8. Core-Stateless Fair Queueing: Achieving Approximately Fair Bandwidth Allocations in High Speed Networks

There are several Fair Queueing (FQ) mechanisms in routers, such as the last paper, and Flow Random Early Drop (FRED). These congestion controls require the routers to maintain state and operate on a per-flow basis, and the authors suggest that it could be the complexity of the per packet processing that prevents from wide-range adoption.

Fair Queueing can be very effective for congestion control. But if it is implemented in every router, it could be very complicated and costly. In responde to this concern, they try to distinguish the core routers and edge routers, giving the former no per-flow state as well as FIFO packet scheduling system, while leaving the latter to do complex Fair Queueing. This is called Core-Stateless Fair Queueing (CSFQ) in this paper.

CSFQ uses a distributed algorithm that core routers do not maintain per-flow state but in stead carry the information throught a label in each packet's header generated from the edge routers.

Secondly, FIFO queueing with probabilistic dropping on input was peformed on the core routers. Fluid Model Algorithm was first demostrated for an idealized bit-by-bit version and later extend to a practical packet-by-packet version.

Simulations between CSFQ and other four additional ones, namely, FIFO, RED (Random Early Detection, drops packets long before buffer full state), FRED (Flow Random Early Drop, drops based on flow state), DRR (Deficit Round Robin, weighted FQ) were performed on single congested links, multiple congested links, and coexistence of different adaptation schemes. The result shows that CSFQ achieves relatively fair bandwidth. It's far better than FIFO or RED and comparable to FRED. But FRED quires packet flow classifcation while CSFQ does not.

The last part of this article gives the reasons why fair allocations are important and introduces punishment for malicious end-notes that uses its own TCP stack to gain more priority of the network data flow.

The traditional approach to achieve fair allocations are very complicated to implement at high speeds among all routers, since each router has to analyse current flow and packet header to reach a decision. CSFQ could be more scalable when the number of routers increases.

Comments: It is understandable to utilize CSFQ in Tier-1 peering since it could simply the processing on the routers since they are obviously the "cores". But do we consider Tier-2 and Tier-3 ISPs to be edges or cores?