State-Time Approach for Analysis and Simulation of Complex Multicomponent Systems Using Future Massively Parallel Computing Systems

科特·S·安德森机械、航空航天和核工程系伦斯勒理工学院NSF提案编号0219734使用未来大规模并行计算资源进行复杂多组件系统分析和仿真的状态时间方法这项工作的主要目标是研究执行复杂动态系统分析、仿真和优化的状态时间方法。目标是产生能够更有效和充分地利用预期的大规模并行计算资源(即106个处理器)的动态系统分析公式。所提出的时空公式允许将运动方程中的时间作为广义坐标来处理。这允许在空间和时间上对执行计算机模拟和相关分析的问题进行并行化，从而产生更高水平的粗粒度并行化，并相应地减少了模拟周转时间。这一显著增长源于这样一个事实，即当今最先进的并行动力学分析算法在时间上是连续的。因此，使用当前推广的方法在系统域上进行并行化只有在使用数量惊人的有限处理器时才是有益的，即使有许多处理器也是如此。在空间和时间上并行化会导致可能分布在可用处理器上的粗粒度计算的数量急剧增加。通过形成方程并以所提出的状态-时间方式求解它们，所需的顺序计算的水平相对于可以并行执行的水平被减少了大约等于使用当前方法执行的时间积分步骤(通常为105步)的数量。拟议的方法如果成功，将允许对更复杂的系统使用更详细的模型，结果只需使用当前公式所需时间的一小部分。

Kurt S. AndersonDepartment of Mechanical, Aerospace and Nuclear EngineeringRensselaer Polytechnic InstituteNSF Proposal No. 0219734State-Time Approach for Analysis and Simulation of Complex Multicomponent Systems Using Future Massively Parallel Computing ResourcesThe principal objectives of this work are to research state-time methods for performing analysis, simulation, and optimization of complex dynamic systems. The goal is to produce dynamic systems analysis formulations which will be able to more effectively and fully exploit anticipated massively parallel computing resources (i.e. 106 processors). The proposed space-time formulation permits the treatment of time within the equations of motion as a generalized coordinate. This permits the problem of performing the computer simulation, and associated analysis to be parallelized over both space and time, resulting in a far greater level of coarse grain parallelization with an associated decrease in simulation turnaround time. This significant increase arises from the fact that the most advanced of today's parallel dynamics analysis algorithms are sequential in time. Thus parallelization over the domain of the system using currently promoted methods is only beneficial when using a surprisingly limited number of processors, even if many are available. Parallelizing over both space and time results in a drastic increase in the number of coarse grain calculations that may be distributed over the available processors. By forming the equations and solving them in the proposed state-time manner, the level of required sequential calculation relative that which may be performed in parallel is reduced by a factor approximately equal to the number of temporal integration steps (often 105) which would have been performed using current approaches. The proposed method, if successful, will allow the use of more detailed models, for more complex systems, with results obtained at a small fraction of the time required using current formulations.