Analysis of Data Reuse in Task-Parallel Runtimes

Abstract. This paper proposes a methodology to study the data reuse quality of task-parallel runtimes. We introduce an coarse-grain version of the reuse distance method called Kernel Reuse Distance (KRD). The metric is a low-overhead alternative designed to analyze data reuse at the socket level while minimizing perturbation to the parallel schedule. Using the KRD metric we show that reuse depends considerably on the system configuration (sockets, cores) and on the runtime scheduler. Furthermore, we correlate KRD with hardware metrics such as cache misses and work time inflation. Overall we found that KRD can be used effectively to assess data reuse in parallel applications. The study also revealed that several current runtimes suffer from severe bottlenecks at scale which often dominate performance.