A Fast and Iterative Migration for GPU Applications

Abstract

Graphic Process Unit (GPU) is widely used for graphics computing and general-purpose computing. Therefore, the cloud environment starts to import GPU as a part of the infrastructure to provide new services, such as cloud gaming. Migration is a core feature in the management of cloud environment. However, it is not easy to migrate the Virtual Machine (VM) with the running GPU application using the current VM migration technique. To this end, we propose FIM, an open source 1 solution which can provide a fast and iterative migration for GPU applications. Our evaluation shows that 1) the downtime of the VM migration is 220-320 ms, 2) different GPU workloads degrade only 1.5-3.5% of performance, and 3) our solution only occupies 100-180 Mbits/Sec bandwidth during execution. 2015 18-19, December, ISCC