MOPED: Orchestrating Interprocess Message Data on CMPs

Speaker: Dr. Junli Gu (AMD)

Time and Date: 10:15-11:00, Apr. 17, 2013

Place: Room 369, Microelectronics Building, Zhangjiang Campus

Abstract

Future CMPs will combine many simple cores with deep cache hierarchies. With more cores, cache resources per core are fewer, and must be shared carefully to avoid poor utilization due to conflicts and pollution. In this talk, we describe a Message Orchestration and Performance Enhancement Device (MOPED) that provides hardware mechanisms to transfer message data efficiently in a timely manner. MOPED thus allows full overlap between communication and computation on the cores. Relative to unmodified MPI execution, MOPED reduces execution time of NAS Parallel Benchmarks by 17-45% and reduce off-chip memory misses are reduced by 43-88% .

Biography

Junli Gu got her PhD degree in Computer Engineering from Tsinghua University while her thesis was done as a visiting scholar in University of Illinois at Urbana-Champaign. Her thesis titles “Design Methodology of a High-Efficiency Cache Communication Mechanism for Chip Multi-Processors”. Junli later joined AMD research China lab. Collaborated with researchers from the US, she conducts research on CPU+GPU based heterogeneous architecture and AMD’s Heterogeneous System Architecture(HSA) . Junli’s research interests include multi-core computer architecture, cache hierarchy and target to increase the efficiency of the data sharing and data movement between cores.