Predicting the Running Times of Parallel Programs by Simulation

Predicting the running time of a parallel program is useful for determining the optimal values for the parameters of the implementation and the optimal mapping of data on processors. However, deriving an explicit formula for the running time of a certain parallel program is a difficult task. We present a new method for the analysis of parallel programs: simulating the execution of parallel programs by following their control flow and by determining, for each processor, the sequence of send and receive operations according to the LogGP model. We developed two algorithms to simulate the LogGP communication between processors and we tested them on the blocked parallel version of the Gaussian Elimination algorithm on the Meiko CS-2 parallel machine. Our implementation showed that the LogGP simulation is able to detect the nonlinear behavior of the program running times, to indicate the differences in running times for different data layouts and to find the local optimal value of the block size with acceptable precision.