Fermi can be fixed.
Pre-Fermi is hosed...
GLYCAM-06EP creates large numbers of adjacent atoms with huge numbers of
exclusions (up to 33 per atom) - more than I've ever seen before.
Fortunately, I was forced to migrate exclusion data for Kepler into L1 so
it handles it correctly OOTB.
Fermi is currently hosed, but it can be switched to do the same but I
suspect a perf hit will ensue.
Today's solution is to run on Kepler. And given how cheap and powerful
GTX680s are, I can't see this causing a lot of frowning.
On Mon, Mar 4, 2013 at 3:31 PM, Jodi Ann Hadden <jodih.uga.edu> wrote:
> Probably best to wait and patch when the fix is complete, just for the
> sake of simplicity. I'm probably the only person in the whole world using
> GLYCAM-06EP right now anyway. ;-)
>
> On Mar 4, 2013, at 1:32 PM, Jason Swails <jason.swails.gmail.com> wrote:
>
> > I will wait until there is a Fermi fix before making the patch (unless we
> > should just release this now and release a Fermi patch when it's
> available?)
> >
> > On Mon, Mar 4, 2013 at 1:26 PM, Jodi Ann Hadden <jodih.uga.edu> wrote:
> >
> >> I've run several MDs with this now using pmemd.cuda. Seems to work, as
> >> long as you're on a Kepler card anyway, as we discussed. Compiled with
> GNU.
> >> Thanks Jason! :-)
> >>
> >> On Feb 28, 2013, at 12:06 PM, Jason Swails <jason.swails.gmail.com>
> wrote:
> >>
> >>> This email is primarily targeted toward the GLYCAM folks, so sorry to
> the
> >>> rest of you for the noise.
> >>>
> >>> Jodi posted a recent reminder of this bug on bugzilla, which made me
> >>> realize how similar it was in some respects to a bug that I recently
> >>> (within the last 2 weeks) found in my own code. Sure enough, it was
> >> caused
> >>> by an array overflow due to an allocation of too little space -- it
> was a
> >>> pretty quick find when I knew what to look for and had the sander code
> to
> >>> compare with. The overflowed array was a 1-4 non-bonded pair array,
> >> which
> >>> is why all the EP-water models were unaffected.
> >>>
> >>> This was more of an informational message to the interested parties not
> >>> following the Amber git emails and/or the relevant bugzilla thread.
> >>>
> >>> I only tested this with the Intel compilers, so if someone in the
> GLYCAM
> >>> community can vet this with their systems and make sure it's the
> 'right'
> >>> fix (for GNU, too, preferably), then I can turn this into a bugfix.
> As a
> >>> side-note, the test system provided in bugzilla had different default
> >> nfft1
> >>> for sander and pmemd, so I had to set the FFT ewald parameters in the
> >> input
> >>> file to compare, but sander and pmemd gave identical answers.
> >>>
> >>> All the best,
> >>> Jason
> >>>
> >>> P.S. pmemd.cuda also appears to work for the test case, so enjoy the
> >>> sander->pmemd.cuda performance boost ;).
> >>>
> >>> --
> >>> Jason M. Swails
> >>> Quantum Theory Project,
> >>> University of Florida
> >>> Ph.D. Candidate
> >>> 352-392-4032
> >>> _______________________________________________
> >>> AMBER-Developers mailing list
> >>> AMBER-Developers.ambermd.org
> >>> http://lists.ambermd.org/mailman/listinfo/amber-developers
> >>>
> >>>
> >>
> >>
> >>
> >> _______________________________________________
> >> AMBER-Developers mailing list
> >> AMBER-Developers.ambermd.org
> >> http://lists.ambermd.org/mailman/listinfo/amber-developers
> >>
> >
> >
> >
> > --
> > Jason M. Swails
> > Quantum Theory Project,
> > University of Florida
> > Ph.D. Candidate
> > 352-392-4032
> > _______________________________________________
> > AMBER-Developers mailing list
> > AMBER-Developers.ambermd.org
> > http://lists.ambermd.org/mailman/listinfo/amber-developers
> >
> >
>
>
>
> _______________________________________________
> AMBER-Developers mailing list
> AMBER-Developers.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber-developers
>
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Mon Mar 04 2013 - 17:00:02 PST
