Communication primitives for unstructured finite element simulations on data parallel architectures

E cient data motion is critical for high performance computing on distributed memory architectures. The value of some techniques for e cient data motion is illustrated by identifying generic communication primitives. Further, the e ciency of these primitives is demonstrated on three di erent applications using the nite element method for unstructured grids and sparse solvers with di erent communication requirements. For the applications presented, the techniques advocated reduced the communication times by a factor of between 1.5 { 3.