Data Locality for Massively Parallel Architectures
Recent years have seen a rise of massive parallelism in modern processors, exemplified by Graphic Processing Units (GPU), Many Integrated Cores (MIC), and so on. As parallelism continues increasing fast, memory bandwidth expansion lags behind, leaving effectively bringing data to cores a critical challenge for computing efficiency and performance scalability.
This talk discusses the important role of data locality enhancement in meeting the needs of future computing. It examines the implications massive parallelism brings to data locality, and some recent progress in the measurement, modeling, and exploitation of data locality. It concludes with a list of open questions and research directions.
Xipeng Shen is the Adina Allen Term Distinguished Associate Professor in the College of William and Mary, and an IBM Center for Advanced Studies (CAS) Faculty Fellow. His research in data locality and massive parallelism won the DOE Early Career Research Award in 2011 and the Best Paper Award at ACM PPoPP 2010. His research in input-centric program dynamic optimizations won the NSF CAREER Award in 2010.
He was a Visiting Researcher at MIT, Microsoft Research, and Intel from 2012 to 2013. Xipeng Shen's research lies in the broad field of programming systems, with an emphasis on enabling extreme-scale data-intensive computing and intelligent portable computing through innovations in both compilers and runtime systems. He has been particularly interested in capturing large-scale program behavior patterns, in both data accesses and code executions, and exploiting them for scalable and efficient computing in a heterogeneous, massively parallel environment.