A Statistics-Based Approach to Fast Performance Evaluation

Although iterative optimization is an effective approach to achieve portable high performance on modern architectures, it takes a lot of time on evaluating the performances of various code versions. This paper proposes a statistics-based approach to accelerate this evaluation process, by testing multiple versions within each execution, so that more points within the target optimization space can be tested in a fixed amount of time. Experimental results are presented to demonstrate the feasibility of this statistics-based approach. It shows that only a few (four to six) samples are needed to accurately evaluate the performance of a given point.