[gmx-users] Load imbalancing

[Mark Abraham]

Hi,

On Mon, 9 Sep 2019 at 12:01, Bratin Kumar Das <177cy500.bratin at nitk.edu.in>
wrote:

> Dear all,
>             I am running REMD simulation with 65 replicas. When the
> simulation is running , I checked the .log file for every replica. In some
> replicas I am getting more than 10% load imbalancing. How to solve that.


That's common to all kinds of simulations, a geometric decomposition of the
work of a simulation is going to have imbalance. How that arises and what
if anything might be done is not related to REMD (Although e.g. NPT REMD
can make it worse.)

If
> I use -ntmpi and -ntomp option ..mdrun is throwing a fatal error
>
>  Fatal error:
> Setting the number of thread-MPI ranks is only supported with thread-MPI
> and
> GROMACS was compiled without thread-MPI
>
> In the document (
> http://www.gromacs.org/Documentation/Acceleration_and_parallelization)...it
> is mentioned that -ntmpi can not be used in multi-simulation runs.....
> So how i can avoid load imbalancing in remd simulations...
>

Changing the number of ranks does affect imbalance, but the relevant thing
is how long you can afford to take rnning the simulation. Running one rank
per simulation has no load imbalance, but lowest throughput.

One can use -ntmpi to change the number of ranks, but as the error message
says, that's only for a thread-MPI build, and you need a real MPI build to
run REMD.

Mark