Performance of Parallel and Storage Intensive Jobs on Clustered Singularity Containers

In this work, we present our findings regarding performance benchmarking of scientific calculations embedded into Singularity containers. We evaluate storage and parallelization performance, running jobs on a personal computer cluster. We demonstrate that containerization does not detrimentally impact the overall utilization and CPU time usage of jobs for intra-node parallelization, especially compared to the overhead of parallelization itself. For inter-node parallelization using MPI, we find that of its configurations, Singularity's bind model exhibits slightly superior efficiency. Regarding storage benchmarking, we find that using the /dev/shm in-memory filesystem yields optimal performance when it is feasible, otherwise the using host native file system with bind mounts serves as a viable alternative.