[gmx-users] Increasing performance of siumulation in cluster

[Mark Abraham]

On Fri, May 31, 2013 at 11:30 AM, Sainitin Donakonda <saigro16 at gmail.com>wrote:

> Hi Mark,
>
> I forgot to mention about simulation it is protein ligand simulation which
> contains one protein and ligand with 20000 solute molecules..


So assuming you mean "solvent", about 60K atoms.


> and regarding
> hardware of cluster which is in my university as follows
>
> CPUs  AMD Opteron 6274   number of cores  5888   th. peak performance  51.8
> TFlops   compute nodes  4-way nodes *Saxonid* with 64 cores   nodes with 64
> GB RAM  48   nodes with 128 GB RAM  24   nodes with 256 GB RAM  12   nodes
> with 512 GB RAM  8
>

Links are better than copy-and-paste...
https://doc.zih.tu-dresden.de/hpc-wiki/bin/view/Compendium/HardwareAtlas


> Can you now suggest some methods which i can use to optimize to increase
> performance
>

Your hardware is quite useful. You should be getting better performance.
Exactly how to do that depends on the simulation (e.g. PME or not, group or
Verlet kernels). You should start with
http://www.gromacs.org/Documentation/Acceleration_and_parallelization and
experiment with the balance of MPI processes and OpenMP threads per node of
your hardware. Look at the timing breakdown at the end of the .log file,
and pay attention to all the information grompp and mdrun writes!

Mark