Dynamic Workload Prediction for Soft Real-Time Applications

Energy consumption is often a hard constraint of embedded real-time systems. Modern processors provide techniques for dynamic voltage and frequency scaling to reduce energy consumption. However, while the processor possibly operates at a lower clock frequency, the running applications should still meet their deadlines and thus set some limits to the use of scaling techniques. In this paper, we propose AutoCorrelation Clustering as a technique to predict the workload of each single iteration of a periodic soft real-time application. Based on this prediction we adjust the processor performance such that all deadlines are met as exactly as possible. We compare our technique to the broadly implemented race-to-idle and identify situations where autocorrelation clustering can gain higher energy savings than race-to-idle. Additionally, autocorrelation clustering can help saving energy in multithreaded processors where race-to-idle can be applied only with a high overhead if at all. We evaluated our approach by simulating the execution of a MPEG decoder on the multithreaded CarCore processor model.