Using realistic factors to simulate catastrophic congestion events in a network

With the rapid growth of the Internet, there has been increased interest in the research literature in the use of computer models to study the dynamics of communication networks. An important example of this has been the study of dramatic, but relatively infrequent, events that result in abrupt, and often catastrophic, failures in the network due to congestion. These events are sometimes known as phase transitions. With few exceptions, the models of such computer communications networks used in previous studies have been abstract graphs that include simplified representation of such important network factors as variable router speeds and packet buffer size limits. Here, we modify this typical approach, adding realistic network factors to a graph model of a single Internet Service Provider (ISP) network that can have more than a quarter million nodes. The realistic factors in our model, including router classes, variable router speeds, flows, the transmission control protocol (TCP), sources and receivers, and packet dropping, can be enabled and disabled in combinations. For each combination of realistic factors, we gradually increase network load, and gauge spread of congestion throughout the network. While there are realistic computer communications network models reported in the literature, to our knowledge none of these have been used to study catastrophic failures in computer networks. We show that the addition of realistic network factors to our model of an ISP network can mitigate catastrophic events. With the addition of variable router speeds or TCP, a phase transition to a congested state, where all routers are congested, does not appear. Yet, as load increases, ultimately the operation of the ISP network appears to decline, along with the ability of its nodes to communicate. The results of this study should be cautionary for other domains, such as electrical power grids, and the spread of viruses or diseases, where abstract graph models are often used to study phase transitions.