Performance Impact of Removing Data Races from GPU Graph Analytics Programs

Some of the fastest CUDA codes contain "benign" data races to boost their performance. However, such races can lead to unpredictable behavior and incorrect results on other hardware and compilers, making their elimination crucial for producing reliable and portable programs. This paper investigates the performance impact of removing data races from six high-end graph analytics codes. We identify and eliminate the races from these GPU programs by adding necessary synchronization and validating their correctness. We present our race-free codes and their original versions as an open-source suite. Comparing the performance of our new codes with their baseline counterparts on multiple inputs and GPUs, we observe that race-free implementations do not always incur a performance penalty. In fact, some race-free versions are faster, with our validated maximal independent set implementation achieving a 5-11% speedup. Our findings indicate that race-free code can reach comparable or even superior performance, supporting the adoption of best practices for parallel programming.