Performance Evaluation of Virtual Machines and Containers in High Performance Computing

Abstract

The High-Performance Computing (HPC) systems demand effective resource-utilization as well as the performance of its execution. Older models of hypervisor-based virtualization provide isolation at the expense of performance, whilst the new models of container-based virtualization (e.g., Docker, Singularity) are lightweight and near-native in terms of performance. This paper will provide a comparative performance analysis of virtual machines (VMs) versus containers on HPC based on CPU, memory, I/O throughput, and execution latency values. With the help of the latest benchmark studies and experimental evidence, we show that container-based environments tend to have lower overhead and better performance than the traditional VMs across different workloads. We also single out the important cases in which hypervisor virtualization is beneficial. The results provide a recommendation to adopt suitable virtualization software in the contemporary HPC clusters. Future work will involve the study of hybrid strategies and methods of orchestration to improve performance when using large cloud and HPC deployments