Theoretical Analysis of Parallel Data Decomposition on Cluster Grid

• Main Authors: Wang, Chun Ching, 王俊清
• Other Authors: Hsu, Ching Hsien
• Format: Others
• Language: en_US
• Published: 2007
• Online Access: http://ndltd.ncl.edu.tw/handle/34006246002457191335

碩士 === 中華大學 === 資訊工程學系(所) === 95 === With the development of cheap personal computers and high-speed network devices, clusters have become the trend in the design of high performance computing environments. As the researches of relative hardware and software technology of Cluster and Grid are constantly improving, the application of Cluster is growing popular. Due to information progress and increasing calculation capacity required by all kinds of applications; the calculation needed by these applications have also extended to cross-network calculation. Through the Internet, which connects several Clusters, the mass calculation platform is combined into Cluster Grid. Data partition and exchange which are delivered by network may happen during the executing program and in Cluster Grid environment. Network speed and communication localization therefore are important factors in programming efficiency. In traditional parallel computer and scatter memory environment, logical processor mapping, communication scheduling, data partitioning, data redistribution are often used to reduce the load of processor exchanging data. But in Cluster Grid environment, data delivery of cross-network has not been taken into consideration. The data, delivered in Cluster Grid environment, can be roughly categorized as external exchange and internal exchange. Since internal exchange data in the same cluster need no cross-networking, so the data transmission speed is better and it close to more efficient. In this thesis, we propose a new mathematical method, as a result, that under conditional circumstances, it achieves excellent data partitioning and maintains data calculation in local environment. In addition, we also conduct some theoretical analysis on the amounts of computing nodes under entire Cluster Grid environment and amounts of date partition in the hope that through the analysis, the results could be applied to practical parallel environment and further to reduce communication cost.