Compiling Geometric Paradigms through Local Communications

This paper presents a compiling technique to generate parallel code with explicit local communications for a mesh-connected distributed memory, MIMD architecture. Our compiling technique works for the geometric paradigm of parallel computation, i.e. a data-parallel paradigm where array data structures are partitioned and assigned to a set of processing nodes, which, to perform their identical tasks, need to exchange some of the data allocated to them. This means that some data dependencies exist between computations mapped onto diierent nodes. Our technique solves them statically, by inserting explicit communications in the parallel code running on each node. We also demonstrate that the technique produces an optimal code as far as the exploitation of the communication media bandwidth is concerned.