high performance computing on graphics processing units: hgpu.org

This work is an attempt to address the problem of bandwidth limited performance of data intensive GPGPU applications. Performance limited by memory bandwidth is common issue faced by general data intensive HPC applications. In case of the GPU, this problem is more pronounced owing to the unique architecture. This problem has been tackled by optimizing basic data rearrangement operations on the GPU. In this direction, methods and approaches have been identified and formulated for optimizing data rearrangement in general on GPU architectures. These are employed to develop near optimal and generic GPU kernels for a set of data rearrangement operations. In particular a library of GPU kernels has been developed, for operations that involve rearranging a generic m-dimensional data into n-dimensions. These kernels have been hand-tuned for maximum throughput, equaling upto 90% of the bandwidth utilization of the intrinsic memcpy function. These kernels are developed as templatized generic implementations allowing for their seamless integration into other existing applications. The target GPU architectures considered in this work are – NVIDIA Tesla c1060 and NVIDIA Tesla c870 and kernels have been developed using NVIDIA CUDA. All the kernels achieve or surpass best known performance in terms of bandwidth utilization. Furthermore, as a case study of a simple CFD Navier-Stokes based flow solver has been developed for the GPU, incorporating the optimal data rearrangement principles. This has been tested for the case of a 2D lid driven cavity flow. The GPU implementation is comprehensively compared with optimal serial and parallel CPU implementations on an Intel Nehalem X5550 platform. A maximum speedup of 252x as compared to the serial code on the CPU and 13x as compared to the parallel CPU code (16 MPI processes on8 cores of 2 Quad Nehalem X5550), has been attained.