Performance issues on many-core processors: A D2Q37 Lattice Boltzmann scheme as a test-case

Performances on recent processor architectures heavily rely on the ability of applications and compilers to exploit a more and more diverse and large set of parallel features. In this paper we focus on issues related to the efficient programming of multi-core processors based on the Sandy Bridge micro-architecture recently introduced by Intel. As a test-case application we use a D2Q37 Lattice Boltzmann algorithm, which accurately reproduces the thermo-hydrodynamics of a 2D-fluid obeying the equations of state of a perfect gas. The regular structure and the high degree of parallelism available in this class of applications make it relatively easy to exploit several processor features relevant for performance, such as, for example, the new Advanced Vector Extension (AVX) SIMD instructions set. However the main challenge is how to efficiently map the application onto the hardware structure of the processor. In this paper we present the implementation of our Lattice Boltzmann code on the Sandy Bridge processor, and assess the efficiency of several programming strategies and data-structure organizations, both in terms of memory access and computing performance. We also compare our results with that obtained on previous generation Intel processors, and with recent NVIDIA GP-GPU computing systems. (C) 2013 Elsevier Ltd. All rights reserved.