[slurm-users] Curious performance results

[Volker Blum]

Hi,

I am testing slurm 20.11.2 on a local cluster together with Intel MPI 2018.4.274 .


1) On a single node (20 physical cores) and executed manually (no slurm), a particular application runs fine using Intel’s mpirun, execution time for this example: 8.505 s (wall clock).

(this is a straight MPI application, no complications)


2) Using slurm and Intel’s mpirun through a queue / batch script,

#SBATCH --ntasks-per-node=20
…
mpirun -n 20 $bin > file.out

the same job runs correcty but takes 121.735 s (wall clock!)


3) After some considerable searching, a partial fix is

#SBATCH --ntasks-per-node=20
...
export I_MPI_PMI_LIBRARY=/usr/lib64/libpmi.so.0
srun --cpu-bind=cores -n 20 $bin > file.out

can bring down the execution time to 13.482 s


4) After changing

#SBATCH --ntasks-per-node=20
#SBATCH --cpus-per-task=2
...
export I_MPI_PMI_LIBRARY=/usr/lib64/libpmi.so.0
srun --cpu-bind=cores -n 20 $bin > file.out

finally, the time is: 8.480 s

This timing is as it should be, but at the price of pretending that an application is multithreaded when it is, in fact, not multithreaded.

***

Is it possible to just keep Intel MPI defaults intact when using its mpirun in a slurm batch script?

Best wishes
Volker

Volker Blum
Associate Professor
Ab Initio Materials Simulations
Thomas Lord Department of Mechanical Engineering and Materials Science
Duke University
https://aims.pratt.duke.edu

volke...@duke.edu
Twitter: Aimsduke

Office: 4308 Chesterfield Building

[Angelos Ching]

I think it's related to the job step launch semantic change introduced at 20.11.0, which has been reverted since 20.11.3, see 

https://www.schedmd.com/news.php For details.

Cheers,

Angelos 

(Sent from mobile, please pardon me for typos and cursoriness.)

26/2/2021 9:07、Volker Blum <volke...@duke.edu>のメール:

Hi,

[Volker Blum]

Thank you! I’ll see if this is an option … would be nice.

I’ll see if we can try this.

Best wishes
Volker