High-performance low-power task scheduling on Multiprocessor System-on-Chip.

摘要

Efficient scheduling is crucial to extract the maximum performance out of the Multiprocessor System-on-Chip (MPSoC). In embedded systems where power is critical, an intelligent task scheduling strategy can guide the power reduction technique while simultaneously satisfying the performance and Quality of Service (QoS) guarantees.;Task scheduling on MPSoC platform has been a well explored research area. However, due to constant technology scaling the challenges in this area are ever increasing. Problems such as temperature gradient and communication contention are not present on single processor systems. Moreover, existing problems like workload variation on uniprocessor systems may worsen on MPSoCs if not dealt tactfully. These problems can affect several system parameters including power, performance, reliability, cost etc.;In this dissertation we consider three distinct problems under the domain of MPSoC task scheduling. First, we consider workload variation problem caused by nondeterministic nature of control intensive applications with conditional branches. We present a condition-aware scheduling framework built upon a set of algorithms and show that it achieves better power savings compared to existing techniques while meeting performance deadlines. We also show that the framework can dynamically adapt to the fluctuation in workload at runtime while achieving greater power savings. Next, we propose a resource-aware scheduling algorithm to minimize resource contention and processing latency on an MPSoC platform. We show that application specific task partition and mapping techniques can guide the scheduling and further boost the performance. Last, we target the temperature variation problem on an MPSoC with large number of processing cores and propose a distributed thermal management policy to balance the thermal gradient and increase the performance. We show that our distributed, migration based and thermal-aware scheduling strategy outperforms existing centralized techniques in many aspects. Finally, we also show how the application specific information can be utilized to further guide power reduction technique. We analyze the MPEG video decoding application and propose a workload prediction model which accurately predicts the future workload and helps achieving better power savings.