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

This paper addresses the drawbacks of the Fair Queuing (FQ) algorithm proposed in the previous paper: the need for state maintenance, buffer management and per-flow packet scheduling. Instead, the authors propose a simpler architecture comprised of edge routers and core routers. Edge routers maintain per-flow state and insert flow rate information into packet headers; core routers use the flow rate information to implement a probabilistic dropping algorithm. The authors analyze the performance of this Core-Stateless FQ scheme and present simulations to illustrate the scheme. The goal here is to achieve reasonable approximations to fair bandwidth allocation rather than matching the FQ scheme perfectly. The packet dropping algorithm is carefully explained using a fluid model, followed by a discussion of rate estimation at the edge routers and determining the fair share. One impressive caveat of this algorithm is the weighted CSFQ mechanism, which enables the use of differential services such as the file server pair that was problematic with the FQ scheme of the previous paper. The authors also assure that each operation that the CSFQ scheme requires can be efficiently implemented. Finally, CSFQ is compared to other queueing algorithms by simulation and evaluation of the results. The authors also explore the effect of high latency in an "all-core" design. Later, the identification approach is introduced as an alternative method of achieving gateway-based congestion control.

I really enjoyed this paper as the concept of separation into edge routers and core routers seems quite innovative. The authors did an excellent job of identifying the areas where this scheme may or may not thrive, rather than presenting a purely positive assessment. This should definitely remain in the syllabus.