A Closed-loop Scheme for Expected Minimum Rate and Weighted Rate Services
David Harrison, Yong Qing Xia, Shivkumar Kalyanaraman, Arvind Venkatesan · 2002
Traditionally QoS capabilities have been constructed out of open-loop building blocks such as packet schedulers and traffic conditioners. In this paper, we consider closed-loop techniques to achieve a range of service differentiation capabilities. Our key contribution is the use of Accumulation-based Congestion Control (ACC) as a data-plane building block to provide an expected minimum rate service which is similar to Frame Relay CIR/PIR and DiffServ assured service. A user with a minimum rate expectation can be interpreted as having a non-concave user utility function. Unfortunately in the context of nonlinear optimization, non-concave utility functions are analytically difficult and often result in multiple optimal solutions. Instead of attempting optimization with non-concave objective functions, we demonstrate a meaningful notion of an expected minimum rate by imposing additional constraints on Kelly’s convex optimization formulated in [14]. Unlike the constraint which simply requires all user allocations to be larger than their respective expected minima, our constraint does not require admission control. The resulting scheme is distributed, requires each control loop to act only on local knowledge and still allows policy-based control over how capacity is allocated during oversubscription. We use ns-2 simulations and Linux implementation experiments to demonstrate that the service performance matches theoretical results. Our scheme does not require Active Queue Management (AQM) at bottlenecks. However, with AQM, we achieve near zero queue with high utilization.