EEWA: Energy-Efficient Workload-Aware Task Scheduling in Multi-core Architectures

Abstract

Modern multi-core architectures offer Dynamic Voltage and Frequency Scaling (DVFS) that can dynamically adjust the operating frequency of each core for energy saving. However, current parallel programming environments and schedulers for task-based programs do not utilize DVFS and thus suffer from energy inefficiency in multi-core processors. To reduce energy consumption while keeping high performance, this paper proposes an Energy-Efficient Workload-Aware (EEWA) task scheduler that is comprised of a workload-aware frequency adjuster and a preference-based task-stealing scheduler. Using DVFS, the workload-aware frequency adjuster can properly tune the frequencies of the cores according to the workload information of the tasks collected with online profiling. The preference-based task-stealing scheduler can then effectively balance the workloads among cores by stealing tasks according to a preference list. Experimental results show that EEWA can reduce energy consumption of task-based programs up to 29.8% with a slight performance degradation compared with existing task schedulers.