[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [ESPResSo-users] Multiple instances of espresso on one node

From: Rudolf Weeber
Subject: Re: [ESPResSo-users] Multiple instances of espresso on one node
Date: Mon, 11 Mar 2019 10:41:51 +0100
User-agent: Mutt/1.9.4 (2018-02-28)

Hi Clemens,
On Thu, Mar 07, 2019 at 09:43:32PM +0100, Clemens Jochum wrote:
> I want to optimize my use of HPC-nodes. Is it possible to have two (or
> more) instances of ESPResSo running on a single node (with slurm in my
> case)?
This is not Espresso specific and highly dependent on the cluster installation.
On our institute's cluster, I used the following for parallel tempering 
simulations, in which several Espresso instances using one core each are 
running on the same node and exchange information via network sockets.
The same will also work for fully independent instances of Espresso, although 
not all cluster operators are happy about that use.

Below script is the run script which is submitted (in this particular case on 
16 nodes).
It launches the individual Espresso instances in the background (via Bash's &) 
and waits for completion of all background tasks at the end.
The launch command for the individual instances is probably cluster specific.


# These may or may not be needed in a particular case
ulimit -n 4096
ulimit -u 4096

# here the Espresso instances are launched
for i in {1..256}
  # on this particular cluster, it was necessary to disable binding of 
  # instances to particular cores. That may or may not be needed on other 
  srun -n1 --cpu_bind=no,v taskset 0xffffffff ./pypresso arguments  
>8output-rank-$i.out &
  sleep 0.1

On a different system, it looked like this:
for i in {1..16}
mpirun -np 1 --bind-to none pypresso arguments >$i.out 2>$i.err &

Some clusters offer queues for smaller jobs with shared usage of a node. If 
that is available, it should be used for individual jobs.
Also note that Espresso's performance tends to be limited by memory bandwidth 
rather than by computing power, so you might not get a linear performance gain 
when running more instances.

Hope that helps.
Regards, Rudolf

Dr. Rudolf Weeber
Institute for Computational Physics
Universit├Ąt Stuttgart
Allmandring 3
70569 Stuttgart
Phone: +49(0)711/685-67717
Email: address@hidden

reply via email to

[Prev in Thread] Current Thread [Next in Thread]