Optimizing Sparse Matrix Assembly in Finite Element Solvers with One-Sided Communication

In parallel finite element solvers, sparse matrix assembly is often a bottleneck. Implemented using message passing, latency from message matching starts to limit performance as the number of cores increases. We here address this issue by using our own stack based representation of the sparse matrix, and a hybrid parallel programming model combining traditional message passing with one-sided communication. This gives an insertion rate up to more than twice as fast compared to state of the art implementations on a Cray XE6.