Performance Analysis of Containerized OrangeFS in HPC Environment

Compares five container runtimes hosting the OrangeFS parallel file system; Singularity came closest to bare metal.

Containerization is on the rise in cloud computing due to its benefits in reproducibility, portability and lightweight properties. The need for scalable and reproducible products and the necessity to interoperate on everything from local machines to large-scale HPC resources have made HPC solutions adopt containers expeditiously. The necessity of developing HPC-focused containers, however, appeared from various scaling and security issues of general-purpose containers such as Docker. This paper specifically examines the I/O performance of different HPC-focused container solutions on OrangeFS, a parallel storage system. Our analysis shows that Singularity covers almost all aspects, from mobility to security concerns, with an almost perfect score. On the other hand, LXC falls behind at I/O performance, whereas Docker registers an unnecessary amount of network packets under the same workload. Our aim is to provide such insights to HPC practitioners so that they can obtain the best I/O performance in their solutions.