Low-cost performance prediction of data-dependent data parallel programs

Current analytic solutions to the execution time distribution of an N-ary parallel composition of tasks having independent and identically distributed execution times are computationally complex, except for a limited number of distributions. In this paper we introduce an analytical solution based on approximating the execution time distributions in terms of a limited number of statistical moments. This approach allows the parallel execution time to be approximated with O(1) solution complexity for a wide range of execution time distributions, while the approximation accuracy outperforms comparable techniques known to date. Experiments show that the error of the predicted mean value of the parallel execution time is even less than 4% for parallel loops comprising up to 10,000 tasks whose execution times are normally distributed. Measurements on real programs (NAS-EP benchmark, PSRS sorter, and WATOR simulator) confirm these results provided the task execution distributions are independent and unimodal.