[gmx-users] About REMD specification in mdrun argument
Mark.Abraham at anu.edu.au
Tue Dec 2 07:16:22 CET 2008
Lee Soin wrote:
> Hi, everyone!
> There's something that I'm not clear about REMD in gromacs. Since all
> the inputs are unified into a unique .tpr file, why should the REMD
> options -multi and -replex be sprcified as mdrun arguments instead of
> being integrated into the same .tpr file?
REMD is a special case of the general capacity to run multiple
independent simulations as one MPI master process (a.k.a. multiple MD).
This so-called "trivial parallelism" can allow better usage of a
parallel environment when a user has access to only coarse multiples of
physical processors (e.g. BlueGene) and needs multiple jobs run that
would be near (or past) processor scaling limits. Using fewer processors
per job can lead to better throughput, so long as one can unite several
jobs into one MPI master process to use the overarching processor
granularity well. Since in such cases, one might have different jobs
with widely different properties, integrating the effect of -multi into
the .tpr file would require that a .tpr file become a container for
(potentially) multiple copies of all the information currently found in
a .tpr file. This would work, but would require significant code
hacking, and doesn't have any clear gain cases compared with -multi and
supplying multiple independently-managed .tpr files.
Since (IIRC) the .tpr contents are REMD-unaware (for example, they might
just have different temperatures for T-coupling), moving -replex into
the .tpr/.mdp complex would require adding checks in mdrun that all the
.tpr files had matching -replex and still not serve a clear purpose.
The .tpr files contain all the inputs necessary to run a system in a
given ensemble, and the .cpt (where supplied when using GROMACS 4)
contains all the state variables for a restart. Since at REMD one can
(equivalently in theory) swap either ensemble descriptors or state
variables at exchanges, it seems elegant to keep the formats of the
files that describe those quantities "clean" with respect to REMD.
More information about the gromacs.org_gmx-users