Software overhead in interconnection network communication has currently become the bottleneck of a cluster system. To reduce it, a user-level communication software UMPS based on real-time OS RTLinux is designed and implemented, which is comfortable with VIA. A new concept of semi-polling driven is presented. With the semi-polling driven mechanism, the interrupts frequency is lowered and the processing performance for short message is significantly ameliorated. By means of the address translation and buffer managing algorithm based on the resource-mapping graph, applications bypass OS and interact with network interface directly using asynchronous DMA. So the overhead and latency in communication are efficiently reduced. Experimental results indicate that the throughputs of UMPS for 64 byte and 1500 byte messages are 394 Mbps and 895 Mbps respectively, and the performance of UMPS surpasses that of other mechanisms.
[1]Chang S-L, Du DHC, Hsieh J, Lin MJ, Tsang RP. Enhanced PVM communications over a high-Speed Local Area Network. IEEE Parallel and Distributed Technology, 1995,3(3): 1405~1431.
[2]Pratt I, Fraser K. Arsenic: A User-Accessible Gigabit Ethernet Interface. Proc. of the INFOCOM, 2001,14(3):64~74.
[3]Pakin S, Karamcheti V, Chien AA. Fast messages (FM): Efficient, portable communication for workstation clusters and massively-parallel processors. IEEE Concurrency, 1997,5(2):60~90.
[4]Zhang Y, Zhang DY, Li SL. The Research and Implementation of Network Intrusion Detection System Based on Cooperative Distributed Agent. Chinese Journal of Computer. 2001,24(7):736~741 (in Chinese with English abstract).
[5]Lifeline K. Building into the Linux network layer. Phrack Magazine, 1999,9(55):55~12.
[6]Wolman A, Voelker G, Thekkath CA. Latency analysis of TCP on an ATM network. Proc. of the Winter USENIX Conf.1994,61(2):125~138.
[7]Blumrich MA, Dubnicki C, Felten EW, Li K, Mesarina MR. Virtual-Memory-Mapped Network Interfaces. IEEE Micro.1995,15(1):21~28.
[8]Welsh M, Basu A, Eicken TV. Incorporating memory management into user-level network interfaces. In: Proc. of Hot Interconnects V. 1997. 618~628
[9]Liu W, Zheng WM, Shen J, Ju DP. Design and Implementation of memory map mechanism in underlying communication. Journal of Software, 1999,10(1):24~28 (in Chinese with English abstract).
[10]Tezuka H, O'Carroll F, Hori A, Ishikawa Y. Pin-Down cache: A virtual memory management technique for zero-copy communication. Proc. of the Int'l Parallel Processing Symp. 1998,26(1):308~341.
[11]Mogul JC, Ramakrishnan KK. Eliminating receive livelock in an interrupt-driven kernel. ACM Transactions on Computer Systems,1997,15(3):217~252.
[12]Ang BS, Chiou D, Rudolph L, Arvind. Message passing support on start-voyager. In: Proc. of the 5th Int'l Conf. on High-Performance Computing. New York: IEEE Computer Society, 1998. 246~257.