{"title":"Configurable PC clusters using a hierarchical complete-connection-based switching network","authors":"N. Tsuda","doi":"10.1109/PRDC.2002.1185632","DOIUrl":null,"url":null,"abstract":"An advanced interconnection network called \"HCC-ABL-tree (hierarchical completely-connected tree by additional bypass linking)\" is proposed for constructing large PC clusters capable of distributed inter-node communication by using Ethernet switches with a small number of I/O-ports. A basic one-level CC-ABL-tree can be constructed by using small subarrays of processing nodes (PCs) providing interconnections with the complete connection scheme, and by connecting the nodes to a tree-structured bypass network with switches with a height of two so that every node of the subarray is allocated to a different second-level switch. A two- or three-level HCC-ABL-tree can be constructed by using the trees with one less number of levels as the components by connecting them to each other with the complete-connection scheme by using the second-level switches of the bypass network as the hierarchical interconnections. The maximum number of processing nodes in a cluster can be increased exponentially by increasing the number of hierarchical levels. The network diameter is two for a one-level tree, three for a two-level tree, and seven for a three-level tree. The proposed network can configure the processing nodes in the cluster as a square-mesh-connected array with any aspect ratio by graph embedding, where a newly proposed distributed routing algorithm can define the paths with no duplicated use of a link. This manner of configuring can also be achieved even when busy or faulty nodes exist in the cluster, while bypassing these nodes in a node-disjoint manner with a small congestion and dilation in the paths.","PeriodicalId":362330,"journal":{"name":"2002 Pacific Rim International Symposium on Dependable Computing, 2002. Proceedings.","volume":"286 1","pages":"0"},"PeriodicalIF":0.0000,"publicationDate":"2002-12-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"2002 Pacific Rim International Symposium on Dependable Computing, 2002. Proceedings.","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.1109/PRDC.2002.1185632","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
引用次数: 0
Abstract
An advanced interconnection network called "HCC-ABL-tree (hierarchical completely-connected tree by additional bypass linking)" is proposed for constructing large PC clusters capable of distributed inter-node communication by using Ethernet switches with a small number of I/O-ports. A basic one-level CC-ABL-tree can be constructed by using small subarrays of processing nodes (PCs) providing interconnections with the complete connection scheme, and by connecting the nodes to a tree-structured bypass network with switches with a height of two so that every node of the subarray is allocated to a different second-level switch. A two- or three-level HCC-ABL-tree can be constructed by using the trees with one less number of levels as the components by connecting them to each other with the complete-connection scheme by using the second-level switches of the bypass network as the hierarchical interconnections. The maximum number of processing nodes in a cluster can be increased exponentially by increasing the number of hierarchical levels. The network diameter is two for a one-level tree, three for a two-level tree, and seven for a three-level tree. The proposed network can configure the processing nodes in the cluster as a square-mesh-connected array with any aspect ratio by graph embedding, where a newly proposed distributed routing algorithm can define the paths with no duplicated use of a link. This manner of configuring can also be achieved even when busy or faulty nodes exist in the cluster, while bypassing these nodes in a node-disjoint manner with a small congestion and dilation in the paths.