Sizing Network Buffers: An HTTP Adaptive Streaming Perspective

摘要

HTTP Adaptive video Streaming (HAS) is the dominant traffic type on the Internet. When multiple video clients share a bottleneck link many problems arise, notably bandwidth underutilisation, unfairness and instability. Key findings from previous papers show that the "ON-OFF" behaviour of adaptive video clients is the main culprit. In this paper we focus on the network, and specifically the effects of network queue size when multiple video clients share network resources. We conducted experiments using the Mininet virtual network environment streaming real video content to open-source GPAC video clients. We explored how different network buffer sizes, ranging from 1xBDP to 30xBDP (bandwidth-delay-product), affect clients sharing a bottleneck link. Within GPAC, we implemented the published state-of-the-art adaptive video algorithms FESTIVE and BBA-2. We also evaluated impact of web cross-traffic. Our main findings indicate that the "rule-of-thumb" 1xBDP for network buffer sizing causes bandwidth underutilisation, limiting available bandwidth to 70% for all video clients across different round-trip-times (RTT). Interaction between web and HAS clients depends on multiple factors, including adaptation algorithm, bitrate distribution and offered web traffic load. Additionally, operating in an environment with heterogeneous RTTs causes unfairness among ompeting HAS clients. Based on our experimental results, we propose 2xBDP as a default network queue size in environments when multiple users share network resources with homogeneous RTTs. With heterogeneous RTTs, a BDP value based on the average RTTs for all clients improves fairness among competing clients by 60%.