[CIG-SHORT] Run on cluster using SLURM

Brad Aagaard baagaard at usgs.gov
Fri Sep 28 13:59:55 PDT 2018


There is no scheduler named slurm, so I suspect that PyLith just used 
the same settings as the default. This may affect how the job was 
distributed among the nodes, or the default scheduler may work by 

We use domain decomposition strategy for parallelism, so to see a 
difference running in parallel, you need an appropriately sized problem. 
If the problem is small to medium sized, running on a machine with a 
faster CPU and faster memory usually provides the best speedup. This 
assumes you are already using the appropriate predonditioner, solver 
settings, and discretization for your particular problem. You may want 
to review the online tutorial on solver settings (see Session V in the 
2016 PyLith tutorial https://wiki.geodynamics.org/software:pylith:cdm2016).


On 09/28/2018 12:57 PM, Bing He wrote:
> Dear all,
> Is there anyone has successfully run Pylith using Slurm? Could you 
> please give me some suggestions?
> I wrote the configuration file and run the code like this (attached). It 
> can use many nodes and cpus, but the computing speed does not improve.
> [pylithapp]
> scheduler =slurm
> launcher =mpich
> [pylithapp.slurm]
> shell =/bin/bash
> [pylithapp.launcher]
> command =mpirun -np ${nodes}
> pylith step06.cfg \
>      --launcher=mpich \
>      --scheduler=slurm \
>      --job.queue=general \
>      --job.walltime=2*day \
>      --job.name=speed \
>      —-nodes=40—-scheduler.ppn=20\
>      --job.stdout=log \
>      --job.stderr=err
> Thanks for your time and have a nice weekend!
> Bing

More information about the CIG-SHORT mailing list