Characteristics of wide-area TCP/IP conversations

Abstract

In this paper, we characterize wide-area networkapplications that use the TCP transport protocol. Wealso describe a new way to model the wide areatraffic generated by a stub network. We believe thetraffic model presented here will be useful instudying congestion control, routing algorithms, andother source management schemes for existing andfuture networks. Our model is based on traceanalysis of TCP/IP wide-area internetworktraffic. We collected the TCP/IP headers of USC,UCB, and Bellcore networks at the point they connectwith their respective regional access networks. Wethen wrote a handful of programs to analyze thetraces. Our model characterizes individual TCPconversations by the distributions of: number ofbytes transferred, duration, number of packetstransferred, packet size, and packet interarrivaltime. Our trace analysis shows that both interactiveand bulk transfer traffic from all sites reflect alarge number of short conversations. Similarly, itshows that that a very large percentage of trafficis bidirectional, even for bulk transfer. Weobserved that interactive applications sendsignificantly different amounts of data in eachdirection of a conversation, and that interarrivaltimes for interactive applications closely follow aconstant plus exponential model. Half of theconversations are directed to a handful of networks,but the other half are directed to hundreds ofnetworks. Many of these observations contradictcommonly held beliefs regarding wide-area traffic.