Modeling Wireless Links for Transport Protocols; Gurtov & Floyd

This paper describes models for wireless links. The 3 types of links considered are cellular (low bandwidth, high latency), wireless LANs (low latency), and satellite (high latency, low bandwidth). The topology given consideration is one in which only one endpoint of a path is a wireless link. The authors say that wireless links on both ends is an unlikely scenario except for the "future" possibility of VoIP being used for telephone calls, which is clearly the case 5 years later.

In considering different types of links, it's also important to keep in mind that each link has a different usage pattern. For example, cellular users will likely be involved in smaller transfers of data, and the ratio of downlink to uplink will vary across the 3 link types.

The paper points to four problems that exist in current modelling techniques. The first problem is that many models are simply unrealistic. Use of TCP over high-loss wireless links, deprecated versions of TCP (possibly due to ease of use in simulators), and retransmission constants for TCP are all representative of practices that just aren't carried out in wireless networks.

The second problem is that some models are realistic but only explore a limited parameter space. For instance, the evaluation might be done using a single flow without regard for the possibility of multiple flows, or the model might employ reverse channel allocation for ACKs without taking into account effects on performance if the ACKs are delayed. On a different extreme, some models are overly realistic in that some use bad implementations with known flaws of components for the simulation. While this practice would be fine for evaluating the badness of said component, it does not generalize well.

The last problem listed is the lack of reproducibility of some models. This problem consists of researchers not fully specifying their experimental settings, which is not so much problem of the model as of poor documentation on the researcher's part.

The paper suggests ideas about how to model a number of different tricky scenarios, including error losses/corruption, delay variation, packet reordering, bandwidth changes and assymetry, and mobility/handover implementation. Graphs showing the effects of mobility show the extent to which small changes like buffer size can affect the accuracy of a graph; in fact, none of the 3 simulations really capture what happens in the actual measurements, which almost seems to suggest that all this modeling is a waste of time. It seems like modeling is only a good idea when a researcher is trying to get a general idea of how new changes will affect a network but certainly not when trying to get an accurate picture of performance.