SC21 Proceedings

The International Conference for High Performance Computing, Networking, Storage, and Analysis

Core-Idling on MPI Intra-Node Communication Channels for Energy Efficiency

Authors: Keon-Woo Kim and Hyun-Wook Jin (Konkuk University, South Korea) and Eun-Kyu Byun (Korea Institute of Science and Technology Information (KISTI))

Abstract: Busy-waiting used to implement parallel programming models can provide low latency but at the expense of energy. There have been several studies to enhance the energy efficiency of MPI libraries and applications by lowering the CPU speed during busy-waiting or turning off CPU components. While most studies have focused on prediction-based schemes that determine to enter an energy-saving mode, a redesign of the MPI library for better supports for suspending and resuming processes over multiple communication channels has been paid less attention. In this study, we suggest a signaling-based core-idling approach for energy efficient MPI over multiple communication channels. Our preliminary implementation supports intra-node communication channels, such as shared memory and memory mapping. The performance measurement results show that our preliminary implementation can reduce the energy consumption of NPB up to 12% in the undersubscribed case and 91% in the oversubscribed case on a NUMA-based system.

Best Poster Finalist (BP): no

Poster: PDF
Poster summary: PDF

Back to Poster Archive Listing