RED in a Different Light

Packet networks require queues (buffers) to absorb short term arrival rate fluctuations. Yet network implementors have always observed that queues at bottlenecks tend to fill and stay filled, which contributes extra delay and removes the ability to absorb bursts. In [1] Floyd and Jacobson proposed the RED (Random Early Detection) active queue management algorithm. RED is simple, robust and quite effective at reducing persistent queues. However, while it has been used widely and successfully on Internet routers, [1] offers little guidance on how to set configuration parameters and RED has gained the reputation of being very difficult to tune. This paper develops RED in different way, treating it as a servo control loop and deriving all the loop parameters from measurable properties of a router. The result is a ‘self-tuning’ RED whose parameters are completely determined by the queue output bandwidth (average departure rate). This new RED performs substantially better than the original version and works for a much wider variety of traffic and link bandwidths. It also admits a substantially simpler and more efficient implementation, one particularly well suited for ASIC forwarding engines. Please note: This is an early draft of an in-progress paper. Several important sections are still missing and the simulation data needs to be reorganized so that the story it tells is clearer. 1.0 Introduction Although queues and buffers are essential to the efficient operation of packet networks, queues at a bottleneck tend to fill up and remain full, adding unnecessary delay to traffic and losing the ability to absorb bursts. To address this problem, the IRTF has urged the deployment of active queue management in the Internet [2]. RED (Random Early Detection) [1] is one of the leading active queue management candidates. The paper defining RED was first published in 1993 and there have since been a number of implementations, variations, imitations, and reports on its use [4][5]. The 1993 paper describes active queue management through randomly dropping (or marking)1 arriving packets when the average queue length indicates congestion and makes a case for the efficacy of such an approach. Briefly, RED works by keeping a run1. [1] suggests that packets be “marked” by a queue management algorithm. Dropping a packet is the most severe form of “marking” or indicating to an end-system that congestion is present. Although other types of marking may be used [ECN], we have focused here on dropping and will use that terminology.