Hi David -
I am anticipating at this point that what will happen with pmemd is roughly
10-50% speedup nve, nvt, less for npt. I've tried everything reasonable.
I've analyzed what namd does, and to a smaller extent what desmond does. I
understand the tradeoffs, our strengths and weaknesses; we are actually
doing quite well all considered, even if namd beats us using ridiculous
amounts of hardware on some platforms (I still need to do some benching on
abe and lonestar to see what we can really do with these platforms now).
Basically, all the low level stuff namd does with the charm++ engine allows
them to completely overlap i/o and computation throughout a step; we can't
do that, being more constrained in a fortran and mpi model. BUT we do
everything a lot more efficiently by minding our p's and q's and getting the
most out of each node. So I am doing a bit more code but no more big
explorations; mostly got to decide on some defaults for autoswitching and
some bug fixes here and there, and test and bench. Got to catch up with Tom
on the sander patches too. I guess my manual will be an extensive readme
detailing differences between 9 and 10 at this point. Sorry this has been
such a mess; I basically got myself committed to making something happen
that really was not feasible within the constraints we have, and then killed
myself trying to force it. I'll keep you posted as I get closer.
Best Regards - Bob
----- Original Message -----
From: "David A. Case" <case.scripps.edu>
To: <amber-developers.scripps.edu>; <vbabin.ncsu.edu>
Sent: Monday, March 03, 2008 1:15 PM
Subject: amber-developers: Deadlines for Amber10
> Hi everyone:
>
> 1. Contributions to the Users' Manuals need to get to me this week. If
> you
> have not yet worked on cleaning up sections related to your contributions,
> please do so now. "Cleanup" includes proofreading for errors, but also
> includes revsing sections to make them clearer, adding in examples where
> they
> are missing, making sure that new functionality is properly described, and
> so on.
>
> 2. Bug fixes and code cleanup are needed as soon as possible. Please
> visit
> and update the bug database at http://bugzilla.ambermd.org. (In the
> initial
> "restricted area" pop-up, user name is amber, password is amberbugs.)
>
> Significant things that need attention:
>
> a. ptraj fails in a number of places
> http://bugzilla.ambermd.org/show_bug.cgi?id=25
>
> b. targetted MD has been failing forever:
> http://bugzilla.ambermd.org/show_bug.cgi?id=13
>
> c. The ti_decomp material is failing in rgroup():
> http://bugzilla.ambermd.org/show_bug.cgi?id=28
>
> d. Divcon tests are failing on many platforms. Some of this is just
> different output formatting, some may be roundoff error. Kennie
> reports that there are bug-fixes that have not been applied in
> CVS :-(.
>
> Other points are listed on the bugzilla page. We are now having users
> testing
> with Pathscale and Portland compilers, as well as on non-sgi ia64
> machines.
> This all leads to the usual mix of (mostly minor) incompatibilities.
>
> We need a final "push" to make all this happen: please do what you can.
>
> ...thanks...dac
>
>
Received on Wed Mar 05 2008 - 06:07:21 PST
