A Model of TCP/IP Suitable for Parallel Simulation of Large Internets

Abstract

The current Internet connects on the order of 105 hosts. What unanticipated network dynamics will arise as the Internet grows to a million or more nodes? This report describes a model of TCP/IP based on extended finite state machines intended to permit simulation on this scale. The model yields the throughput and delay averaged over all connections of a workload consisting of client and server hosts interconnected through a subnet of gateways. A variety of parameters characterize the workload, protocol software speed, subnet, and internet topology. The model presented embodies a number of conjectures on what aspects of a protocol should be omitted to permit simulation of large numbers of nodes, and conjectures on what protocol aspects should be included to give qualitative, rather than quantitative, insight into scaling the Internet.