• Article
  • | |
  • Metrics
  • |
  • Reference [10]
  • |
  • Related [20]
  • |
  • Cited by [6]
  • | |
  • Comments
    Abstract:

    The software overhead in IN (interconnection network) communication has become currently the bottleneck of cluster system performance. In this paper, a method is proposed to reduce the communication software overhead by providing a fast message passing mechanism at user level——ULFM (user level fast message), that has been implemented in the IN TH-GBNet of cluster system. In ULFM, a new type of network access interface is provided to form a communication buffer centric application access to the cluster system IN. Several techniques, such as user level network access, zero copy message passing and simplified communication protocol, are integrated together for ULFM, that can effectively reduce the communication software overhead and achieve high communication performance with low latency and high bandwidth for application programs. Experimental results show that the latency at application level between different nodes for 16-byte message can be reduced to only 6μs, and the bandwidth can reach to 40MB/s for 2K byte length message.

    Reference
    [1] Horie T. Improving AP1000 parallel computer performance with message communication. In: Proceedings of the 20th International Symposium of Computer Architecture. New York: ACM Press, 1993. 314~325.
    [2] Boden, N.J. Myrinet: a gigabit-per-second local area network. IEEE Micro, 1995,15(1):21~28.
    [3] Blumrich, M.A. Network interface for protected, user-level communication [Ph. D. Thesis]. Princeton University, 1996.
    [4] Du, Yi, Li, San-li. Design and routing algorithm of the high-speed TH-switch in k-array n-cube interconnection network. Chinese Journal of Computers, 1998,21(10):873~880 (in Chinese).杜毅,李三立.可扩展高速互连网络TH-GBNet的设计与实现.计算机学报,1998,21(10):873~880.
    [5] Hwang, Kai. Advanced Computer Architecture: Parallelism Scalability Programmability. McGraw-Hill, Inc., 1993.
    [6] Von Eicken, T., Basu, A., Buch, V., et al. Vogels: U-Net: a user-level network interface for parallel and distributed computing. In: Proceedings of the 15th Symposium of Operating System Principles. 1995. 40~53.
    [7] Warschko, T.M. The ParaStation project: using workstations as building blocks for parallel computing. In: Arabnia, H., ed. Proceedings of the Parallel & Distributed Processing Technique and Applications 1996, Vol I. 1996. 375~386.
    [8] Shen, Jun. FMP: a fast messages passing for workstation clusters. Chinese Journal of Computers, 1998,21(7):595~602 (in Chinese).申俊.FMP:一种适用于机群系统的快速消息传递机制.计算机学报,1998,21(7):595~602.
    [9] Dubnicki, C., Bilas, A., Li, Kai, et al. Design and implementation of virtual memory-mapped communication on myrinet. In: Proceedings of the 1997 Interational Parallel Processing Symposium. 1997. 388~396.
    [10] Welsh, M., Basu, A., von Eicken, T. Low-Latency communication over fast ethernet. In: Proceedings of the Euro-Par'96. Springer Verlag, Lyon, France, 1996.
    Comments
    Comments
    分享到微博
    Submit
Get Citation

周桂林,戈弋,李三立,黄震春,马群生.一种适用于机群系统的用户层消息传递机制.软件学报,2001,12(5):689-697

Copy
Share
Article Metrics
  • Abstract:3952
  • PDF: 4935
  • HTML: 0
  • Cited by: 0
History
  • Received:June 21,1999
  • Revised:February 18,2000
You are the first2045158Visitors
Copyright: Institute of Software, Chinese Academy of Sciences Beijing ICP No. 05046678-4
Address:4# South Fourth Street, Zhong Guan Cun, Beijing 100190,Postal Code:100190
Phone:010-62562563 Fax:010-62562533 Email:jos@iscas.ac.cn
Technical Support:Beijing Qinyun Technology Development Co., Ltd.

Beijing Public Network Security No. 11040202500063