Testing MPI Applications Performance in Singularity Containers

Abstract

The papers covers field experience of the testing performance of MPI applications when running in Singularity containers compared with bare metal. Using computing containers is a major trend in many areas, including high performance computing. Traditionally Docker is the most commonly used container provider. However, using Docker in HPC is not that native, compared with Internet or Enterprise use cases. One of the major concerns is the fact that HPC clusters depend heavily on high speed network, that is based usually on Infiniband interconnect, and Docker was not designed with this environment in mind. Singularity container technology was created recently as an alternative to Docker, but designed from ground up for HPC use case. Among other features, Singularity uses completely different approach to organizing network communication between containers that should fit better for distributed applications that rely on MPI over high speed low latency network. This paper concentrates on testing Singularity containers performance when running High Performance Linpack test suite and using MPI over Infiniband for network communication. The paper shows that Singularity containers allow achieving performance metrics that corresponding to the values for bare metal in the default container setup, without having to perform a separate network configuration on the host system side. At the same time, it was found that libraries mismatch on the host system and on container side may cause unexpected significant performance degradation.