Scheduling Parallel Tasks on Multicore Nodes for Energy Saving
Jason Mair, Kai‐Cheung Leung, Zhengfeng Huang · Parallel and Distributed Computing: Applications and Technologies · 2010
In this paper, we have further explored the novel metrics and policies of Speedup per Watt (SPW), Power per Speedup (PPS), Energy per Target (EPT), Sharing Policy, the Hare and the Tortoise Policies, which were introduced in our previous work. Each policy leverages application parallelism and Dynamic Voltage and Frequency Scaling (DVFS) to reduce energy consumption in multicore computers. Our experiments show that, running the Sharing Policy on a busy multicore computer can make savings of 25% in energy and reduce execution time by 26%. The Hare Policy made savings up to 73% on a system with low utilization. Combining these two methods with task migration on a small four-node cluster made savings of 58%.