Chinese Journal of Computers   Full Text
Title	Research on Component Parallel Technology Based on Fuzzy Clustering Analysis
Authors	DU Jing1) AO Fu-Jiang2) YANG Xue-Jun1) YANG Can-Qun1)
Address	1)(School of Computer, National University of Defense Technology, Changsha 410073) 2)(School of Mechatronics and Automation, National University of Defense Technology, Changsha 410073)
Year	2007
Issue	No.11(1939—1946)
Abstract &   Background	Abstract This paper proposes a new scientific computing-oriented component technology—Component Parallel Technology Based on Fuzzy Clustering Analysis, aiming at improving parallelism and data locality among components, and avoiding communication bottleneck. The technology is composed of two parts: Domain partition and sub-component combination. Domain partition uses data dependence analysis technique during compile time. Then considering the effect of access stride, the concept of interval overlap degree is proposed by using indefinite equation. Based on this, it implements the classification and combination of sub-components by using fuzzy clustering algorithm for interval overlap degree designed by the authors, and presents the formal description of the algorithm. The experimental results show that the algorithm is efficient and scalable for scientific component programs in terms of fine data locality, moderate granularity and high parallelism. keywords communication bottleneck; parallelism; data locality; interval overlap degree; fuzzy clustering background With the advancement of high performance computing, the complexity of scientific computing software has grown tremendously. It is significant to use component technology in high performance scientific computing for the needs of large-scale, complex, high-performance scientific software. The key technique of using scientific component is how to execute the components in parallel more efficiently. The existing systems just reorganize the data set within components at run-time to perform data parallelism, and thereby increase the run-time overhead, restrict data parallelism among components and cause the communication bottleneck between components. To address the problems, the authors propose a new compile-time scientific computing-oriented component technology—Component Parallel Technology Based on Fuzzy Clustering Analysis, aiming at improving parallelism and data locality among components, and avoiding communication bottleneck. This paper presents the details of the new technique, including domain partition via data dependence analysis, classification and combination of sub-components by using fuzzy clustering algorithm for interval overlap degree designed by the authors and the corresponding formal algorithm description. The experimental results show the efficiency of the new scientific computing-oriented component technology. The work of this paper was supported by the Natural Science Foundation of China under grant Nos.60621003 and 60633050, in which the efforts in this paper focus on solve the problems of programming complexity and parallel efficiency. The research group has tested plentiful experiments and gotten a lot of good efforts in scientific computing components.