The International Conference for High Performance Computing, Networking, Storage and Analysis
Gkleepp: Parallelizing a Symbolic GPU Race Checker.
Student: Mark S. Baranowski (University of Utah)
Supervisor: Ganesh L. Gopalakrishnan (University of Utah)
Abstract: The increased usage of GPGPU programs necessitates tools to detect errors in the program. Existing tools which use hardware instrumentation may miss certain bugs due to limitations in the memory they observe and the requirement that tests must produce the but. Formal methods found in tools such as Gkleep can find data races without users providing test cases. The primary disadvantage of using such tools is the slowness of execution when compared to the speed at which conventional tools run. We present a parallelized version of Gkleep, christened Gkleepp, which better uses computation resources through parallel execution. We can attain speedups of 2-8 times while retaining most detection of data races.