Development and Application of High-End Engineering and Planetary Computational Fluid Dynamics(CFD) Models on Optimized PC Clusters

Computational fluid dynamics (~FD) is a discipline that is perpetually in search of grea~er co~puter power and better algonthms. This search has intensified as the phenomena stu~Ied wIth CFD model~ have expan~ed from ~aditional problems studied by aerospace ~ngIneers and meteorologIsts to mor~ dlv~rse desIgns and more demanding challenges. For Instance, th~ C~D gr~up at the UmversIty of Kentuck-y (UK) has worked on traditional chal.l~nges lIke s~mulatIng 1?e flow through a turbine cascade [Suzen et aI, 2000] and on nontr~ dItI?nal ones lIke ev~luatmg means for noise reduction in inkjet printers [Xiong et aI, 2000]. LIkeWIse, the C~mparatIve .Planetology L~boratory (CPL) at the University of Louisville (VoiL) has ~xp~ded hIgh-resolutIOn a~osphenc modeling from ten-estrial applications to planetary applIcatIOns such as Venus, JupIter [Showman & Dowling 2000], and Neptune [LeBeau & Dowling 1998]. High computational-cost CFD problems typically are solved at national supercomputer centers or similar facilities with limited accessibility. However, CFD design problems of a practical nature fall into this class regularly, requiring large corporations and universities to purchase expensive shared-memory supercomputers such as those built by SGI and HP. The high up-front cost of CFD analysis places the technology bcyond the reach of many mainstream engineers and researchers. Thus, many projects that would gain long-term economic advantages from CFD simulations never realize their potential. CFD code and system developers who find a practical way to tip the balance of this equation to affordable simulations will enjoy strong demand for their results. Even those orgarrizations with substantial computer budgets, such as large laboratories, industries, and operational weather-forecasting centers, stand to gain measurably from significantly cheaper and more efficient means of performing CFD calculations. We propose to develop inexpensive, high-performance clusters of PCs, or "Beowulfs" [Becker, et aI, 1995], specialized for CFD ,applications, using the novel approach that the hardware, operating system, and application code are optimized together rather than separately. We recognize that in the past few years some positive claims have been made in favor of PC clusters in which the hype was bigger than the performance of the actual systems. However, we will prove by construction that by expertly engineering the PC cluster design, using and building tools to improve application performance, and restructuring the application code for the cluster, it is possible to make grand-challenge-class CFD modeling accessible and affordable to mainstream engineers, scientists, and industrialists. A powerful prototype of such a system (KLAT2) was built at UK this year for only $40,000. This approach will be applied to three operational CFD codes developed in Kentuck~ that serve a range of NAsA thermo-fluids 'investigat~ons. The first is OVERFLOW; an establIshed CFD solver based on Reynolds-Averaged Navier-Stokes (RANS) turbulence models. The second is LESTool, a next-generation CFD model employing advanced turbulence modes. The third is EPIC (Explicit Planetary Isentropic-Coordinate), an atmospheric model used to study planetary atmospheres. Three research groups from UK and VofL will combine their talents with colleagues at NASA Ames, Langley, and JPL to carry out this concerted, groundbreaking CFD development and application effort. .