 
Summary: Steven Kappes
CoreStateless Fair Queueing: Achieving Approximately Fair Bandwidth
Allocations in High Speed Networks
This paper presents another method of fair queueing by implementing maxmin
fairness. The network is divided into edge routers, which have fewer flows, and core
backbone routers, which need to be able to operate at higher speeds. The edge
networks calculate the rate of each flow. Every packet for a flow is then labeled with its
rate. The core routers will then approximate maxmin fairness by estimating the fair
share rate of each flow. If there is congestion, packets will be probabilistically dropped
to ensure flows operating above their fair share get scaled back.
The benefit of separating these operations is that the complexity is reduced at core
routers. Core routers will not need to keep any perflow state, only averaging the fair
share rate, the aggregate arrival rate, the accepted traffic rate. This allows the
complexity per packet to be only O(1) at core routers, while the edge routers should be
able to handle classifying each packet into a flow. Simulations show that this
approximation performs much better than non fair queueing algorithms such as FCFS
and RED. CSFQ even outperforms deficit roundrobin when there is a large number of
flows.
Implementing such a scheme would certainly be more difficult than any of the other
queueing methods, however. The network would need to be explicitly structured
