Minimizing queue variance using randomized deterministic marking

Na Li, Gustavo de Veciana, Sangkyu Park, Marissa Borrego, San-qi Li

Abstract: Previous work on congestion control in TCP/IP networks combines improved end-user transmission mechanisms with active queue management schemes at network routers. An active queue management scheme consists of two stages. In order to stabilize the queues at a router, one must first determine an appropriate packet marking probability given the current degree of congestion. Second, in order to realize the desired marking probability, an effective packet marking algorithm needs to be implemented to decide which packets should be marked. Researchers have increasingly focused on the first stage, namely determining the fraction of packets to mark, overlooking the fact that for a given marking probability, various possible marking algorithms result in different queue variance, and thus loss, delay, and jitter. We propose a marking algorithm DREAM. DREAM decouples the functions of reducing queue variance and randomizing the phases of flows. Compared to existing schemes, it significantly reduces queue variance while avoiding flow synchronization. Based on a simple Markov chain model we explain why our scheme is superior. Our simulation results confirm its effectiveness. Furthermore, DREAM is simple to implement and has a much lower overhead as compared with existing mechanisms.