On the Long-Run Behavior of
Equation-Based Rate Control. M. Vojnovic
and J. Y. Le Boudec
(EPFL)
We consider unicast equation based rate control, where a source estimates the
loss event ratio p, and, primarily at loss events, adjusts its sending rate
to f(p). Function f is assumed to represent the loss-throughput relation that
TCP would experience. When no loss occurs, the rate may also be increased
according to some additional mechanism. We assume that the loss event
interval estimator is non-biased. If the loss process is deterministic, the
control is TCP-friendly in the long run, i.e, the average throughput does not
exceed that of TCP. If, in contrast, losses are random, it is not a priori
clear whether this holds, due to the non-linearity of f, and a phenomenon
similar to Feller`s paradox. Our goal is to identify the key factors that
drive whether, and how far, the control is TCP friendly (in the long run). As
TCP and our source may experience different loss event intervals, we
distinguish between TCP-friendly and conservative (throughput does not exceed
f(p)). We give a representation of the long term throughput, and derive that
conservativeness is primarily influenced by various convexity properties of
f, the variability of loss events, and the correlation structure of the loss
process. In many cases, these factors lead to conservativeness, but we show
reasonable lab experiments where the control is clearly non-conservative.
However, our analysis also suggests that our source should experience a
higher loss event ratio than TCP, which would make non-TCP friendliness less
likely. Our findings provide guidelines that help understand when an equation
base control is indeed TCP-friendly in the long run, and in some cases,
excessively so. The effect of round trip time and its variation is not
included in this study.
Papers are provided as
a service to all by the members of ACM SIGCOMM.
This
paper is available in Adobe PDF format.
|