An Enhanced Algorithm for Fair Traffic Conditioning in Differentiated Services Networks Abiola Adegboyega, Rupinder Makkar, Kayvan Mosharaf, Ioannis Lambadaris Broadband Networks Laboratory Carleton University Ottawa, Ontario, Canada, K1S 5B6 {agboyega, rup, mosharaf, ioannis}@sce.carleton.ca Abstract—Fair bandwidth sharing among traffic flows with different characteristics in Differentiated Services (Diffserv) networks is the focus of the current research. This paper examines and enhances an algorithm developed to enforce fairness among disparate TCP flows in the Assured Forwarding (AF) service in Diffserv. Equation Based Marking (EBM) [1] was introduced by K. Shin et al to enforce fairness in AF by monitoring existent network conditions used in marking decisions. The estimation of packet losses by the algorithm is integral to marking. The loss rates of different connections were demonstrated to converge hence enforcing a fair marking regardless of the metrics of individual flows. In this paper, EBM is analyzed for fairness and enhanced by implementing a more efficient technique for loss rate estimation. Comparison is made between EBM and the enhanced technique with results showing appreciable improvements in the maintenance of fairness. Furthermore, a service definition required by QoS standards is met with the implementation of the additional algorithm to EBM. Keywords-Diffserv, QoS, TCP, Fairness, loss rate I. INTRODUCTION Differentiated Services (Diffserv) [2] offers a scalable service provisioning architecture in IP networks by the allocation of network resources based on application requirements. Diffserv philosophy, in providing end-to-end Quality of Service (QoS) moves architectural complexity to the ‘edge’ of the Diffserv domain where ingress traffic is conditioned and marked for differentiation into classes while the ‘core’ forwards aggregated traffic streams and enforces QoS during network congestion. The Assured Forwarding (AF) Per-Hop Behavior (PHB) [3] is a Diffserv forwarding mechanism that has been standardized by the Internet Engineering Task Force (IETF). AF maintains traffic profiles for ingress traffic and marks packets at the network edge as IN-profile provided they adhere to the service profile of the network or OUT-of-profile otherwise. The basis of this marking is the comparison of the subscription rate of incoming streams with the Committed Information Rate (CIR) and Peak Information Rate (PIR) assigned for the Diffserv domain. If the ingress traffic rate is lower than the CIR, it is marked as IN and assigned the lowest drop precedence AFx1. When the traffic rate exceeds the CIR but not the PIR, packets are marked as OUT with drop precedence AFx2. Ingress traffic with rates greater than the PIR are marked as OUT with the highest drop precedence AFx3. The AF PHB can be achieved in core routers by a tiered dropping scheme that ensures service differentiation by dropping packets during congestion beginning with the highest drop precedence AFx3, then AFx2 and eventually AFx1 only when all other precedences are dropped. A common implementation of this dropping scheme is done by an active queue management (AQM) scheme like Random Early Detection (RED) [4] with some variations e.g. RED with IN and OUT (RIO), with two thresholds one each for conformant and non-conformant traffic [5]. It was shown in [6] that AF is unable to provide bandwidth assurance when competing TCP traffic streams are different based on the flow metrics of packet size, Round Trip Time (RTT), target rate and the number of microflows aggregated in a flow. Furthermore, while TCP flows are able to slow down their sending rates during congestion, non- responsive flows (e.g. UDP traffic) maintain their network subscription rates regardless of congestion. One approach of mitigating unfairness is by using ‘TCP- aware’ traffic conditioners, employing one or more of TCP’s characteristics [7][8][9]. This paper presents an enhanced version of the traffic conditioner in [1] that employs the inverse of a modeled steady-state TCP equation in marking functions. Our contributions to the existing algorithm include the implementation of a more efficient loss rate estimation technique, which is integral to packet marking. Loss History Discounting provides enhancements to EBM with more details given in section IV. The rest of the paper is organized as follows. Section II discusses related work in mitigating unfairness in AF services. Section III presents EBM [1] and its use of loss rate estimation in marking decisions. Section IV presents our version of the modified conditioner. We compare the original and enhanced algorithms with simulations in section V and conclusions are made in section VI. II. RELATED WORK Intelligent traffic conditioning for the provision of fairness in Diffserv networks has been researched extensively. In [6], Seddigh et al employed simulations to illustrate the impact of the flow metrics of RTT, packet size, microflows and target on fairness and the distribution of surplus bandwidth. In [10] simulations were employed to demonstrate that TCP/UDP interactions and RTT have a direct impact on achievable throughput of competing flows. Globecom 2004 1728 0-7803-8794-5/04/$20.00 © 2004 IEEE IEEE Communications Society