Room 1034 ETB
Donald S. Fussell
University of Texas at Austin
Abstract: Grain size is a fundamental aspect of a parallel program because it affects communication, synchronization, load balance, and idle time. Grain size selection can be difficult because it depends on characteristics of both the program and the hardware. We introduce a programming model and runtime system that simplifies grain size selection for a wide range of programs, particularly for dynamic irregular computations running on heterogeneous hardware. The key idea is a division of labor in which the programmer provides a method of reducing a computation’s granularity while the runtime system uses this dicing method to adaptively identify the best work distribution and grain size for the given computation and execution environment.
We evaluate our system by applying it to a series of increasingly irregular workloads, demonstrating that this automatic granularity control allows programmers to more easily write portable and efficient code for heterogeneous platforms that consist of CPUs and GPUs.
BIO: Donald S. Fussell is Trammell Crow Regents’ Professor in the Department of Computer Sciences, Director of the Laboratory for Graphics and Parallel Systems, a member of the Computer Engineering Research Center of the Electrical and Computer Engineering Department, and a member of the Institute for Computational Engineering and
Sciences at the University of Texas at Austin. He obtained a BA in Mathematics and Social Sciences at Dartmouth College in 1973, and an MS and PhD in Mathematical Sciences at The University of Texas at Dallas in 1977 and 1980, respectively. His research interests include Computer Graphics, Computer Games, Computer Architecture, and Computer Systems Design.
Host: Dr. Khatri