GPU Implementation of ConeTorre Algorithm for Fluid Dynamics Simulation

TL;DR: A fluid simulation code based on the Lattice-Boltzmann method with a performance that surpasses state-of-the-art solutions is developed.

Abstract: LRnLA algorithms allow simulation of large problems with performance that exceeds the memory-bound limit of the traditional stepwise algorithms, that is, algorithms without any kind of temporal blocking. We show how the ConeTorre LRnLA algorithm that was successfully implemented for various CPU codes may be ported to work with CUDA framework and implemented the Lattice-Boltzmann Method (LBM) for fluid dynamics. As the standard tools and guidelines do not comply with the LRnLA paradigm, we have performed manual optimization of the communication between main memory levels of GPU and reduce overhead for data access patterns. We have made the performance estimate of the LRnLA implementation with the use of the Roofline model. The computation remains memory-bound, but with the ConeTorre algorithm the operational intensity is increased several times, and the maximum achievable performance for the chosen algorithm parameters is 9 billion cell updates per second on Tesla V100. We have achieved more than 66% of the estimate. As a result, we have developed a fluid simulation code based on the Lattice-Boltzmann method with a performance that surpasses state-of-the-art solutions.