A large amount of resources is spent writing, porting, and optimizing scientific and industrial High Performance Computing applications, which makes autotuning techniques fundamental to lower the cost of leveraging the improvements on execution time and power consumption provided by the latest software and hardware platforms. Despite the need for economy, most autotuning techniques still require large budgets of costly experimental measurements to provide good results, while rarely providing exploitable knowledge after optimization. This talk discusses a user-transparent autotuning technique based on Experimental Design, that operates under tight budget constraints by significantly reducing the measurements needed to find good optimizations. Our approach enables users to make informed decisions on which optimizations to pursue and when to stop. An experimental evaluation of our approach will be presented, first on the configuration of a GPU Laplacian kernel, and then on kernels from the SPAPT benchmark suite. We will also discuss our ongoing work to apply low-discrepancy sampling and Gaussian Process Regression to search spaces where global structure is hard to detect and exploit.
Pedro Bruel is a Ph.D. student on High Performance Computing at the University of Grenoble Alpes, advised by Arnaud Legrand and Alfredo Goldman. He got his B.Sc. in Computer Science in 2015 from the University of São Paulo, Brazil. Pedro works on autotuning, applying the experimental design methodology to High Performance Computing problems, and is currently on a two-month internship program at Hewlett Packard Enterprise Labs.
This seminar will be streamed. https://anlpress.cels.anl.gov/cels-seminars/