Juan J. Alonso, Associate Professor, Department of Aeronautics & Astronautics
Francisco Palacios, Engineering Research Associate, Department of Aeronautics & Astronautics
Thomas Economon, PhD Candidate, Department of Aeronautics & Astronautics
High-Performance, Open-Source CFD Analysis and Design on the Intel® Xeon Phi™ Coprocessor
The solution of Partial Differential Equations (PDEs) is the basis for predictive simulations in Computational Fluid Dynamics (CFD) to analyze of a wide range of problems including turbulence, acoustics, heat transfer, vertical flows, and combustion. CFD simulations run the gamut of computational expense, from simple, single-processor jobs to highly-complex computations distributed over millions of cores. Improving the performance of such simulations will allow more accurate predictions, and permit the use of methodologies that are currently prohibitively expensive. As improvements in single-core clock-speed have stalled, improvements in the parallelization of CFD codes provide the greatest opportunity for improvements in wall-clock solution time.
Heterogeneous computers with millions of accelerated cores are becoming widespread in scientific computing and will be the future of exascale architectures. Unfortunately, it is increasingly difficult to achieve high levels of performance, and the scalability required by aerospace applications will only be realized through investment in algorithmic improvements to fully utilize all resources. The Intel® Xeon Phi™ coprocessor architecture, with many cores and low communication bandwidth, promises to be a revolution for CFD.
We believe that not every engineer needs to become an expert on the diversity of architectures populating the market today. For this reason, and over the past 3 and half years, we have developed and supported an open-source software tool suite, the Stanford University Unstructured (SU2) suite, that enables engineers to analyze complex shapes that interact with a fluid, such as an aircraft wing or wind turbine blade, and to optimize it to obtain high levels of performance. We believe that open-source solutions such as SU2 can serve as a platform to leverage research in high-performance parallel computing for use in the broader community. Our Intel® Parallel Computing Center will focus on high-fidelity applications in the engineering fields described above, and on breaking current paradigms in three significant ways using the Xeon Phi™ coprocessor architecture:
- Create an optimized implementation of the complete SU2 suite on the Intel® Xeon Phi™ coprocessor, through research and adoption of best practices including vectorization, hybrid OpenMP/MPI programming, and full parallel I/O.
- Understand the suitability of and customize algorithms for implementation on Intel® Xeon Phi™ co-processors.
- Enable scalability to very large numbers of Xeon Phi™ coprocessors.
The result of our research will be included in the SU2 suite and available to the community at large.