Watching Stars in Pixels: The Interplay Of Traffic Shaping and YouTube Streaming QoE over GEO Satellite Networks

Abstract

Geosynchronous satellite (GEO) networks are an important Internet access option for users beyond terrestrial connectivity. However, unlike terrestrial networks, GEO networks exhibit high latency and deploy TCP proxies and traffic shapers. The deployment of proxies effectively mitigates the impact of high network latency in GEO networks, while traffic shapers help realize customer-controlled data-saver options that optimize data usage. However, it is unclear how the interplay between GEO networks’ high latency, TCP proxies, and traffic-shaping policies affects the quality of experience for commonly used video applications. To address this gap, we analyze the quality of over 2 k YouTube video sessions streamed across a production GEO network with a 900 Kbps shaping rate. Given the average bit rates of the videos, we expected streaming to be seamless at resolutions of 360p, and nearly seamless at resolutions approaching 480p. However, our analysis reveals that this is not the case: $30\%$ of both TCP and QUIC sessions experience rebuffering, while the median average resolution is only 404p for TCP and 360p for QUIC. Our analysis identifies two key factors that contribute to sub-optimal performance: (i) unlike TCP, QUIC only utilizes $70\%$ of the network capacity; and (ii) YouTube’s chunk request pipelining neglects network latency, resulting in idle periods that disproportionately harm the throughput of smaller chunks. As a result of our study, Viasat discontinued support for the low-bandwidth data-saving option in U.S. business and residential markets to avoid potential degradation of video quality—highlighting the practical significance of our findings.

Udit Paul was a PhD student at the University of California, Santa Barbara when the work was performed.

A Appendix

1.1 A.1 Ethical Considerations

Although our work involves HTTP log analysis on an operational GEO satellite network, our work is not human subjects research. At no point is any data collected from the customers of the network. We collect and analyze only our own experimentally generated traffic.

Fig. 12.

Idle time of campus network.

Fig. 13.

Post 30 s throughput without considering idle time. Vertical line is 900 Kbps.

Fig. 14.

Chunk size vs throughput over TCP.

Fig. 15.

Achieved throughput. Vertical line is 900 Kbps.

Fig. 16.

Time to first byte of each chunk. Vertical line is 600 ms.

1.2 A.2 Supplementary Results

In this section we include some additional, supplementary graphs that were briefly described in the main body of the paper. The median idle time for both TCP and QUIC was short in our campus network experiment, around 15 ms, as shown in Fig. 12. This result suggests that the pipelining inefficiency is magnified by the high round trip time of the GEO satellite network. Figure 13 shows the $T_{network}$ after 30 seconds of playback, in order to eliminate any effect due to slow start. The figure indicates that QUIC throughput still varies well below the shaped bandwidth 900 kbps. This indicates that congestion control, and specifically the initial slow start, are not the source of the low throughput. Figure 16 shows the TTFB of each chunk. We can observe that almost all chunks have a TTFB larger than 600 ms; QUIC in particular forms a cluster close to 600 ms. The correlation between achieved throughput ($T_{idle}$) and chunk size for TCP is illustrated in Fig. 14. The Pearson statistic for correlation of achieved throughput and log(chunk size) is 0.62. Finally, Fig. 15 shows that the $T_{idle}$ of TCP outperforms that of QUIC in GEO networks; the median TCP throughput is 0.58 Mbps, while QUIC’s median throughput is 0.47 Mbps. Importantly, however, neither reach the shaped bandwidth rate. Figure 16 shows the TTFB of each chunk. The median chunk TTFB for TCP is 1.21 s, while it is 0.78 s for QUIC.