Design Considerations for Rate Control of Aggregated TCP Connections

We study rate control of aggregated TCP connections, i.e., multiple TCP connections treated as a single aggregate for purposes of rate control, via traffic conditioning mechanisms such as traffic policing and shaping. This is a likely scenario given the current trends in policy-based service differentiation on the Internet (e.g., Differentiated Services) and content aggregation on the Web (e.g., virtual hosting). Traffic aggregation is increasingly necessary for cost-effective, scalable provisioning and management of network and server resources. We propose and evaluate a set of mechanisms for fair sharing of an aggregate’s allocated bandwidth between connections comprising the aggregate, for traffic conditioning via policing with marking and shaping. We propose logical token buckets (common token bucket with logical partitions) that account for the round-trip times of individual connections to provide fair bandwidth sharing while achieving high aggregate throughput and bandwidth utilization. We propose modifications to TCP’s congestion window increase during congestion avoidance to achieve fairness between short and long-lived connections, and introduce the notion of aggregate fairness. We demonstrate that the proposed mechanisms provide a high degree of fairness and bandwidth utilization while limiting an aggregate’s bandwidth usage to the desired rate. We also describe the key protocol stack extensions for our AIX-based prototype implementation to enable efficient rate control of TCP connection aggregates.