Extend of my previous work. Consider stochastically TCP-friendliness that defined as:

$X^\textrm{TCP} \le_U X^\pi \textrm{, if } E[u(X^\textrm{TCP})]\le E[u(X^\pi)] \forall u\in U$

where $$\pi$$ denotes the control for UDP traffic and $$X$$ are the rates of a set of sessions; $$U$$ is a set of utility function and the comparison $$\le$$ is component-wise to the vector $$X$$.

If $$U$$ is defined as the set of all increasing functions ($$U=st$$), then the stochastic ordering $$X^\textrm{TCP} \le_{st} X^\pi$$ is the usual stochastic order, which means

$X^\textrm{TCP} \le_{st} X^\pi \Leftrightarrow Pr[X^\textrm{TCP} \le x] \ge Pr[X^\pi \le x]$

The protocol in this paper, TCP-Friendly CBR-Like Rate Control or TFCBR, is not to apply admission control but to smooth the rate of sending to a longer time scale, say, minutes. There are two parameters in this scheme, namely, $$\alpha$$ the fraction of TCP rate to send by the TFCBR; and $$\beta$$ the time constant. The protocol checks TCP rate for previous $$\beta$$ seconds using equation and adapt to the $$\alpha$$ of that rate. Thus it is like a moving average and it claims that, by appropriately adjusting the parameters, it can be stochastically TCP-friendly.

The paper also points out, $$\alpha$$ cannot be greater than 1 or otherwise it is not TCP-friendly at all.

## Bibliographic data

@inproceedings{
title = "TCP-Friendly CBR-Like Rate Control",
author = "Feng and Xu",
booktitle = "Proc ICNP",
year = "2008",
}