Using a master and slave approach for GPGPU computing to achieve optimal scaling in a 3D real-time simulation

Abstract

With the ever increasing computational demand of scientific research and data analysis, there has been a migration towards GPU computing. GPU are now the primary source of compute power in most top supercomputers. But in order to make use of the power programs must utilize more than a single GPU. Within this paper we will explain various approaches we have taken to utilize multiple GPU, and attempt to reach close to perfect scaling on a multi-step simulation. The result of this is having developed our simulation to be computed on a master and slave setup of GPU. Our simulation mentioned is being developed for the purpose of simulating microtubule dynamics on a gliding assay.