I am benchmarking my coarse-grained MD simulations with 30k beads. I noticed that MPI-GPU based mdruns used less memory and did not change much as I increased the number of nodes. Where as for MPI-OpenMP based mdruns showed increased memory usage with number of nodes.
Nodes MPI-OpenMP(GB) MPI-GPU (GB)
1 0.694 1.19
2 1.45 1.27
4 2.34 1.27
5 3.15 1.27
6 3.86 1.28
8 4.86 1.25
9 5.5 1.25
10 5.92 1.24
For all the benchmarks I have used 3 CPUs per task and 8 Tasks per node. For MPI-GPU simulations, 4 GPUs were used. I am trying to understand why MPI-GPU based mdruns are so memory efficient.
Thank you in advanced.
I am running my jobs using SLURM and I used the 'seff' command to see the total memory usage. I am not sure if it reports GPU memory.
When you use GPU version of most MD codes, the long-range interaction calculation will be transfered to GPU (and used GPU memory) for doing, only the final results will be transfered to and stored at memory on mainboard. That is understandable about your observation.
- Badges
- Science topic