Stochastic Fairness Queuing

Fairness queuing has recently been proposed as an e ective way to insulate users of large computer communication datagram networks from congestion caused by the activities of other (possibly ill-behaved) users. Unfortunately, fair queuing as proposed by Shenker et al. [DKS89] requires that each conversationz be mapped into its own queue. While there are many known methods of implementing this type of mapping, they are relatively slow, requiring numerous memory references, and thus do not lend themselves to a software or rmware implementation capable of operating in high-speed networks. This paper presents a class of algorithms collectively called \stochastic fairness queuing" that are probabilistic variants of fair queuing. These algorithms do not require an exact mapping, and thus are suitable for high-speed software or rmware implementation. Furthermore, these algorithms span a broad range of CPU, memory, and fairness tradeo s. zA \conversation" consists of all packets with a given source-destination address pair. Input Packets Distribute By Address Pair ---FCFS Queue FCFS Queue FCFS Queue sss FCFS Queue ----? Output Packets Figure 1: Fairness Queue