Latency analysis of TCP on an ATM network
WTEC'94 Proceedings of the USENIX Winter 1994 Technical Conference on USENIX Winter 1994 Technical Conference
Hi-index | 0.00 |
Modern workstations increasingly rely on distributed software such as NFS and NIS, yet the speed of networking software is not improving as rapidly as the workstation and networking hardware, leading to a network software bottleneck. We present detailed measurements of various components of the TCP/IP and UDP/IP protocol stack on a DECstation 5000/200 running Ultrix 4.2a. Measurements are by layer (i.e. socket, transport, IP, data-link) and by function (i.e. checksum computation, data copying, buffer management, protocol processing, and operating system interaction), with further breakdowns within each category. We show that checksum computation and data transfers dominate component times for a real LAN workload, that using large packet MTUs (maximum transmission units) is very important to achieving high through- put, and that given the distribution of component times for small sized messages, it will be difficult to improve latency. TCP and UDP time breakdowns are shown to be quite similar, suggesting that "light- weight" transport protocols are not likely to greatly decrease processing time. Finally, analytical models for network software processing times are presented.