Extending the Problem Data Size for GPU Simulation Beyond the GPU Memory Storage with LRnLA Algorithms

Abstract

To use the CPU RAM to store the data of a GPU simulation, the data exchange between CPU and GPU is required. With LRnLA algorithms, the computation without data exchange is made longer, and the data exchange may be concealed. In this work, the concept is demonstrated with the Lattice Boltzmann method simulation with ConeTorre LRnLA algorithm on GPU. A new method of data storage is proposed. The data on the synchronization planes are stored in tiles, and on the wavefront slopes a new data structure is used, which is optimized for cell data exchange between LRnLA sub-tasks. The performance tests show less than 5% overhead for CPU-GPU communication, and the GPU performance persists for simulations where the main storage site is the CPU RAM.