Katherine Yelick

Katherine Yelick, who will assume the post of NERSC Center Division Director in January 2008, is Professor of Electrical Engineering and Computer Sciences at the University of California at Berkeley and a Faculty Research Scientist at Lawrence Berkeley National Laboratory. She received her bachelor’s, master’s, and Ph.D. degrees from the Massachusetts Institute of Technology. Yelick joined the UC Berkeley faculty in 1991, and has been with Computing Sciences at LBNL since 1996, most recently as Future Technologies Group Lead.

Yelick has devoted her career to making parallel machines easier to use through the use of libraries, languages, compilers, and other software tools. Her parallel language and compiler projects include Split-C and UPC, which are parallel extensions of C, and Titanium, a high performance scientific computing language based on Java. She is co-author of the main textbook on UPC and leads the Berkeley UPC compiler effort, an open source compiler that is widely used in applications and productivity experiments and is shipped as the standard UPC compiler for some systems from some vendors.

In 2006, Yelick was named one of 16 “People to Watch in 2006” by the newsletter HPCwire. The editors noted that “Her multi-faceted research goal is to develop techniques for obtaining high performance on a wide range of computational platforms, all while easing the programming effort required to achieve high performance. Her current work has shown that global address space languages like UPC and Titanium offer serious opportunities in both productivity and performance, and that these languages can be ubiquitous on parallel machines without excessive investments in compiler technology.” All three of her languages give programmers the convenience of shared data structures on machines with or without hardware support for shared memory.

She has also worked on the DARPA High Productivity Language Systems (HPLS) project, and has done foundational work in combining and mapping parallel programming models including mixed task and data parallelism, distributed data structures, a provably optimal dynamic load balancing algorithm, and compiler support for speculative parallelism. She has extensive work on interdisciplinary applications projects including symbolic algorithms (the Knuth-Bendix completion procedure for theorem proving, a Gröbner basis algorithm for symbolic algebra, and BDD construction), biological computations (a simulation of blood flow in the heart and Phylogeny construction algorithm for genomics), circuit simulation (using speculation), accelerator modeling, adaptive mesh refinement, unstructured mesh generation, FFTs, dense linear algebra, and sparse linear algebra (factorization using an event-driven, multithreaded model).

She has also worked on the design and use of novel architectures, from the Berkeley IRAM project (a processor in memory vector system) to techniques for running scientific applications on architecture emerging from games, graphics and media processing. She led the Future Technologies Group at LBNL, which covers many aspects of the parallel computing problem: architecture, operating systems (K42), fault tolerance (BLCR), communication libraries (GASNet), compilers (UPC), performance modeling (APEX-Map), and application scaling. The LBNL team was instrumental in writing the “Berkeley View” report and will be close collaborators on the ParLab activity.

Yelick is a world expert in automatic tuning systems. She leads the automatic tuning effort in the Performance Engineering Research Institute (PERI) project; PERI is a DOE SciDAC project involving over 30 researchers from 10 institutions. She and her student developed the first automatic tuning library for sparse matrix computations, Sparsity. Prior to her work, search-based tuning was only use for dense matrices and FFTs, domains where the computation depends only in the size of the input problem, and not its data. Sparsity and the follow-on OSKI library optimize for the structure of the sparse matrix, adding explicit zeros and replacing data structures to improve performance. She is now working on extending these ideas to structured grid computations and collective communication.

Yelick speaks extensively on her research, with over 15 invited talks and keynotes over the past three years. She is involved with a National Research Council committee on the Future of Computing Performance and was a co-author on the WTEC report on High End Computing in Japan, commissioned by NSF, DOE, NITRD, and NASA. Her awards include best-paper awards, the George M. Sprowls Award for an outstanding Ph.D. dissertation at MIT, the ARO Young Investigator Award, the Okawa Foundation award, and teaching awards from the EECS Departments at both MIT and Berkeley. She regularly teaches a parallel computing course (CS267) to graduate students, over half of whom are from departments outside computer science. Yelick also recently developed an undergraduate parallel programming course to address the need for programmers able to effectively use multicore processors.