Observations on Equation-Based Rate Control
We consider unicast equation based rate control defined as follows. A source adjusts its rate primarily at loss events to f(p) where p is an estimator of loss event ratio. Function f is typically the loss-throughput formula of a TCP source. In absence of loss events, the send rate can be also increased by an additional mechanism. We suppose the estimator is unbiased estimator of the loss event interval (amount of data sent between two consecutive loss events). Indeed, for the loss event intervals fixed to some value, the throughput x satisfies x=f(p). However, if loss process is random, it is not clear how the throughput would relate to f(p). If x<=f(p), we say the control is conservative. We derive a representation of the throughput, and obtain that conservativeness is primarily due to various convexity properties of function f, and variability and correlation structure of the loss process. In many cases, these factors drive the control to be conservative, but we also show some reasonable cases of non-conservative control. However, having observed that our source should experience larger long-run loss event ratio than TCP would, non TCP-friendliness becomes less likely. In our study we do not consider the effects involved due to randomness of the round-trip time.
Record created on 2004-08-31, modified on 2016-08-08