A Locality and Memory Congestion-Aware Thread Mapping Method for Modern NUMA Systems

Authors: Mulya Agung (Tohoku University), Muhammad Alfian Amrizal (Tohoku University), Ryusuke Egawa (Tohoku University), Hiroyuki Takizawa (Tohoku University)

Abstract: On modern NUMA systems, the memory congestion problem could degrade performance more than the memory access locality problem because a large number of processor cores in the systems can cause heavy congestion on memory controllers. In this work, we propose a thread mapping method that considers the spatio-temporal communication behavior of multi-threaded applications to improve the locality and to reduce the memory congestion on modern NUMA systems. We evaluate the proposed method using NPB applications on a NUMA system. Experiments show that our proposed method can achieve up to 20% performance improvement compared with locality-based and balance-based methods.

Best Poster Finalist (BP): no

Poster: pdf
Poster summary: PDF


Back to Poster Archive Listing