynski Relation code that Adrian put into sander a while back.
The problems spotted in extra pnts by Dave Case earlier have been fixed
(Dave - thanks!); the Jarzynski stuff has been run against the amber9 jar
test case (which is GB). I have not had a chance to test jar + pme;
presumably no issues but you never know for sure... When using GB, pmemd is
roughly 10-15% faster for 8-16 procs; the jar test case is sufficiently
small that you can't run more processors than that (there is a restriction
on the minimum number of residues per processor currently). I would expect
pme + jar to be faster than sander but not exceedingly high scaling because
it is a nmropt .eq. 1 case, which basically means we don't do as efficient a
job of coordinate/force distribution due to uncertainties about the nmr
code. It is not logically necessary for the nmr code to be slower; there
just are inadequate test cases to mount a full assault on this ancient code
and properly parallelize it.
Best Regards - Bob
Received on Sun Apr 29 2007 - 06:07:23 PDT
