Summary: The paper introduces a new Fair Queuing scheme that achieves fairness without compromising the per packet
processing time (work required is O(1) per packet). The Fair Queuing algorithm that was proposed earlier required O
(log n) per packet and implementing it on hardware was also expensive. Similarly fairness based on pure round robin
scheduling suffers from the fact that packet lengths are not considered and approaches based on topology perform
DRR Fair Queuing derives ideas from Fair Queuing and Stochastic FQ. It uses hashing to determine the queue to which
a flow has to be assigned and collisions automatically reduce the bandwidth guaranteed to the flow. Each queue is
assigned a quantum and can send a packet of size that can fit in the available quantum. If not, the unused quantum gets
added to this particular queue's deficit and the packet can be sent in the next round. The quantum size is a very critical
parameter in the DRR scheme, determining the upper bound on the latency as well as well as the throughput. The
quantum size is also central to the O(1) assurance provided by the algorithm. Only when all quantum sizes are above the
maximum packet size, can this be guaranteed. The paper also formalizes two metrics - Fairness Index (measures the
fairness of the queuing discipline) and Work Quotient (the effort required to process each packet), and these metrics are
used to compare DRR with other queuing schemes.
· The paper uses a very simple yet elegant idea to achieve superior performance with respect to throughput fairness, and
this algorithm can also be implemented in gateways in an efficient and cost effective manner.
· Through Deficit Round Robin, they provide a generic framework to implement fair queuing efficiently, and they dont
advocate broad policies that would suit environments. They are careful to show the mechanism and leave the policy
decisions to the implementor.