Scheduling strategies for real-time tasks in thermally constrained processing environments.

摘要

This dissertation is motivated by rapid and increasing rise in power density of modern processing platforms. High power densities cause thermal hotspots which degrade reliability, performance, and efficiency of a system. The research in this dissertation aims to alleviate these temperature related problems in the context of real-time computing (systems where tasks have deadlines). Traditional temperature management strategies, such as Dynamic Voltage and Frequency Scaling (DVFS), cause performance degradation which may cause deadline violations in real-time systems. Depending on the criticality of a real-time system, deadline violations and other failures can have catastrophic consequences. Therefore, it is important to devise thermal management schemes specific to real-time systems; which is the subject of this dissertation. Real-time systems have been an active area of research for the past four decades. However, majority of research has focused on meeting deadline constraints of real-time tasks. Strong theoretical results for thermal feasibility of real-time tasks have been absent and current solutions are restrictive in terms of power model and solution strategies. In this dissertation, we counter these limitations and provide strong theoretical results pertinent to thermal feasibility of real-time tasks.;We propose the concepts of Thermal Impact/Utilization which are used to prove several important theoretical results. Specifically, we prove that thermal utilization of less than or equal to 1 is a necessary and sufficient condition for thermal feasibility of periodic real-time tasks on uni-core systems. This is similar to the computational feasibility condition proposed by Liu and Layland in 1973. Apart from periodic task scheduling, we also propose optimal scheduling algorithms for aperiodic tasks. In line with current architectural trends, the concept of Thermal Impact/Utilization is generalized to a multi-core processing platform and thermally optimal scheduling strategies for multi-core platforms are proposed. DVFS strategies are also proposed which enable further temperature reduction. We also evaluate the proposed multi-core solution on a hardware test-bed. The theoretical results, simulations and hardware evaluations show that the proposed concepts can be used to significantly reduce system temperature. The application of the proposed concepts to other application domains also has significant potential.