Performance Assessment of Fluid Dynamics Codes on Different Computer Architectures

Raymond P LeBeau, Daniel Reasor, Xiaolong Deng, Swetha Panguluri, Narendra Beliganur

University of Kentucky, Lexington, KY

Thomas Hauser

Utah State University, Logn, Utah

The application of Computational Fluid Dynamics (CFD) to more demanding problems has brought along a complementary interest in achieving high efficiency, cost-effective simulations. This objective is obviously dependent on the available computer system as well as on the code. However, determining the best hardware for a given code has grown more challenging in the past few years as chip manufacturers have moved from the paradigm of ever-accelerating clock speeds to new architectures with multiple cores, faster and more complicated cache structures, and other similar structural improvements. In this enviroment, different CFD codes may find different architectures more suitable for optimal performance. This presentation will cover the systematic evaluation of several CFD codes across numerous platforms, with a primary focus on commodity cluster systems. The codes tested include two in-house incompressible CFD codes, one two-dimensional structured and one three-dimensional unstructured, the widely used three-dimensional structured code OVERFLOW, and the planetary general circulation model EPIC. The platforms will range from classic 32-bit to 64-bit, mult-core processors, covering the past few year of architecture development. General themes relating cache efficiency and cross-platform performance are investigated since this currently appears to play a critical role in processor performance and the determination of the best processor choice for a given application.