Performance Measurement and Modeling of Parallel Computer Systems

这项研究计划补助金的目的是进行初步研究，并制定一个完整的建议，将提交给国家科学基金会在1995年秋季。 研究的主要目标是并行计算机系统的测量和建模研究。 所使用的平台包括位于橡树岭国家实验室的Kendall Square KSR 1和Intel Paragon多处理器系统，以及位于菲斯克大学的Sun Solaris 2.3工作站网络。 评价方法是利用并行计算机系统的备选互连网络模型来度量和评价各种系统设计方案。 选择嵌入式网络模型是因为，与诸如马尔可夫模型和Petri网模型的其他方法不同，嵌入式网络模型不需要枚举系统可能进入的所有可能状态，而是使用计算高效的技术来确定系统的平均性能度量。 并行和分布式计算机和通信系统的使用正变得普遍。 然而，作为本研究重点的计算机系统具有难以测量和建模的特性。 这些系统可能在任务图中有分叉和连接，在多道程序设计级别中有变化，并使用各种同步技术。 诸如线程调度和放置之类的系统问题很难精确建模。 因此，开发和应用计算效率高的技术来评估这些类型的系统的性能是一个重要的和有意义的研究领域。 该项目包括在菲斯克大学的本科生研究经验的组成部分。 给予适当的指导，本科生有机会改变参数，运行实验，并参与测量和建模过程。

The purpose of this Research Planning Grant is to perform preliminary research and to develop a full proposal which will be submitted to NSF in the fall of 1995. The primary objective of the research is a measurement and modeling study of parallel computer systems. The platforms to be used include the Kendall Square KSR1 and the Intel Paragon multiprocessor systems located at Oak Ridge National Laboratory, and a network of Sun Solaris 2.3 workstations located at Fisk University, an HBCU/MI. The method of evaluation is to measure and evaluate various system design alternatives using alternative queueing network models of parallel computer systems. Queueing network models are chosen because, unlike other methods such as Markov models and Petri net models, queueing network models do not require the enumeration of all possible states that the system may enter, but rather, use computationally efficient techniques to determine mean performance metrics of the system. The use of parallel and distributed computer and communications systems is becoming commonplace. However, the computer systems which are the focus of this research have characteristics which are difficult to measure and model. The systems may have forking and joining in the task graph, variations in the multiprogramming level and use various synchronization techniques. System issues such as thread scheduling and placement are difficult to model accurately. Thus, the development and application of computationally efficient techniques for evaluating the performance of these types of systems is an important and significant area of research. The project includes a component for undergraduate research experiences at Fisk University. Given appropriate guidance, undergraduate have the opportunity to alter parameters, run experiments, and participate in the measuring and modeling process.