M. E. Gómez, J. Flich, A. Robles, P. López, J. Duato
{"title":"VOQ/sub SW/: a methodology to reduce HOL blocking in InfiniBand networks","authors":"M. E. Gómez, J. Flich, A. Robles, P. López, J. Duato","doi":"10.1109/IPDPS.2003.1213134","DOIUrl":null,"url":null,"abstract":"InfiniBand is a new switch-based standard interconnect for communication between processor nodes and I/O devices as well as for interprocessor communication. InfiniBand architecture allows switches to support up to 15 virtual lanes per port for data traffic. To route packets through a given virtual lane (VL), packets are labeled with a certain service level (SL) at injection time, and SLtoVL mapping tables are used at each switch to determine the VL to be used. Many previous works in the literature have shown that separate virtual lanes are able to reduce the influence of the well-known head-of-line (HOL) blocking effect on network performance. However, using virtual lanes to form separate virtual networks is not enough to eliminate the HOL blocking problem. Alternative solutions such as Virtual Output Queuing (VOQ) are able to eliminate it at the expense of modifying the switch buffer organization. In this paper, we propose an effective strategy to implement the VOQ scheme in IBA switches by using virtual lanes. This strategy does not require to modify the switch architecture, simply SL to VL tables must be properly filled. Evaluation results show that our proposed VOQ scheme is able to outperform the results obtained with the virtual network approach using the same number of resources. Moreover, the methodology proposed to implement the VOQ scheme in IBA only requires a small number of resources in order to significantly improve network throughput.","PeriodicalId":177848,"journal":{"name":"Proceedings International Parallel and Distributed Processing Symposium","volume":"54 1","pages":"0"},"PeriodicalIF":0.0000,"publicationDate":"2003-04-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"13","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Proceedings International Parallel and Distributed Processing Symposium","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.1109/IPDPS.2003.1213134","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
引用次数: 13
Abstract
InfiniBand is a new switch-based standard interconnect for communication between processor nodes and I/O devices as well as for interprocessor communication. InfiniBand architecture allows switches to support up to 15 virtual lanes per port for data traffic. To route packets through a given virtual lane (VL), packets are labeled with a certain service level (SL) at injection time, and SLtoVL mapping tables are used at each switch to determine the VL to be used. Many previous works in the literature have shown that separate virtual lanes are able to reduce the influence of the well-known head-of-line (HOL) blocking effect on network performance. However, using virtual lanes to form separate virtual networks is not enough to eliminate the HOL blocking problem. Alternative solutions such as Virtual Output Queuing (VOQ) are able to eliminate it at the expense of modifying the switch buffer organization. In this paper, we propose an effective strategy to implement the VOQ scheme in IBA switches by using virtual lanes. This strategy does not require to modify the switch architecture, simply SL to VL tables must be properly filled. Evaluation results show that our proposed VOQ scheme is able to outperform the results obtained with the virtual network approach using the same number of resources. Moreover, the methodology proposed to implement the VOQ scheme in IBA only requires a small number of resources in order to significantly improve network throughput.