Array-Based Reduction Operations for a Parallel Adaptive FEM

For many applications of scientific computing, reduction operations may cause a performance bottleneck. In this article, the performance of different coarseand fine-grained methods for implementing the reduction is investigated. Fine-grained reductions using atomic operations or fine-grained explicit locks are compared to the coarse-grained reduction operations supplied by OpenMP and MPI. The reduction operations investigated are used for an adaptive FEM. The performance results show that applications can gain a speedup by using fine-grained reduction since this implementation enables to hide the reduction between calculation while minimising the time waiting for synchronisation.