Analyzing the Impact of Bufferbloat on Low-Latency Internet Traffic
TL;DR
Findings show that uncontrolled buffers have a significant negative effect on the performance of low-latency traffic in congestion but that AQM mechanisms can be carefully tuned to achieve a substantial reduction in queueing delay without seriously impacting throughput.
Abstract
Bufferbloat has become a severe degradation to performance in contemporary IP networks, especially to low-latency Internet traffic that has strong delay and jitter requirements. Over buffering in network equipments which initially is supposed to eliminate the loss of packets and enhance the throughput, tend to create continuously long queues, and consequently cause swelling end-to end latency and subpar Quality-of-Experience (QoE). In this paper, the author examines the effects of bufferbloat on the application performance of cars relying on latency sensitive applications: Voice over IP (VoIP), online gaming, augmented and virtual reality (AR/VR) and real-time teleoperation. We initial discuss the underlying causes of bufferbloat, namely buffer size, conservative drop policy and the interaction between transport layer congestion control and transport layer congestion control. The paper subsequently defines low-latency traffic needs and points out that it is susceptible to queueing and jitter delays, as well as bursty packet loss. A controlled testbed based on access-network conditions with mixed traffic loads is used as an experimental evaluation. The drop-tail buffering is contrasted with Active Queue Management (AQM) methods to measure their impacts in terms of latency, throughput, fairness and coexistence with the legacy TCP flows. Findings show that uncontrolled buffers have a significant negative effect on the performance of low-latency traffic in congestion but that AQM mechanisms can be carefully tuned to achieve a substantial reduction in queueing delay without seriously impacting throughput. Lastly, the paper explains the open issues regarding bufferbloat in wireless networks and mobile networks, new applications based on ultra-low-latency, and the correct measurement and detection of operational networks. The results offer empirical evidence to the network designers and researchers whose goal is to develop latency-sensitive Internet infrastructures.