Hybrid job (MPI+OpenMP) for different AWS EC2 instances

GROMACS version: 2019.2
GROMACS modification: No
Here post your question : I have been trying to accelerate a gmx_mpi jobs on single and multiple AWS EC2 instance nodes. I am trying specific combinations of # slots and # threads. I can’t seems to get much acceleration. My cpu usage and gpu usage remains low. I have consulted online docs and course notes. I have done some benchmarking by varying numbers of slots, thread, vCPU counts, gpu counts but I cannot seem to get acceleration of even full use of the cpus/gpus. I am running Metadynmaics with the Plumed patch

Plumed gathers information to one rank of the simulation whenever the collective variables are updated. That doesn’t scale. If you want to see g that’s part of your problem, compare performance with a non plumed run on the same system. Then consider whether your collective variable action can be of lower frequency.

Thanks mabraham

Also take a look at the work by Carsten Kutzner on GROMACS performance on AWS, e.g. https://fosdem.org/2021/schedule/event/hpc_containerized_ensemble/attachments/slides/4498/export/events/attachments/hpc_containerized_ensemble/slides/4498/TalkFosdem2021.pdf