Equipment in Support of Parallel Processing Research

该奖项是对将Ncube并行处理器升级到下一代Ncube处理器的奖励。研究项目包括用于计算机辅助设计的并行分层模拟系统、用于信号处理的并行迭代算法和并行有限元方法。为了解决现实世界的问题，升级后的Ncube处理器是必要的。90年代的S面临的计算挑战是将并行化应用于解决重要的计算密集型问题。下一波超级计算机浪潮将是使用数千至数百万个独立处理器并行工作来解决问题。必须有某种机制来在处理器之间平均分配工作，以便最大限度地利用并行性。这项研究着眼于三个特别适合并行体系结构的问题领域，在并行体系结构中，每个处理器都有自己的内存，并执行自己的程序。第一个项目是医学图像重建的信号处理。为了实现这种医学图像重建技术，需要在快速的并行处理器上运行良好的并行算法，以获得用于诊断目的的最高质量的图像是非常重要的。第二个项目涉及VLSI电路的设计和仿真。这是一个计算特别密集的过程，模块化程度足够高，因此可以在Ncube处理器上进行编程。为了在实际的VLSI电路上运行仿真，需要进行升级。最后一个项目是探索一套偏微分方程解的有限元方法。当并行化时，这些技术可能在各个计算元件之间表现出令人望而却步的通信成本。新的Ncube处理器上的额外内存和更好的通信允许进一步探索这些数值解技术。

This award is for the upgrade of an NCUBE parallel processor to the next generation NCUBE processor. The research projects are in parallel hierarchical simulation systems for use in computer-aided design, parallel iterative algorithms for signal processing, and parallel finite element methods. The upgraded NCUBE processor is necessary in order to tackle real world problems. The computing challenge of the 90's is to put parallelism to work in solving important computational intensive problems. The next wave of super computers will be using thousands to millions of individual processors working in parallel to solve problems. There must be some mechanism to split the work up evenly amongst the processors in order to utilize parallelism to the fullest extent. This research looks at three problem domains particularly suited for parallel architectures where each processor has its own memory and executes its own program. The first project is signal processing for medical image reconstruction. It is very important the highest quality images be available for diagnostic purposes and good parallel algorithms running on a fast parallel processor are required in order to implement this medical image reconstruction technology. The second project involves the design and simulation of VLSI circuits. This is a particularly computation intensive process that is modular enough so that it can be programmed on an NCUBE processor. The upgrade is needed to run the simulations on real VLSI circuits. The final project is the exploration of a set of techniques known as finite element methods for partial differential equation solution. These techniques, when parallelized, may exhibit a prohibitive communication cost between the individual computing elements. The additional memory and better communications on the new NCUBE processor allows further exploration of these numerical solution techniques.