Strategies to minimise the total run time of cyclic graph based genetic programming with GPUs

In this paper, we describe our work to investigate how much cyclic graph based Genetic Programming (GP) can be accelerated on one machine using currently available mid-range Graphics Processing Units (GPUs). Cyclic graphs pose different problems for evaluation than do trees and we describe how our CUDA based, "population parallel" evaluator tackles these problems. Previous similar work has focused on the evaluation alone. Unfortunately large reductions in the evaluation time do not necessarily translate to similar reductions in the total run time because the time spent on other tasks becomes more significant. We show that this problem can be tackled by having the GPU execute in parallel with the Central Processing Unit (CPU) and with memory transfers. We also demonstrate that it is possible to use a second graphics card to further improve the acceleration of one machine. These additional techniques are able to reduce the total run time of the GPU system by up to 2.83 times. The combined architecture completes a full cyclic GP run 434.61 times faster than the single-core CPU equivalent. This involves evaluating at an average rate of 3.85 billion GP operations per second over the course of the whole run