earticle

영어

It is still difficult and cumbersome to locate where a program runs into data races, because they may lead to unpredictable results of the program. To detect data races occurred during an execution of parallel programs, previous work provides large runtime and space overhead or focuses on reducing false positives. Thus, the prior detectors are still imprecise and inefficient, when applied to large scale parallel programs. This paper presents an efficient data race detection technique, called EDP that analyzes conflicting accesses to every shared memory location. Our technique precisely reports data races because it guarantees to locate at least one data race for each shared memory location, if there exists any, and provides a significant improvement of efficiency as O(1) space and time overheads for each access history.