Hi Paul,
> How do I enable the output of the "ewald error estimate" along with
> ntpr
> info in amber11's pmemd.cuda
You can't unfortunately. It is disabled for performance reasons, read not
actually implemented. If you want to check the ewald error estimate I
recommend taking a restart file for the point you are at and running a few
steps on the CPU to see.
> I am seeing a very big increase in performance by setting nfft(1-3) to
> 64.
Yeap this is normal and is a limitation of the current NVCC compilers. Take
a read of the notes on performance on
http://ambermd.org/gpus/
Right now power's of 2 are good so 32, 64, 128, 256 are your options. It is
hoped that this will be addressed by NVIDIA with a new cufft library in the
next nvcc release.
> My box size is ~160 angstroms, so I am wondering if I can get away with
> the
> large grid size associated with using nfft=64. My benchmark runs are
> finishing ~10% faster with nfft=64 vs. nfft=128, and setting these
> values
> rather than letting pmemd set them gets me ~80-100% performance
> increase.
That sounds dangerously small for me. With PMEMD setting them by hand you
probably get something like 160 or so which is not a power of 2 and thus
causes problems. Setting it by hand to 128x128x128 should at least improve
things without too much sacrifice in accuracy. If you go all the way to 64^3
then you probably need to increase the cutoff to offset this. Man, I wish I
was an expert in PME accuracy. We need Tom Darden here to re-educate us all
in what was originally all sorted out 15 years ago.
> Any advice on assessing whether or not using nfft=64 or 128 values are
> acceptable will be greatly appreciated.
Note, if your box size is not changing much then your Ewald error estimate
should be more or less stable. I would suggest running a single point
(nstlim=1) on the CPU code with the settings you want, FFT size, cutoff,
order etc and note down what the ewald error estimate is. Then running with
these settings on the GPU will give you the same ewald error.
All the best
Ross
/\
\/
|\oss Walker
| Assistant Research Professor |
| San Diego Supercomputer Center |
| Tel: +1 858 822 0854 | EMail:- ross.rosswalker.co.uk |
|
http://www.rosswalker.co.uk |
http://www.wmd-lab.org/ |
Note: Electronic Mail is not secure, has no guarantee of delivery, may not
be read every day, and should not be used for urgent or sensitive issues.
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Tue Jun 01 2010 - 16:30:04 PDT
