Gmx mdrun performance on cluster

User discussions
1

I am submitting a simulation on a node with 48 cores, but I am choosing only 40 cores, each with 4 threads for my job as following:
mpirun -np 40 gmx_mpi mdrun -deffnm prod -ntomp 4 -plumed plumed.dat

(I am just using one node)

During the run I am getting following warning message, and the performance is also poor:
WARNING: On rank 0: oversubscribing the available 48 logical CPU cores per node with 160 threads.
This will cause considerable performance loss.

I looked at different help pages to improve the performance but keywords like -nt are not recognised on my node. Could you please help me?
Thanks in advance.

2

Hi!

You’re launching 40 ranks (-np 40), each using 4 threads (-ntomp 4), so you get 160 threads. You must ensure that the product of -np and -ntomp does not exceed the number of cores. If you need to use 40 ranks for your plumed simulation, you must use -ntomp 1.

but keywords like -nt are not recognised on my node

-nt and -ntmpi are for use with a thread-MPI build (not the “real” MPI you have). Instead of -ntmpi, you should use the -np flag of mpirun.

1 Like
3

Thanks for your reply, it really helped and resolved my issue.