high performance computing on graphics processing units: hgpu.org

Stochastic simulations involve multiple replications in order to build confidence intervals for their results, and Designs Of Experiments (DOEs) to explore their parameters set. In this paper, we propose Warp-Level Parallelism (WLP), a GPU-enabled solution to compute Multiple Replications In Parallel (MRIP) on GPUs (Graphics Processing Units). GPUs are intrinsically tuned to process efficiently the same operation on several data, which is not suited to parallelize MRIP or DOEs. Our approach proposes to rely on small thread groups, called warps, to perform independent computations such as replications. This approach has proved to be efficient on three classical simulation models, but originally lacked the transparency users might expect. In this work, we enhance WLP using Aspect Oriented Programming (AOP). Our work describes the way to combine CUDA and AOP, and brings forward the techniques available to exploit AOP in a CUDA-enabled development.