Locality Enhancement and Dynamic Optimizations on Multicore and GPU

Colloq: Speaker: 
Dr. Xipeng Shen
Colloq: Speaker Institution: 
College of William and Mary
Colloq: Date and Time: 
Tue, 2011-09-27 10:00
Colloq: Location: 
5100, Room 128 JICS Lecture Hall
Colloq: Host: 
Dr. Jeffrey Vetter
Colloq: Host Email: 
vetter@ornl.gov
Colloq: Abstract: 
Recent years have seen two prominent changes in computing, namely, the advent of Chip Multiprocessors (CMP) and the rapid adoption of GPU for high performance computing. They bring program optimizations some novel opportunities, but also many new challenges---such as the complex cache sharing among computing units, the high sensitivity of performance to thread divergences and irregular memory references, and a substantially expanded parameter space for optimizations.This talk presents some recent progresses in addressing these complexities. It particularly focuses on advances in three aspects: a CPU-GPU pipelining scheme for streamlining GPU applications on the fly, the locality analysis and job co-scheduling on CPU with non-uniform cache sharing, and an input-centric paradigm for dynamic optimizations that benefit both CPU and GPU applications. These techniques show promises in significantly enhancing software-hardware matching, creating synergy among (heterogeneous) computing units, and stimulating further innovations for the maximization of efficiency in future computing
Colloq: Speaker Bio: 
Xipeng Shen is an Assistant Professor of Computer Science at the College of William and Mary. He received his Ph.D. and Master degrees in Computer Science from University of Rochester. He is an IBM CAS Faculty Fellow, a recipient of the CAREER Award from the National Science Foundation and the Early Career Award from the U.S. Department of Energy. His work on cache sharing on multicore received the best paper award from the ACM PPoPP 2010.