Performance-aware task scheduling in multi-core computers

Multi-core systems become more and more popular as they can satisfy the increasing computation capacity requirements of complex applications. Task scheduling strategy plays a key role in this vision and ensures that the task processing is both Quality-of-Service (QoS, in this thesis, refers to deadline) satisfied and energy-efficient. In this thesis, we develop task scheduling strategies for multi-core computing systems. We start by looking at two objectives of a multi-core computing system. The first objective aims at ensuring all tasks can satisfy their time constraints (i.e. deadline), while the second strives to minimize the overall energy consumption of the platform. We develop three power-aware scheduling strategies in virtualized systems managed by Xen. Comparing with the original scheduling strategy in Xen, these scheduling algorithms are able to reduce energy consumption without reducing the performance for the jobs. Then, we find that modelling the makespan of a task (before execution) accurately is very important for making scheduling decisions. Our studies show that the discrepancy between the assumption of (commonly used) sequential execution and the reality of time sharing execution may lead to inaccurate calculation of the task makespan. Thus, we investigate the impact of the time sharing execution on the task makespan, and propose the method to model and determine the makespan with the time-sharing execution. Thereafter, we extend our work to a more complex scenario: scheduling DAG applications for time sharing systems. Based on our time-sharing makespan model, we further develop the scheduling strategies for DAG jobs in time-sharing execution, which achieves more effective at task execution. Finally, as the resource interference also makes a big difference to the performance of co-running tasks in multi-core computers (which may further influence the scheduling decision making), we investigate the influential factors that impact on the performance when the tasks ii are co-running on a multicore computer and propose the machine learning-based prediction frameworks to predict the performance of the co-running tasks. The experimental results show that the techniques proposed in this thesis is effective