Task-Based Parallelization of Unstructured Meshes Assembly Using D&C Strategy

This paper introduces our work on parallelizing unstructured mesh applications in shared memory. Our approach uses task-based runtime and Divide and Conquer recursive bissection strategy for the work decomposition. We demonstrate this parallelization strategy on the unstructured mesh assembly step of an industrial CFD application. Preliminary results show a speedup of 1.8x on a 12 cores node and a performance on Xeon Phi equivalent to 15 Ivy Bridge cores, with a parallel efficiency higher than 90% on the physical cores.