Exploiting parallelism in many-core architectures: Lattice Boltzmann models as a test case

In this paper we address the problem of identifying and exploiting techniques that optimize the performance of large scale scientific codes on many-core processors. We consider as a test-bed a state-of-the-art Lattice Boltzmann (LB) model, that accurately reproduces the thermo-hydrodynamics of a 2D-fluid obeying the equations of state of a perfect gas. The regular structure of Lattice Boltzmann algorithms makes it relatively easy to identify a large degree of available parallelism; the challenge is that of mapping this parallelism onto processors whose architecture is becoming more and more complex, both in terms of an increasing number of independent cores and – within each core – of vector instructions on longer and longer data words. We take as an example the Intel Sandy Bridge micro-architecture, that supports AVX instructions operating on 256-bit vectors; we address the problem of efficiently implementing the key computational kernels of LB codes – streaming and collision – on this family of processors; we introduce several successive optimization steps and quantitatively assess the impact of each of them on performance. Our final result is a production-ready code already in use for large scale simulations of the Rayleigh-Taylor instability. We analyze both raw performance and scaling figures, and compare with GPU-based implementations of similar codes.