A Fluid Simulation Methodology for Complex, Large Scale Networks

马萨诸塞大学阿默斯特分校和达特茅斯学院的研究人员联合提议开展流体模拟方法的基础研究，以实现下一代互联网等复杂网络的高效、多时间尺度、分层性能建模。所提出的努力将使网络更容易建模，并大大加快对其性能的评估；它将使建模比以前更大的网络成为可能，并允许在建模精度和解决速度之间进行定量评估。提出的流体建模方法的新特点包括：多时间尺度的建模。提出的统一建模框架提供了在不同时间尺度上对系统行为进行建模，允许在较粗时间尺度上建模时显着节省仿真执行时间。在较粗的时间尺度上对快时间尺度的行为进行建模时，开发用于定量表征建模精度潜在损失的形式化方法是该框架的重要组成部分。分层流体建模。pi表明，多层次（分层）流体建模允许通过更简单的模型抽象模拟的潜在复杂组件，并将该更简单的模型替换为更大的模拟。pi还开发了严格的技术，可以自动识别何时可以在不显着损失建模精度的情况下执行此类替换。复杂系统建模的效率。pi表明，除了提供多时间尺度、多层次的建模外，流体仿真可以比传统的离散事件仿真更有效。pi的初步工作表明，与离散事件相比，单处理器流体模拟速度提高了50倍，而建模精度几乎没有损失。在许多问题领域的适用性。pi证明了流体模拟方法是“即插即用”的，因为它们可以很好地应用于当前的互联网协议套件、未来的实时和多播协议或ATM网络。

Researchers from the University of Massachusetts at Amherst and DartmouthCollege jointly propose to perform fundamental research on developing afluid simulation methodology that will enable efficient, multi-time-scale,hierarchical performance modeling of complex networks such as the nextgeneration Internet. The proposed effort will allow networks to be moreeasily modeled and their performance to be evaluated significantly faster;it will make it possible to model significantly larger networks than beforeand to allow the tradeoffs between modeling accuracy and solution speed tobe quantitatively assessed. The novel features of the proposed fluidmodeling methodology include:Modeling at multiple time-scales. The proposed unified modeling frameworkprovides for modeling system behavior at different time-scales, allowing forsignificant savings in simulation execution times when modeling at coarsertime-scales. Development of formal means for quantitatively characterizingthe potential loss of modeling accuracy when fast-time-scale behavior ismodeled at a coarser time-scale is an important part of the framework.Hierarchical fluid modeling. The PIs demonstrate that multi-level(hierarchical) fluid modeling allows for abstraction of a potentiallycomplex component of the simulation by a simpler model and substitution ofthis simpler model into the larger simulation. The PIs also developrigorous techniques that can automatically identify when such substitutioncan be performed without significant loss of modeling accuracy.Efficiency in modeling complex systems. The PIs demonstrate that inaddition to providing for multi-time-scale, multi-level modeling, fluidsimulation can be significantly more efficient than traditional discreteevent simulation. The PIs' initial work shows a 50-fold speedup in auniprocessor fluid simulation speed over the discrete-event counterpart withlittle loss in modeling accuracy. Applicability in many problem domains. The PIs demonstrate that fluidsimulation methodologies are ``plug and play'' in the sense that they can beapplied equally well to the current Internet protocol suite, futurereal-time and multicast protocols, or ATM networks.