GPGPU virtualisation with multi-API support using containers

Abstract

Virtualisation of GPGPUs using PCI-Passthrough is limited to costly specialised hardware. API-Interception provides an alternative software-based approach to GPGPU virtualisation that has been shown to provide good performance and increased utilisation on High Performance and High Throughput Computing systems. Furthermore, user applications can transparently access many non-local GPGPU resources. However, current API-Interception implementations have either limited batch system support or none at all. This paper introduces a new multicomponent system that supports multiple API-Interception implementations on several batch systems. The system consists of: (a) a factory component that produces lightweight Linux Containers using Docker, where each Container supports one or more API-Interception implementations and controls a single GPGPU; (b) a registry service that manages the Container resources; and, (c) a set of plugin scripts that bridge between the batch system and the registry. This paper also evaluates the performance of benchmarks on the prototype.