Performance Analysis of the Cilk Locality Runtime

Many parallel programming platforms, such as Cilk Plus, implement a work-stealing scheduler to load-balance parallel computations, since it provides a provably good execution time bound and performs well in practice. However, most work-stealing schedulers, including the Cilk Plus scheduler, do not utilize shared caches and memory efficiently. The Cilk Locality runtime was developed to address the poor locality in the Cilk Plus runtime with the aim of improving overall performance. In this paper, the performance of the Cilk Locality runtime is compared to the Cilk Plus runtime using a number of different metrics. The results show that the Cilk Locality runtime achieves a modest improvement over the Cilk Plus runtime due to improved locality