A methodology for the optimal configuration of TCP traffic in network simulations under link load constraints

Given that over 90% of the Internet load is carried by TCP, most network simulation studies use TCP flows to generate the background traffic. A basic, but unresolved, question however is: how can one decide how many TCP flows to simulate from one network node to another? Simulating too many flows on a link can cause an unrealistically high loss rate on that link, while simulating too few flows can result in undesirably light load conditions. Similarly, to simulate realistic network conditions, one has to carefully control the load distribution on various network links (e.g., edge vs. core links), as well as the hop count (path length) of the simulated TCP flows. Previous simulation studies have dealt with these issues in a trial-and-error manner, experimenting with several traffic configurations until a realistic distribution of link load and loss rate is achieved. In this paper, we present a methodology that determines the number of TCP flows that should be simulated between each pair of nodes in a network, based on the network topology, a specification of the utilization and loss rate for certain links, and an average number of hops for the TCP flows. Our methodology is based on a linear program formulation that, while meeting the utilization and loss rate specifications, minimizes the number of required TCP flows. This optimization criterion minimizes the memory requirement of the simulation. Our evaluations show that the proposed methodology can closely approximate the specified link conditions in terms of utilization and loss rate. We also analyze the largest approximation errors and reveal their causes.