Performance diagnosis for inefficient loops

Abstract

Writing efficient software is difficult. Design and implementation defects can cause severe performance degradation. Unfortunately, existing performance diagnosis techniques like profilers are still preliminary. They can locate code regions that consume resources, but not the ones that waste resources. In this paper, we first design a root-cause and fix-strategy taxonomy for inefficient loops, one of the most common performance problems in the field. We then design a static-dynamic hybrid analysis tool, LDoctor, to provide accurate performance diagnosis for loops. We further use sampling techniques to lower the run-time overhead without degrading the accuracy or latency of LDoctor diagnosis. Evaluation using real-world performance problems shows that LDoctor can provide better coverage and accuracy than existing techniques, with low overhead.