Seminars & Events

Mathematics and Computer Science Division
"A Quantitative Performance Model for the GPU"

DATE: September 1, 2011
TIME: 10:30 AM - 11:30 AM
SPEAKER: Yao Zhang, LANS Postdoc Interviewee
LOCATION: Building 240 Seminar Room 4301, Argonne National Laboratory
HOST: Boyana Norris

Description:
With the advent of C-based programming environments like CUDA and OpenCL, the recent years saw a great deal of interest in developing high-performance general-purpose applications for GPUs. However, today’s performance tuning practice by GPU programmers still demands manual measurement and paper-and-pencil analysis. In this talk, we will present our work on GPU performance modeling. The model we developed is able to help GPU architects and programmers to identify performance bottlenecks, suggest them solutions, and quantitatively predict/evaluate the effectiveness of the proposed solutions. By dynamic program simulation based on the native GPU instruction set, our model is able to handle data-dependent applications and predict performance with a 5–15% error. We will also demonstrate our performance tool for a variety of numerical algorithms including dense/sparse matrix multiply and tridiagonal system solvers. Finally,we will discuss our future work on building a performance modeling infrastructure, and the opportunities and challenges in model-based performance autotuning.

Save the event to your calendar [schedule.ics]