high performance computing on graphics processing units: hgpu.org

The proposed paper will present details on recent advancements in scientific computing in terms of integrating new hardware and software to greatly enhance the computational efficiency of comprehensive rotorcraft analysis. The focus will be on showing the tremendous computational accelerations that are possible (i.e., orders of magnitude speed up) by using software developments in the form of the fast multipole method (FMM) combined with hardware accelerations by using graphics processors (GPUs). Even today, for many applications a comprehensive rotorcraft analyses may prove too computationally expensive (e.g., for several aspects of design and analysis) because of the significant computational requirements that are needed and wall-clock times incurred. A good example is when free-vortex wake analyses are exercised within such comprehensive codes, especially when the simulations are conducted for maneuvering flight and/or over long time scales. To avoid high computational costs, the fidelity of the analysis may have to be compromised by using prescribed wake or small disturbance linearized models instead, e.g., incarnations of dynamic inflow. These simpler models may also be "tuned" to predict solutions for specific flight conditions and may not have enough predictive capabilities (although their failure may not be immediately recognized) in other flight conditions. Recent interest in the prediction of brownout dust clouds is a good example of a complex problem in aeromechanics that requires the use of comprehensive rotorcraft models. For such problems, which involves the simulation of maneuvering flight near the ground, the use of a free-vortex wake method is required at a minimum. However, because of the tracking of substantial number of vortex elements (N) in the flow simulation and a much larger number of particles (M) for the brownout simulations, the basic computational cost of such algorithms is O(N^2 + NM). This cost may translate into weeks of compute time within some comprehensive analysis environments. In the proposed paper, two methods will be discussed to deal with the practical implementation of such problems: 1. Development of efficient algorithms to reduce the computational cost, and 2. Utilization of parallel hardware for highperformance computing. In particular, the fast multipole method (FMM) has complexity O(N logN) and is shown to be very appropriate for this type of problem. The hardware acceleration was implemented using graphics processors (GPUs) that have a low cost highly parallel architecture consisting of hundreds of cores. The results that will be presented in the final paper will show feasibility of this approach for several applications, and how to obtain the computational accelerations that are up to two orders of magnitude faster. However, the potential for much faster algorithmic accelerations is also shown to be dependent on the nature of the aeromechanics problem and the precision that is required.