2014 Volume 27 Issue 1 Pages 81-93
ABSTRACT In each SIMD (Single Instruction, Multiple Data) group, called a `warp of a GPU (Graphics Processing Unit), all the ?xed number of threads execute the same instruction concurrently at each unit period of time. We consider a class of probabilistic algorithms designed for use on GPUs, including a wide variety of Monte Carlo methods,such that each thread contains a loop iterated stochastically variable times, and that the life-cycle of a warp ends when the slowest thread completes its requested task.A run-time model is proposed in order to explain the distributions of execution time observed in SIMD parallel computations using the algorithms of this class. Asymptotic properties of those distributions are also presented.