Re: NAMD is too slow?!

From: Axel Kohlmeyer (akohlmey_at_cmm.chem.upenn.edu)
Date: Sun Mar 29 2009 - 10:23:44 CDT

On Sun, 2009-03-29 at 13:37 +0400, DimitryASuplatov wrote:
> Hello,
>
> I am calculating a system of 700 aa protein in a water box (15
> angstroms around the protein), using electrostatic switch 10A and
> cutoff 15 A. PME is off. The estimation on 64 Intel Xeon 3Gz cores is
> 0.44 days/ns and about 4 days for 10ns. I was using intel compilers,
> mvapich and the following options
> CHARMM: ./build charm++ mpi-linux-ia64 icc ifort --no-shared -O
> -DCMK_OPTIMIZE=1 -I/usr/lib/mvapich-intel-x86_64/include/
> NAMD: ./config tcl fftw Linux-ia64-MPI-GM-icc

that is probably not too far off, but you should note that ia64
refers to itanium architecture and not x86_64. in namd2.6 that
is labeled as amd64.

> and I use mpirun -np 64 namd2 <config> to run the program
>
> I was using gromacs previosly and I find it to be 10 times faster.

there are a number of issues here. first of all, you should make a
scaling test, i.e. run 2, 4, 8, 16, 32, 64 cpus to see whether you are
still within the range where namd scales well. due to the advanced load
balancing and latency hiding options in namd, which help a lot
normally, the performance goes down much more rapidly, if you use too
many cpus.

second you should make sure that you are comparing apples and apples.
gromacs uses a lot of tricks to cut corners. you have SIMD assembly
loops for certain targets and gromacs uses a square root approximation
from a table lookup plus one iteration of newton raphson.

to get the equivalent with NAMD you should download the latest beta,
enable the SIMD/SSE intrinsics for the non-bonded loops, which also
requires that you use the correct compilation target and finally you
should investigate other speed enhancing techniques, like multi-timestep
or the settings for the langevin thermostat.

cheers,
    axel.
>
> My question is -
> is this a common estimate for a system of that size or I have problems
> with my cluster or my installation?

i'd say you need to learn more about how namd works and what kind of
hardware you have. ;-)
>
> Thanks
> SDA

-- 
=======================================================================
Axel Kohlmeyer   akohlmey_at_cmm.chem.upenn.edu   http://www.cmm.upenn.edu
   Center for Molecular Modeling   --   University of Pennsylvania
Department of Chemistry, 231 S.34th Street, Philadelphia, PA 19104-6323
tel: 1-215-898-1582,  fax: 1-215-573-6233,  office-tel: 1-215-898-5425
=======================================================================
If you make something idiot-proof, the universe creates a better idiot.

This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:52:32 CST