Equation-Based Congestion Control for Unicast
Applications
Sally Floyd, Mark Handley, Jitendra Padhye, Jörg Widmer (SIGCOMM 2001)
Summary:
When packets arrive at a router faster than they can be placed on the appropriate outgoing link, they must be stored in a router buffer. Should this condition persist, the router buffer can become full, at which point any additional packets requiring storage must be dropped. When packet losses for this reason become sufficiently frequent, congestion is said to occur on the network. To address this problem, TCP makes use of Additive Increase/Multiplicative Decrease (AIMD), a congestion control mechanism in which the sending rate is limited by a congestion window (i.e. an upper-bound on the number of unacknowledged packets) which is halved for every window of data containing a packet drop and increased by roughly one packet per window of data otherwise. While AIMD is sufficient to avoid the problem of congestion, the unnecessary severity of the multiplicative decrease factor of ½ results in an unstable sending rate that may be unsuitable for some real-time applications (e.g. streaming video).
This paper introduces the TCP-Friendly Rate Control (TFRC) protocol, a best-effort alternative to TCP that although slower in claiming available bandwidth, achieves a more stable sending rate while still being responsive to congestion over longer time periods. TFRC makes use of a technique called equation-based congestion control, in place of AIMD. In this approach, the sending rate is limited by a control equation that expresses the maximum acceptable sending rate T as a function of the loss event rate, where a loss event consists of a number of packets dropped within a round-trip time. In the case of TFRC, the control equation used is the TCP Response Function, which characterizes the steady-state sending rate of TCP as a function of the steady-state loss event rate p and the round-trip time RTT; use of this function ensures that TFRC will not starve, or be starved by, TCP when deployed on the Internet. By carefully calculating the parameters p and RTT, TFRC ensures the smoothness of T, while preserving responsiveness to congestion. Through appeals to extensive tests run over the Dummynet network emulator, the ns network simulator and the public Internet, this work provides strong evidence that TFRC behaves as expected.