Minimizing expected energy consumption for streaming applications with linear dependencies on chip multiprocessors

Abstract

Dynamic voltage scaling (DVS) is a widely applied power management mechanism in real-time systems. We propose an algorithm for scheduling periodic hard real-time streaming applications with linear dependencies and known probability distributions of computational requirements on chip multiprocessors (CMP). The goal of the scheduling is to minimize the expected energy consumption while satisfying two quality of service (QoS) requirements: throughput and response time. Our experiments show significant energy savings (up to 55%) over scheduling when only the worst case computational requirements are known. In addition, while dynamically reclaiming processor idle time across multiple processors yields small benefit when scheduling is based on the probability distribution of computational requirements, it results in significant energy savings when scheduling for the worst case, especially for applications with short deadlines.