Earticle

Detecting data races is important in debugging shared memory parallel programs, because the races could exhibit unpredictable results in execution of programs. Unfortunately, previ- ous race detection techniques cannot guarantee that the detected accesses could be involved in the rst races to occur in parallel programs which contain random synchronizations. This paper presents a monitoring algorithm which collects ltered key accesses of local thread blocks that are involved in races comparing with the latest ltered accesses of the other concurrent thread blocks in a particular execution of parallel programs. We discard all the accesses other than key accesses in a block to perform the race condition determination, which makes our algorithm more ecient in time and space. We also demonstrate the cor- rectness of our algorithm by showing at least one of the collected candidate accesses would be included in the actual first races.