Traffic Modeling and Network Simulation

Overview

We are developing a closed-loop aggregate IP traffic simulation system using TCP congestion control. Our goal is to make the synthetic output traffic stochastically similar to that from the actual live wire of an Internet link.

Statistical Modeling of Traffic

In our system, a TCP simulator is driven by stochastic inputs: requests for transfers of information from one host to another and propagation delay and losses for a cloud of hosts that feed requests to the network that is simulated. The inputs are generated by statistical models that are developed based on mathematical theory and on empirical studies of packet header data collected on Internet wires. For example, for HTTP transfers, we model several time processes: start times, request file sizes from the client, downloaded file sizes from the server.

Our statistical models of traffic incorporate the long-range dependence and the nonstationarity that is pervasive in Internet traffic. The former is well known and much studied. The latter, however, has received much less attention. Traffic on Internet wires is a superposition of traffic sources. The cause of the nonstationarity is a changing number of superposed traffic sources. As the number changes, the statistical properties change. And the change is far more profund that just simly an increase in rate as the number of sources increases. Marginal distributions and autocorrelation change as well.

The Simulator

The output of the simulator is packet traffic that interacts with a network environment. The simulated TCP, which might spawn hundreds or thousands of TCP flows, responds to network congestion signaled via feedback packets. To mimic the behavior of TCP, we take the source code of an actual TCP implementation from the BSD kernel. Unlike the open-loop approaches which directly control the packet arrival times, packets are generated by TCP based on the processing of returning acknowledgment packets. The simulated network environment models propagation delays, link capacities, switching, routing, and congestion control inside the network. Packets traverse the simulated network and get dropped or arrive at the end host. This end host is also modeled by the TCP simulator.

Traffic Modeling: Status

We have developed a fundamental approach to modeling traffic that rests on a the notion of connection-rate superpostion, which allows us to specify models based both on the theory of superposed stocastic processes and on empirical study of Internet measurments. See the papers xxx for a discussion. We have completed the modeling of start times for HTTP, In progress is the modeling of HTTP request and downloaded file sizes and round trip times at the start times of each HTTP connection. Using tools and appoaches developed in the HTTP modeling, we are building models for SMTP inputs. After HTTP and SMTP we move to other major applications running under TCP and then on to UDP.

Simulation: Status

We are about to run our first simulation with just HTTP traffic and for a single link where the output will be synthetic traffic for the real traffic of MHWire1. Due data is 7/31/00.
Home   Papers   Data Collection   Visualization   S-Net   Helios   Network Tomography   Multicast   Web Performance