This paper tackles an interesting research problem: it is well-known that traditional transport protocols (e.g. TCP) don’t perform well over wireless links (e.g. because they conflate packet losses with congestion; I reviewed an earlier paper that evaluates TCP improvements for wireless links). But in order for transport protocol designers to take wireless network performance into account, they need simple models of wireless performance that are nonetheless accurate over the wide array of different wireless variants and environments. This was reflected in the MACAW paper, for example: the proposals made by the paper are very dependent on their assumptions about the behavior of the wireless link layer, and it is hard to know how well those assumptions will generalize.
Modeling is something of an art: a good model should balance realism, generality, and detail. There are four basic parameters to a wireless model, from the perspective of the transport protocol designer:
- Type of link: Cellular, wireless LAN, or satellite.
- Network topology: where does the wireless link(s) appear in the network topology?
- Traffic model: for example, is traffic bidirectional and how large is a typical transfer size?
- Performance metrics: what should the designer optimize for. For example, throughput, goodput, fairness, or delay.
Additionally, wireless links possess some additional characteristics that are typically reflected in models:
- Error losses and packet corruption. This should be modeled by having the link drop packets according to some random distribution (e.g. a uniform distribution over packets or bits).
- Variation in delay. Modeled by temporarily suspending data transmission over the simulated link.
- Packet reordering. This can be modeled by swapping the position of two packets in a queue, but this has the disadvantage that there must be a non-zero queue length for reordering to occur. Perhaps better is to simulate reordering by imposing delay on one packet but not others.
- On-demand resource allocation. For example, in GPRS, radio channels are released when the queue size falls below a certain threshold; the channel must be allocated when new data arrives, introducing delay. This can be modeled by introducing an additional delay into the model when data arrives and the channel has been idle for longer than a configurable period.
- Bandwidth variation. This can be modeled by simply changing the bandwidth of the link.
- Asymmetry in bandwidth and latency. This can be directly reflected in the model: uplink and downlink bandwidths and latencies can be allowed to differ.
- Queue management. To model current cellular and wireless LAN networks, the authors suggest using a “Drop-Tail” policy. To model satellite or “future cellular and wireless LAN networks”, RED may be more appropriate.
Modeling mobility depends on the goals of the researcher. For those concerned only with the effect that intersystem handoff has on transport protocols, it suffices to change the link characteristics in the network model to effect a handoff. If the researcher wants to model handoffs in more detail, a more exhaustive model is required (accounting for topology, home and foreign agents, Mobile IP, etc.)