Performant Container Support for HPC Applications

The demand for the ability to easily customize, reproduce and migrate applications and workflows has been steadily increasing among the HPC community as software environments and applications grow in complexity. Lightweight containers that are suitable for HPC applications at scale are considered to be a viable approach to meet this demand. Previous studies have addressed the performance aspects of most existing containers using microbenchmarks and revealed the performance overheads of the best implementations to be small. However, the feasibility of providing containerized, real-world HPC applications on HPC systems, and the impact on overall application performance at scale, has not yet been explored. Here we present a basic feasibility and performance study using the Singularity container. We evaluate what is required to enable container images to utilize the high-speed fabric present on most HPC systems and explore their performance by comparing real-world applications run both within a container and in the absence of a container (natively). The results indicate lightweight Singularity images are a promising approach to the HPC communities demands for not only customizability, reproducibility and portability, but also performance.