Accommodating polymorphic data decompositions in explicitly parallel programs

Explicitly parallel programs have the potential for greater performance than their implicitly parallel counterparts. However, this benefit can be accompanied by additional programming difficulties. We address one particular problem that has implications for both scalability and portability: the need for programs do accommodate diverse data decompositions. We explain why programs with explicit communication have difficulties in handling changes in data decomposition, and we present a solution to this problem which involves the notions of derivative functions and configuration parameters. We illustrate the technique by using three different data decompositions to solve the Modified Gram-Schmidt method on four parallel machines.<
>