Thread scheduling within paravirtualised domains

Abstract

The virtualisation of commodity hardware, notably that of systems based on the x86 architecture, is currently a source of great interest for both the scientific and commercial communities alike. Recent efforts have minimised the overheads of virtualisation through improved hardware support as well as through the adoption of virtualisation paradigms such as paravirtualisation. Research has also shown that microkernels, long thought to be impractical due to their supposedly inadequate performance, could indeed serve as viable operating system designs. These developments could enable the design of a secure operating system composed of a network of communicating microkernel components, each of which exists within its own virtual machine. This project is concerned with the creation of a single virtualised component, namely a thread scheduling framework that can support multiprocessor virtual machines. This will serve both as a platform for the evaluation of different scheduling policies as well as a possible foundation over which future operating system components could be built. The system can broadly be seen as having two aspects, namely the thread scheduling component and the supporting memory management functionality. The focus of this project is on the former aspect, with the latter implemented to support essential memory operations such as block allocation and freeing as well as fundamental dynamic data structures. The scheduler achieves a clean separation between mechanism and policy by keeping the thread switching and selection procedures separate. This allows the same framework to support the implementation of multiple scheduling policies and topologies, both symmetric and asymmetric. In turn, this enables different scheduling policies to be evaluated under consistent conditions. This project evaluates several scheduling policies, and attempts to quantify the effects of contention arising both within and amongst virtual machines.