An Improved Abstract GPU Model with Data Transfer

摘要

GPUs are commonly used as coprocessors to accelerate a compute-intensive task, thanks to their massively parallel architecture. There is study into different abstract parallel models, which allow researchers to design and analyse parallel algorithms. However, most work on analysing GPU algorithms has been software based tools for profiling a GPU algorithm. Recently, some abstract GPU models have been proposed, yet they do not capture all elements of a GPU. In particular, they miss the data transfer between CPU and GPU, which in practice can cause a bottleneck and reduce performance dramatically. We propose a comprehensive model called Abstract Transferring GPU which to our knowledge is the first abstract GPU model to capture data transfer between CPU and GPU. We show via experiments, that existing abstract GPU models cannot sufficiently capture all of the actual running of a GPU algorithm time in all cases, as they do not capture data transfer. We show that by capturing data transfer with our model, we are able to obtain more accurate predictions of the GPU algorithm actual running time. It is expected that our model helps improve design and analysis of heterogeneous systems consisting of CPU and GPU, and will allow researchers to make better informed implementation decisions, as they will be aware how data transfer will affect their programs.