Congestion Control and Sizing Router Buffers in the Internet

摘要

The Internet is made of communication links andpacket switching nodes named routers. Routers are equipped with buffers that hold packets during congestion and feed output links with packets during underutilization. A rule largely known in literature is the "bandwidth-delay rule", which states that, in order to guarantee full link utilization, it is necessary to provide each link with a buffer B velence RTT(centre dot)C, where RTT is the round trip time and C is the link capacity. The bandwidth delay rule requires buffer size that increases linearly with link capacity. With the recent introduction of 10 Gbps Routers and Ethernet cards, buffer requirements as dictated by the bandwidth-delay rule become extremely large. For instance, a 10Gbps router link with a RTT of 200ms would require a 2Gbits buffer size, which is a challenging requirement for manufacturers. Moreover, such a large buffers would introduce large and time-varying queuing delays that are harmful for time sensitivity traffic such as audio and video. In this paper we investigate the relation between the TCP congestion control and the buffer size required to guarantee full link utilization. We consider two TCP congestion control algorithms: the standard TCP NewReno and the recently proposed TCP Westwood+. Analytical results show that while classic TCP Reno/NewReno requires buffer of order size approx 1/n~(1/2), where n is the number of coexisting flows, Westwood+ TCP, in principle, can provide full link utilization for any buffer size. Discrete event simulations and experiment on real 10 Gigabit per second wide area network confirm theoretical results.