Higher order delay functions for delay-loss based TCP congestion control

摘要

TCP-Illinois aims to address TCP's low throughput when operating in high-speed, high-delay networks. Previous research has shown that, due to its linear increase behaviour and to its relatively long congestion epochs, TCP-Illinois exhibits suboptimal scaling behaviour with an increasing path Bandwidth- Delay Product (BDP). This paper discloses our contributions towards improving the aggressiveness and responsiveness of loss-based TCP congestion control algorithms. We formally show that higher order versions (of power n) of the delay functions used by TCP-Illinois become more aggressive and responsive with an increasing value of n. Based on this finding, we propose three variants: i) a second order (quadratic) version of additive increase and multiplicative decrease (TCP-Q), ii) a second order multiplicative decrease only (TCP-Fq) and iii) a sub-linear multiplicative decrease only (TCP-Fs). By modifying the TCP-Illinois code in the GNU/Linux kernel, we obtained the three corresponding modules and used them for our simulations using the TCP/Linux patch for ns2. Based on standardised congestion control metrics, we conducted a comparative analysis between our variants and a number of relevant high speed TCP algorithms. Simulation results agree with our analytical findings; compared to TCP-Illinois, TCP-Q exhibits shorter congestion epochs and thus better responsiveness and convergence.