Performance analysis for stencil-based 3D MPDATA algorithm on GPU architecture

Abstract

EULAG (Eulerian/semi-Lagrangian fluid solver) is an established computational model for simulating thermo-fluid flows across a wide range of scales and physical scenarios. The multidimensional positive defined advection transport algorithm (MPDATA) is among the most time-consuming components of EULAG. The main aim of our work is to design an efficient adaptation of the MPDATA algorithm to the NVIDIA GPU Kepler architecture. We focus on analysis of resources usage in the GPU platform and its influence on performance results. In this paper, a performance model is proposed, which ensures a comprehensive analysis of the resource consumption including registers, shared, global and texture memories. The performance model allows us to identify bottlenecks of the algorithm, and shows directions of optimizations. The group of the most common bottlenecks is considered in this work. They include data transfers between host memory and GPU global memory, GPU global memory and shared memory, as well as latencies and serialization of instructions, and GPU occupancy. We put the emphasis on providing a fixed memory access pattern, padding, reducing divergent branches and instructions latencies, as well as organizing computation in the MPDATA algorithm in order to provide efficient shared memory and register file reusing. © 2014 Springer-Verlag.