Source models of network game traffic

We study the traffic generated by sessions of a popular multi-player network game. Our analysis indicates that empirical game traffic can be characterized well by certain analytical models. While clients and servers as well as hosts with different configurations, produce different models, all models from the game are well modeled by the same families of distributions. We find that some data sets are best modeled with split distributions; that is, one portion of the data is well modeled with one particular distribution, and the rest of the data with another. We describe how our models can be simulated and discuss how host processing speed influences packet interarrival processes, which in turn influence playability. The latter is a clear indication that end user quality of service is more than just a network issue--host characteristics must be considered as well. These are empirical results that have been rarely incorporated into theoretical analyses of network traffic. As Internet gaming becomes more popular, we expect that our models will be useful for testing hardware and protocols that support gaming. (C) 2000 Elsevier Science B.V. All rights reserved.