Accelerating Computational Fluid Dynamics Codes on Multi-/Many-Core Intel Platforms

In this paper, we present optimization techniques that are crucial to unlock parallelism and vectorization in modern computational fluid dynamics (CFD) codes thereby significantly improving their performance on emerging Intel multi-/many-core platforms such as the Intel R © Xeon R © processors and Intel R © Xeon Phi coprocessors. We focus on unstructured-mesh finite-volume codes and restrict the discussion to fine-scale optimizations with the objective to improve the strong-scaling behavior of these classes of algorithms. We present the key architectural features of the Intel Xeon Phi coprocessor and describe strategies to exploit them for improving performance of three widely used CFD codes. Our benchmarking results show substantial performance advantages to speed up time-to-solution.