Characterizing Netflix bandwidth consumption

The widespread deployment and adoption of the Dynamic Adaptive Streaming over HTTP (DASH) standard is making Internet video-on-demand a `standard´ Internet application similar in impact as email and web browsing. While video streaming has been widely deployed and studied for decades, DASH-based streaming is very different as it involves adaptation both by the application and by TCP. The dynamics and implications of multiple levels of end-to-end congestion control are not well understood. The contribution of the research presented in this paper is twofold: first, we characterize the bandwidth consumption of a widely deployed DASH application (i.e., Netflix); second, we provide insight in how different implementations and different access networks can impact bandwidth consumption. Our results suggest that Netflix adaptation defaults to underlying TCP mechanisms during periods of heavy, sustained network congestion. However, the application algorithm is clearly intertwined with the underlying TCP mechanisms during periods of volatile network conditions. In one network scenario, we observed that a backlogged TCP flow achieved a throughput of 6 Mbps while a Netflix session (under similar path conditions) consumed less than 3 Mbps of bandwidth.