high performance computing on graphics processing units: hgpu.org

High-performance computing systems are more and more often based on accelerators. Computing applications targeting those systems often follow a host-driven approach in which hosts offload almost all compute-intensive sections of the code onto accelerators; this approach only marginally exploits the computational resources available on the host CPUs, limiting performance and energy efficiency. The obvious step forward is to run compute-intensive kernels in a concurrent and balanced way on both hosts and accelerators. In this paper we consider exactly this problem for a class of applications based on Lattice Boltzmann Methods, widely used in computational fluid-dynamics. Our goal is to develop just one program, portable and able to run efficiently on several different combinations of hosts and accelerators. To reach this goal, we define common data layouts enabling the code to exploit efficiently the different parallel and vector options of the various accelerators, and matching the possibly different requirements of the compute-bound and memory-bound kernels of the application. We also define models and metrics that predict the best partitioning of workloads among host and accelerator, and the optimally achievable overall performance level. We test the performance of our codes and their scaling properties using as testbeds HPC clusters incorporating different accelerators: Intel Xeon-Phi many-core processors, NVIDIA GPUs and AMD GPUs.