Parallel FEM Adaptation on Hierarchical Architectures

The parallel FEM package NuscaS allows us to solve adaptive FEM problems with 3D unstructured meshes on distributed-memory parallel computers such as PC-clusters. In our previous works, a new method for parallelizing the FEM adaptation was presented, based on using the 8-tetrahedra longest-edge partition. This method relies on a decentralized approach, and is more scalable in comparison to previous implementations requiring a centralized synchronizing node. At present nodes of clusters contain more and more processing cores. Their efficient utilization is crucial for providing high performance of numerical codes. In this paper, different schemes of mapping the mesh adaptation algorithm on such hierchical architectures are presented and compared. These schemes use either the pure message-passing model, or the hybrid approach which combines shared-memory and message-passing models. Also, we investigate an approach for adapting the pure MPI model to hierarchical topology of clusters with multi-core nodes.