ATRG Academic Session

September 19th, 2012
Room 214, Faculty of Science Bldg.7,
the University of Tokyo

13:50 - 14:00
Introduction by Professor Suda

14:00 - 15:00
Surrogate-Assisted Software Auto-Tuning

Professor Weichung Wang
Department of Mathematics
National Taiwan University

Software auto-tuning is a board area with various applications. In auto-tuning, a set of parameters are determined to achieve particular goals such as performance optimization and load balance. On the other hand, many computer experiments can be formulated as auto-tuning problems. In this talk, we focus on surrogate-assisted approaches on parameters determinations for particular goals in different computer experiment problems. By using limited computational resources, surrogates are constructed to approximate the true response functions that are usually unknown and/or complicated. To construct the surrogates, existed statistical tools such as Kriging (Gaussian process) and expected improvement can be useful in many smooth response functions. We also develop the "overcomplete basis surrogate methodology" for complicated and oscillatory response functions.
Several examples, e.g. chaotic light sources, eigenvalue solver, and photonic crystal bandgap, will be presented to illustrate the efficiency of the surrogate-assisted approaches.

15:00 - 15:15 Break

15:15 - 16:15
New Directions in Auto-tuning Higher-level Motifs

Dr. Shoaib Kamil
Computer Science Engineer
Future Technologies Group
Computational Research Division
Lawrence Berkeley National Laboratory

Auto-tuning has been traditionally applied to numeric kernels that are relatively simple and solve a fixed problem that cannot be substantially changed by the programmer. In this talk, I present some insights for auto-tuning higher-level motifs such as graph algorithms and others that require higher-level functions. Furthermore, some results of co-tuning from auto-tuning kernels that use other auto-tuned kernels inside will be presented. The talk's purpose is to spur discussion on techniques for co-tuning that are more tractable than the current brute-force approaches.

16:15 - Closing by Professor Katagiri