“No-Compile-Time Knowledge” distribution of finite element computations on multiprocessors

This paper addresses partitioning and scheduling of irregular loops arising in finite element computations on unstructured meshes. Target computers are distributed memory parallel computers that provide a global address space. We introduce the concept of “conditioned iterations loop” which distributes the iterations dynamically according to a runtime condition. This technique is improved by a learning approach. It is also combined with a renumbering of the mesh to improve data locality. Execution times on KSR1 with up to 64 processors on a model loop show good performances