Slack-Time-Aware Energy Efficient Scheduling for Multiprocessor SoCs

Variable task execution time causes more complexity in task scheduling and thermal management. In this paper, we introduce a discrete probability distribution model to capture the variation of task execution time. And then we propose a slack-time-aware two-phase scheduling framework for parallel applications running on heterogeneous MPSoCs. It exploits slack time among tasks with precedence constraints and scale up or down processor frequency dynamically to control processor power and temperature. In the offline phase, considering expected execution times of tasks and a total deadline, an integer linear programming (ILP) method is designed to create an allocation solution for all tasks to minimize expected energy consumption. In the online phase, the scheduler dynamically scales frequency or migrates tasks to reduce hotspots of all processors. In particular, the slack-time-aware framework takes into account static and dynamic slack time to schedule tasks such that the energy consumption is minimized. Experimental results show that our proposed algorithm decreases energy consumption by 5.2% in comparison with the pessimistic solution (using the worst case execution time), while maintaining a low deadline miss ratio.