On Tue, 2012-03-27 at 14:35 +0200, Elisa Frezza wrote:
> Thank you very much.
>
> I would like to ask you another question.
> In your opinion, how is the performance of distance and angle restraints
> with nmropt=1 to make umbrella sampling in the new version of Amber?
>
> Elisa
The approach taken to get it working is to synchronize coordinates and
forces to the CPU in order to carry out the nmropt calculations there.
The upside of this is the ease of implementation. The downside is that
it amounts to 3 communications of 3*natom doubles across the PCIe bus
_each_ time step (when normally we communicate maybe 2 or 3 doubles).
Therefore, I would expect the performance hit to depend on the number of
atoms you're simulating -- the more atoms, the higher the performance
hit will be. It also depends on your hardware. The better hardware you
have, the larger the performance hit will be (since the bus hasn't
gotten faster, but the GPU has). Tentative estimates put the hit around
10-20% for GB simulations, and upwards of 35-50% for explicit solvent
simulations. The C1060 I tried was at the low end of these performance
hits and the numbers for a C2090 I saw were at the higher end. Our
tests were certainly not exhaustive when testing the number of atoms in
the system, hence the 'tentative'.
The "correct" way of doing this is to write a kernel for the restraints
so all of these calculations are done on the GPU, in which case it
becomes negligibly more expensive than a standard simulation without
restraints. Due to time constraints, though, this won't happen before
the release.
HTH,
Jason
--
Jason M. Swails
Quantum Theory Project,
University of Florida
Ph.D. Candidate
352-392-4032
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Tue Mar 27 2012 - 06:30:03 PDT
