CPU scheduling for virtual desktop infrastructure

Abstract

Virtual desktop infrastructure has been shifting the way of provisioning individual desktops from dedicated workstation deployment to on-demand virtual machine distribution. Inheriting versatile functionalities from ever-evolving virtualization technology, virtual desktop infrastructure enables cost-effective desktop environments in organizations such as enterprise. Dominant characteristics of desktop workloads and user behaviors drive the infrastructure to highly consolidated environments where many virtual machines are co-located on a single hardware platform in order to make the best use of available resources. For the high consolidation ratio of virtual desktops, CPU is a crucial resource that should be carefully managed to meet the computing demands of desktop users, since it can be both time-shared and space-shared in a very fine-grained manner. Although desktop workloads have been increasingly diversifying and challenging to be characterized, CPU schedulers in virtualization layer have aimed mainly at fairly partitioning CPU bandwidth for performance isolation between virtual machines without properly handling diverse workloads.

This dissertation proposes CPU scheduling schemes that identify and satisfy the computing demands of consolidated desktops by intelligently handling various workloads ranging from interactive to parallel workloads. In order to transparently bridge the semantic gap between operating systems and virtualization layer, the invented schemes extract workload characteristics from limited information accessible by the virtualization layer. The well-established correlation between architectural events and workload demands guides the CPU scheduler to effective decisions for enhancing the desired performance of desktop workloads. As opposed to dedicated environments, the scheduling decisions not only improve the performance of individual desktops but also preserve fair CPU sharing between them considering the multi-tenant nature of conso...