RIA: Cost-effective Parallel Computing Platforms Based on SCI-connected Distributed Systems

9410063 Li本建议研究以SCI(Scalable Coherence Interface，IEEE1596标准)连接的个人工作站网络的系统架构模型和系统软件支持。与传统的局域网连接的NPW不同，SCI连接的NPW可以以高带宽(1G字节/秒)和低延迟(微秒)的方式在整个网络中的工作站之间提供物理共享的内存和缓存一致性。这项研究将提供对新的并行结构的理解，这种结构突出的是由于共享存储器和高速缓存一致性而产生的紧密耦合的性质与由于工作站之间的物理距离和工作站的独立操作而产生的松散耦合的性质之间的矛盾。这项研究将产生各种SCI连接NPW设计的模型及其比较，支持此类体系结构的操作系统设计，以及基于典型并行应用程序的性能数据。这一结果将是朝着廉价但强大和广泛可用的并行计算平台迈出的重要一步。研究工作将分三个步骤进行：(1)建立SCI-Connected NPE的建筑特征模型。(2)根据硬件和应用需求，设计并实现操作系统软件支持。还将研究系统级性能。(3)设计并实现应用原型，获取性能数据。仿真将用于预测大规模SCI-NPWS的性能。***

9410063 Li This proposal research is to study the system architectural model and system software support of network of personal workstations (NPW) connected by SCI (Scalable Coherence Interface, IEEE Standard 1596). Oppose to the traditional LAN-connected NPW, SCI- connected NPW can provide physically shared memory and cache coherence among workstations across a network with high bandwidth (1Gbyte/sec) and low latency (microseconds). The research will provide the understanding of the new parallel architecture which is highlighted by contradiction between its closely-coupled nature due to shared memory and cache coherence and its loosely- coupled nature due to the physical distance between workstations and independent operations of the workstations. This research will result in models of various designs of SCI-connected NPW and their comparisons, operating system designs to support such architectures, and performance data based on typical parallel applications. The result will be an important step towards inexpensive but powerful and widely available parallel computing platforms. The work will be done in three steps: (1) Model the architectural characteristics of SCI-connected NPE. (2) Design and implement operating system software supports based on the hardware and applications requirements. System level performance will also be studied. (3) Design and implement application prototypes to obtain performance data. Simulation will be used to predict the performance of large scale SCI-NPWs. ***