Partitioning the global space for distributed memory systems

Abstract

Partitioning the iteration space can signijcandy affect lhe execution time of a loop. In this paper, we propose an improvement over previous partitioning methods for single loops with unform data dependencies. For distributed memory systems, partitioning each loop separately does not guarantee an efjcient execution of the code because of across loop data dependence. As a result, a global iteration space is formed so that all loops in a program are considered when partitioning the global space.In addition, a new and general form of expressing data dependence called hyperplane dependence is introduced and used in the partitioning. It is a dependence whose source and destination are subspaces (of any dimension) of the global iteration space.

Automatic Data Distribution for Composite Grid Applications

“…Since the description of the grid assembly is normally part of the input, user-supplied data distribution would imply user intervention for each new grid assembly input. Other automatic distribution systems consider only a single mesh or general graphs., thereby loosing regularity for compilation [6][7][8][9][10][11][12]. The algorithms presented here explicitly address composite grid applications and preserve regularity to support compile-time optimization of these applications.…”