Performance predictability of divide and conquer skeletons

Abstract

Parallel divide and conquer computations, encompassing a wide variety of applications, can be modeled and encapsulated as a high level primitive called skeleton. The paper deals with a skeleton designed for parallel divide and conquer algorithms that provide hypercubical communications among processes The paper also introduces an accurate timing model designed for prediction of proposed primitive. The timing analysis model presented here still characterizing the communication time through architecture parameters but introduces a few novelties. The proposal is to introduce different kinds of components to the analytical model by associating a performance constant for each specific conceptual block of the skeleton. The trace files obtained from the execution of the resulting code using the skeleton are used by lineal regression techniques giving us, among other information, the values of the parameters of those blocks. An extended example showing the relative accuracy of the proposed approach concludes the paper.