NSF-CGP Fellowship: Performance Studies of the EM-X Distributed-Memory Multiprocessor

9722187 Sohn该奖项提供资金支持新泽西理工学院计算机与信息科学系Andrew Sohn博士与日本筑波电工实验室计算机科学部Yoshinori Yamaguchi博士进行两次为期9个月的研究访问。他们的研究涉及开发算法，以调查何时，何地以及如何多线程和多线程架构可以减少高性能计算机中并行处理器之间的通信时间。为了实现他们的目标，合作者有一个由以下五个步骤组成的技术计划：(1)为选定的基准问题设计单线程和多线程算法；(2)制定不同线程粒度的数据和工作负载分配策略；(3)实现算法以比较三种多处理器的性能，即多线程的EM-X（由日本电工实验室设计和制造）和单线程的IBM SP-2和Cray T3E；(4)分析表征单线程和多线程机器性能的参数；(5)建立预测单线程和多线程性能的模型。这项研究的主要结果将是一套工具和方法，能够识别多线程和多线程并行机器与单线程机器的性能。这组工具和方法将帮助算法设计者将单线程算法转换为多线程版本。研究结果还将帮助未来的超级计算机架构师开发多线程机器，以减少通信时间，实现高性能计算。孙博士和他的日本合作者在这一研究领域享有很高的声誉。本项目由美国国家科学基金会和日本财团全球伙伴关系中心科学奖学金计划资助。＊＊＊

9722187 Sohn This award provides funds to support two research visits, for a total of nine months, by Dr. Andrew Sohn, Department of Computer and Information Science, New Jersey Institute of Technology, for collaboration with Dr. Yoshinori Yamaguchi, Computer Science Division, Electrotechnical Laboratory, Tsukuba, Japan. Their research is concerned with developing algorithms to investigate when, where, and how multithreading and multithreaded architecture can reduce communication time between parallel processors in high performance computers. To achieve their objectives, the collaborators have a technical plan consisting of the following five steps: (1) design single-threaded and multithreaded algorithms for selected benchmark problems; (2) develop data and workload distribution strategies with varying thread granularity; (3) implement the algorithms to compare the performance of three multiprocessors, the multithreaded EM-X (designed and built by the Electrotechnical Laboratory, Japan), and the single-thread IBM SP-2, and Cray T3E; (4) analyze parameters that characterize the performance of single-threaded and multithreaded machines; and (5) formulate a model to predict the performance achieved by single threading and multithreading. The principle result of this research will be a set of tools and methodologies which enable the identification of the performance of multithreading and multithreaded parallel machines against single-threaded machines. This set of tools and methodologies will help algorithm designers to convert single-threaded algorithms to a multithreaded version. The results also will help future supercomputer architects develop multithreaded machines for reducing communication time towards high performance computing. The Dr. Sohn and his Japanese collaborator are highly regarded investigators in this field of research. This project is supported under the Science Fellowship Program between the National Science Foundation and the Japan Foundation Center for Global Partnership. ***