PhD Thesis Final Defense to be held on July 17, 2019, at 12:00

The examination is open to anyone who wishes to attend.

Thesis Title: Efficient Accelerator Virtualization Techniques for Cloud Environments

Abstract: The objective of this thesis is to explore methods and develop frameworks for heterogeneous architectures aiming at integrating them into virtualization environments. In this context, we developed and evaluated techniques in order to support GPUs and Xeon Phi accelerators into the virtualization ecosystem. Following these methods, we primarily focused on i) executing accelerator-based applications on virtual machines, retaining the virtualization overhead at acceptable rates and ii) sharing of an accelerator between multiple virtual machines. Initially, we utilized a remote CUDA acceleration framework in a single-node virtualization platform and combined it with our optimized, low overhead intra-node framework which results in efficient application offloading in virtualized environments. Next, we followed the paravirtualization technique targeting the low-level transport layer of the Intel Xeon Phi, to enable the corresponding accelerator device to be shared by multiple virtual machines running on the same physical node.

Supervisor: Nectarios Koziris, Professor

PhD student: Stefanos Gerangelos