Implementing the lattice Boltzmann model on commodity graphics hardware

Abstract

Modern graphics processing units (GPUs) can perform general-purpose computations inaddition to the native specialized graphics operations. Due to the highly parallel nature ofgraphics processing, the GPU has evolved into a many-core coprocessor that supports highdata parallelism. Its performance has been growing at a rate of squared Moore’s law, andits peak floating point performance exceeds that of the CPU by an order of magnitude.Therefore, it is a viable platform for time-sensitive and computationally intensiveapplications.The lattice Boltzmann model (LBM) computations are carried out via linear operations atdiscrete lattice sites, which can be implemented efficiently using a GPU-basedarchitecture. Our simulations produce results comparable to the CPU versionwhile improving performance by an order of magnitude. We have demonstratedthat the GPU is well suited for interactive simulations in many applications,including simulating fire, smoke, lightweight objects in wind, jellyfish swimmingin water, and heat shimmering and mirage (using the hybrid thermal LBM).We further advocate the use of a GPU cluster for large scale LBM simulations and for highperformance computing. The Stony Brook Visual Computing Cluster has been the platformfor several applications, including simulations of real-time plume dispersion in complexurban environments and thermal fluid dynamics in a pressurized water reactor.Major GPU vendors have been targeting the high performance computing market withGPU hardware implementations. Software toolkits such as NVIDIA CUDA provide aconvenient development platform that abstracts the GPU and allows access to itsunderlying stream computing architecture. However, software programming for a GPUcluster remains a challenging task. We have therefore developed the Zippy framework tosimplify GPU cluster programming. Zippy is based on global arrays combined with thestream programming model and it hides the low-level details of the underlying clusterarchitecture.