Reducing Data Motion of Lattice Boltzmann Simulations through Application of Boundary Conditions on GPUs

Student: Griffin A. Dube (Clemson University, Dept of Electrical and Computer Engineering)
Supervisor: John P. Gounley (Oak Ridge National Laboratory)

Abstract: Lattice Boltzmann simulations are commonly used for solving many computational fluid dynamics problems. The main computation of many of these simulations are separated into two parts: collision and streaming of the lattice Boltzmann method particle distribution function, and application of boundary conditions. Most lattice Boltzmann performance studies on GPUs focus on the core collision and streaming kernels. In this study, we investigate how porting lattice Boltzmann Method boundary conditions to GPUs influences CPU-GPU data motion and overall application performance by porting the boundary conditions code of a lattice Boltzmann proxy application to CUDA. We complete all parts of the simulation on GPUs, reducing the number of data transfers required and improving time-to-solution by up to 49× by reducing the amount of time spent communicating data between CPU and GPU.

ACM-SRC Semi-Finalist: no

Poster: PDF
Poster Summary: PDF


Back to Poster Archive Listing