Monte Carlo optimisation auto-tuning on a multi-GPU cluster

In this paper we investigate Monte Carlo optimisation of the fitness function on a multi-GPU cluster. Our main goal is to develop auto-tuning techniques for the GPU cluster. Monte Carlo or random sampling is a technique to optimise a fitness function by giving random values to function parameters. When execution of the fitness function requires a high amount of computational power Monte Carlo sampling becomes both very time and computational power consuming. A developer who is not familiar with the application, hardware, and the CUBA runtime cannot determine the optimal execution parameters. This makes GPU auto-tuning well suited to achieving better performance and reducing computing time.