Hi everyone:
I've been working a bit with desmond, getting ready for a visit to Bob Duke
next week. Here are a few initial notes.
1. It is not hard to convert an amber prmtop file to desmond format. I've
done this for the "jac production NVE" benchmark that Ross put together; (note
that this uses dt=0.002, so ns/day are roughly twice that of the original
jac benchmark). I've put these in the git repo in the amber11/benchmarks/jac
subdirectory -- see the "desmond.note" file.
2. Running this on a single node (dual-quad intel 5570, 2.9 GHz) yields
the following:
a. pmemd: 11.2 ns/day (compiled with Intel 11.1)
b. desmond: 15.5 ns/day (using binary provided by DEShaw)
Remember that pmemd is running double precision, whereas desmond is single
precision. Also, these are rough numbers -- I did not run multiple times
and average, etc. But they should be in the ballpark. I used 16 threads
for both pmemd and desmond, even though there are only 8 physical cores;
the hyperthreading capability of the newer Intel cpus makes that a
significant win.
By comparison, a C1060 Tesla gpu (in SPDP mode) gives 11.9 ns/day, and a
C2050 gives 20.7 ns/day. (These are numbers from Ross' web page; my machines
give about 12.6 ns/day for S1070 and 18.6 ns.day for C2050; this means you
should probably expect some variaton in GPU speed depending on details of the
machine configuation you have.)
3. Does anyone have enough experience with NAMD or Gromacs to run this
benchmark with those codes? It would be nice to have some rough ideas of
relative performance. (I suppose using a particular teragrid machine for
all tests would be more valuable than using a particular machine I happen
to have available.)
4. The desmond tutorial (
http://deshawresearch.com/resources_desmond.html)
is really nice, and might serve as an inspiration. One thing that comes to
mind: our basic tutorial (B1) generates its input coordinates with NAB. This
is fine, execpt that it doesn't introduce users to the steps generally needed
to massage a PDB file to provide starting coordinates. Since starting from
a pdb structure is the most common way Amber is used, our tutorials are
failing to provide some key guidance. It would be very helpful (volunteers?)
to have a basic tutorial using a protein and a pdb file.
5. Istvan reports that there will be a new release of desmond within the next
month or so, but that it will not have major changes from what is available
now. Still, if you thinking of playing with this, you might want to hold off
until the newer version is available.
...dac
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Wed Jul 21 2010 - 06:30:04 PDT
