Scalable increase adaptive decrease: congestion control supporting low latency and high speed.

Congestion control in the Internet has been an open research issue for more than two decades. A large number of proposals already exist that especially address the scalability problem of traditional congestion control in high-speed networks. However, more and more applications with narrow latency requirements are emerging which are not well addressed by existing proposals. In this work, we present TCP SIAD, a new congestion control scheme supporting both high-speed networks and low latency, based on a new design principle called Scalable Increase Adaptive Decrease (SIAD). More precisely, our algorithm aims to provide high utilization under various network conditions, and therefore allows operators to configure small buffers for low latency support. We designed TCP SIAD based on a new approach that aims for a fixed feedback rate independent of the available bandwidth and provides full scalability. Further, our approach introduces a configuration knob for the induced feedback rate and thereby controls the aggressiveness. This can be used by a higher-layer control loop to impact capacity sharing, e.g., for applications that need a minimum rate. Increasing the aggressiveness can lead to higher throughput but also induces more congestion experienced by the transmission as well as competing traffic. However, we argue that fairness cannot be addressed by congestion control, that only influences the instantaneous share for each flow, and therefore do not aim for TCP-friendliness. Instead, fairness must be policed on a per-user basis over longer time scales supported by a configurable aggressiveness of the used congestion control scheme as provided by TCP SIAD. We evaluate TCP SIAD’s scalability, adaptivity, capacity sharing, and convergence properties against well-known high-speed congestion control schemes, such as Scalable TCP and High Speed TCP, as well as H-TCP that among other goals targets small buffers. We show that only TCP SIAD is able to utilize the bottleneck with arbitrary buffer sizes while avoiding to unnecessarily buffer packets in network queues. Further, only TCP SIAD and Scalable TCP implement a fixed feedback rate independent of the link speed, where Scalable TCP reaches this on the cost of inducing a standing queue and high loss rate. Moreover, we demonstrate the capacity sharing properties of SIAD depending on the configured feedback rate. Further, due to a new Fast Increase phase TCP SIAD can quickly allocate newly available capacity and converges reasonably fast. In addition, TCP SIAD provides a much higher resilience to non-congestion losses than all other schemes in test. We conclude that TCP SIAD fulfills the stated requirements and shows high robustness to perform further testing in the Internet.