A dynamic approach for workload partitioning on GPU architectures

Workload partitioning and the subsequent work item-to-thread mapping are key aspects to face when implementing any efficient GPU application. Different techniques have been proposed to deal with such issues, ranging from the computationally simplest static to the most complex dynamic ones. Each of them finds the best use depending on the workload characteristics (static for more regular workloads, dynamic for irregular workloads). Nevertheless, no one of them provides a sound tradeoff when applied in both cases. Static approaches lead to load unbalancing with irregular problems, while the computational overhead introduced by the dynamic or semi-dynamic approaches often worsens the overall application performance when run on regular problems. This article presents an efficient dynamic technique for workload partitioning and work item-to-thread mapping whose complexity is significantly reduced with respect to the other dynamic approaches in literature. The article shows how the partitioning and mapping algorithm has been implemented by fully taking advantage of the GPU device characteristics with the aim of minimizing the involved computational overhead. The article shows, compares, and analyses the experimental results obtained by applying the proposed approach and several static, dynamic, and semi-dynamic techniques at the state of the art to different benchmarks and over different GPU technologies (i.e., NVIDIA Fermi, Kepler, and Maxwell) to understand when and how each technique best applies.