C o n t r o l as a B u i l d i n g B l o c k for Q o S David Harrison, Shivkumar Kalyanaraman and Sthanunathan Ramakrishnan I Dept of E C S E & Dept of CS, Rensselaer Polytechnic Institute, Troy, N Y 12180 Email: {harrisod, shiukurna}~networks.ecse.rpi.edu URL: http://www.ecse.rpi.edu/Homepages/shivkuma/ Congestion Riviera algorithm is rate-based since it is easier to m a p services in a rate-based architecture, and the control trafficis much lesser than window-based schemes. $~nd~r IbmD B; d i l l ,,..=~ .Patina r ,'.u.mB ~ ,i I ¢ rmrMii Polnl =~ Figure 2: Accumulation Measurement We realize such an algorithm using the concept of accumulation. Accumulation is the number of packets backlogged in the network and measured as shown in Figure 2. We use a thresholding scheme to detect congestion epochs. When not in congestion we increase additively over the output rate and decrease multiplicatively from the output rate otherwise. This policy causes the queue to increase slowly and to decrease quickly (within an RTT in a single bottleneck case) while maintaining high utilization. We have proved the stability of this scheme in a multi-bottleneck scenario and have also shown t h a t the scheme achieves proportional fairness. Blle,.Io-ad|e ~ l r , I ~ (trd~) Figure I: Example Model In this work, we focus on the question: H o w to provide QoS for private networks in an inter-domain multiprovider environment. Traditional QoS mechanisms like diff-serv,C S F Q require deployment of the QoS mechanisms at every potential bottleneck. Scalable performance as well as complex coordination and deployment issues have inhibited the large scale deployment of QoS mechanisms. Our solution is to use a new overlay congestion control algorithm (Riviera)to push back all congestion to the edge (e.g access router) and hence called - ~ .... _--T~.--?y.i-~. an edge-to-edge control algorithm (see Figure 1). Then | ~ I). . . . . . . QoS issues likequeueing, bandwidth sharing, packet loss distribution are all pushed to the edge allowing the core to operate using largely stateless mechanisms. Use of (L)~.ILI~ 'L~,l~pu,~'~:l+~JLqml~.-F.,Ih't (bi'lli2w/~]~'r.-.~'t congestion control as a QoS building block has two implications: Figure 3: Assured Service with Closed Loop Blocks , , r,;-;2:2T T ! I. The differentiation between flows occurs in R T T We have implemented basic services like "Better than timescales unlike in schedulers where differential best effort service", and "Isolating misbehaving flows" service is provided on a packet-by-packet basis. and advanced services like assured services (Eg: Figure 3) and quasi-leased line (QLL) by modifying the in2. The queues at the interiorbottlenecks are all consolidated and distributed at the edges. These smaller crease and decrease functions. Advanced services need queues at the edge are much easier to handle and policing at edges. In the case of a QLL we need admisbuffer management schemes m a y be used to con- sion control to prevent over-subscription. The signaltrol them more effectively (as opposed to directly ing/configuration complexity of the basic scheme (without the advanced services) is minimal. ISPs just need controlling the queue at the interior bottleneck). to provide a single class for all edge-to-edge flows. For such an architecture to be realized we need the algorithm to tranaparentl~l detect congestion, have near lossless behavior, not involve upgrades of interior routers and have minimum configuration requirements. Our tThis work was funded in part by NSF ANIR grants: ANI9806660, ANI9819112 and a grant from Intel Corp. Our scheme supports bounded scalability, incremental deployment, simpl~ed interior configuration and a simple set of overlay services. Future work includes understanding the dynamic range of services possible using such building blocks, design of control plane and management plane functions for these services etc. ACM SIGCOMM Computer Communication Review

/lp/association-for-computing-machinery/congestion-control-as-a-building-block-for-qos-Jn8ngPIye3