Presentation
Improve and Stabilize Classification Results of DataRaceBench
DescriptionDataRaceBench is a benchmark using small kernel applications to classify the detection capabilities of data race detection tools. During our experiments of applying Archer to the benchmark suite we observed different short-comings. With recently added kernels, the turn-around time of a basic benchmark run increased from several minutes to more than an hour. Furthermore, we observed non-deterministic and unexpected results. In this presentation, we propose several changes to existing kernels to address these short-comings. In addition, we propose to use variants of the kernels with non-deterministic runtime schedules that explicitly enforce these different schedules. Finally, we provide an evaluation of the updated benchmark with Archer running in thread-centric and task-centric mode.
