[gmx-users] Re: High frequency output slow down the simulation?

[chris.neale at utoronto.ca]

This is one reason why I developed the checkpointing script that is  
available in the wiki under beginners -> checkpointing_jobs

Often a cluster node allows you to write to ${TMPDIR} which creates a  
unique directory that is actually local and then you can copy the  
contents to your final storage space.

Chris.

-- original message --

Yang Ye wrote:
> Writing on local or remote certainly makes a difference. How about  
> the speed with no output at all?

Absolutely. Depending on the cluster setup, the best option is often
along the lines of

1) Write to disk physically associated with the CPUs (IIRC for GROMACS,
preferably the disk associated with MPI node 0, where applicable).
2) Write minimally often for your needs - using XTC rather than TRR
output wherever high precision is not required.
3) Arrange for the file to be returned to more convenient or permanent
file space at the end of the job.

This last step might be as simple as adding some cp commands to your
batch script, or perhaps issuing some "stage file out" instructions to
your batch system.

Your first port of call should be the cluster administrator /
documentation for how best you should be doing these steps. We can't
help much there.

Mark