Video Conferencing and Flow-Rate Fairness: A First Look at Zoom and the Impact of Flow-Queuing AQM

Congestion control is essential for the stability of the Internet and the corresponding algorithms are commonly evaluated for interoperability based on flow-rate fairness. In contrast, video conferencing software such as Zoom uses custom congestion control algorithms whose fairness behavior is mostly unknown. Aggravatingly, video conferencing has recently seen a drastic increase in use – partly caused by the COVID-19 pandemic – and could hence negatively affect how available Internet resources are shared. In this paper, we thus investigate the flow-rate fairness of video conferencing congestion control at the example of Zoom and influences of deploying AQM. We find that Zoom is slow to react to bandwidth changes and uses two to three times the bandwidth of TCP in low-bandwidth scenarios. Moreover, also when competing with delay aware congestion control such as BBR, we see high queuing delays. AQM reduces these queuing delays and can equalize the bandwidth use when used with flow-queuing. However, it then introduces high packet loss for Zoom, leaving the question how delay and loss affect Zoom’s QoE. We hence show a preliminary user study in the appendix which indicates that the QoE is at least not improved and should be studied further.

This work has been funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy – EXC-2023 Internet of Production – 390621612. We would like to thank the center for teaching- and learning services at RWTH Aachen University for issuing further Zoom Licenses. We further thank the anonymous reviewers and our shepherd Mirja Kühlewind for their valuable comments.

Authors and Affiliations

1. Communication and Distributed Systems, RWTH Aachen University, Aachen, Germany

   Constantin Sander, Ike Kunze, Klaus Wehrle & Jan Rüth

Corresponding author

Correspondence to Constantin Sander .

Editors and Affiliations

1. Brandenburg University of Technology (BTU), Cottbus, Germany

   Oliver Hohlfeld

2. Telefonica Research, Barcelona, Spain

   Andra Lutu

3. Iribe Center, University of Maryland, College Park, MD, USA

   Dave Levin

Appendix

In the following, we present results of a small-scale user study which we conducted to analyze whether our findings regarding packet loss but also improvements regarding delay have positive or negative impact on Zoom’s subjective quality. However, as our study was performed with a limited number of participants due to COVID-19 restrictions, we had to restrict the number of scenarios that we could investigate. Thus, the results and their generalizability are limited and this study should be regarded as an initial step in understanding how QoE, queuing and Zoom interact.

A QoE Impact of Flow-Queuing AQM

As we have shown in Sect. 4.4, flow-queuing AQM can achieve more equal flow-rates and reduce latency when Zoom and TCP share a bottleneck. However, this means lower bandwidths for Zoom, so likely worse video quality. In contrast, lower latencies should probably mean better interactivity. As the exact correlation w.r.t. perceived experience is hard to grasp, we perform a small-scale user study to capture the influence of flow-rate equality and AQM reduced latency on Zoom’s QoE.

Limitations of this Study. However, our study is limited, as we had to limit the number of participants (n = 10) due to COVID-19 restrictions. As such, we also restricted the number of scenarios to keep the individual study duration to roughly 25 min. Additionally, we had to switch from synthetically generated videos (noise to maximize bandwidth utilization) that we used throughout Sect. 4 to real video-conferences. This makes it difficult to compare the video-flows’ demands from our synthetic evaluation to this user study as the bandwidth demand varies with the compression rate (higher compression for actual webcam video). In summary, our study should only be regarded as an initial step.

In the following, we introduce the design and stimuli of our study and which metrics we are interested in. Subsequently, we present the results.

1.1 A.1 User Study Design

We perform a video conference where the subject interacts with an experiment assistant via Zoom focusing on interactivity and understandability to rate the quality and whether potentially degraded quality is acceptable when a concurrent download is active. The assistant reads short paragraphs of texts and the subject shortly summarizes them once the paragraph ended. This way, we test whether the video conference allowed for easy understanding but also represent the typical condition where conference attendees interrupt each other unintentionally. After 5 repetitions of summarizing, the subject and assistant alternately count to 10 to get a feeling for the delay, as proposed by the ITU [4]. Lastly, the assistant reads random numbers and the subject stops the assistant at a given number (unknown to the assistant) for the same reasons.

Quality Rating. After every run, the subject rates the overall, audio, video, and interactivity quality on a seven-point linear scale [5] (c.f., y-axis in Fig. 8). Moreover, the subject decides (yes/no) if communicating was challenging, whether the connection was acceptable at all, whether the quality was acceptable if they were downloading a file during a business or private call or when someone else was downloading documents or watching movies in parallel.

Test Conditions. We test 3 different scenarios using our previously described testbed; for all conditions, we shape the subject’s link to 0.5 Mbps, adjust the min. RTTs to 50 ms and use a queue size of 10\(\times \)BDP. The scenarios differ in whether an extra flow competes on the downlink and whether the queue is managed. In detail, in Scenario 1 (Tail-Drop) only Zoom is active using a tail-drop queue. Scenario 2 (Tail-Drop + Flow) adds a TCP CUBIC flow on the downlink, representing, e.g., a movie download. Scenario 3 (FQ_CoDel + Flow) adopts the TCP flow, but switches to the flow-queuing variant of CoDel.

Study Details. We perform a “within subject” lab study: each subject rates every test condition selected from a latin square to randomize the order. Each experiment takes about 5 min and is repeated for the 3 scenarios plus a training phase at the start using Scenario 1. In total, the 4 experiments plus rating take about 25 min. Although conducting studies with members familiar to the study is discouraged [4], we stick to the same experiment assistant to reduce variations.

Subject Recruitment. Our subjects are 10 colleagues from our institute which volunteered to take part and are strongly familiar with Zoom. We limited our study to these participants to reduce contacts during the pandemic. As such we were able to hold the conferences in the participant’s first language.

User study quality rating and votes

Full size image

1.2 A.2 Results

Figure 8a shows the mean opinion score and 95% confidence intervals of the quality rating (distributions checked for normality via a Shapiro-Wilk test). The confidence intervals are computed via the t-distribution due to our small sample size. Further, Fig. 8b shows the distributions of “Yes” (positive, to the right) and “No” (negative, to the left) answers for the different questions.

Generally looking at the plots we can see that the worst results stem from using FQ_CoDel, while using a tail-drop queue with no concurrent flow results in the best quality ratings. For the overall quality of the video conference and the video quality this difference is statistically significant as the confidence intervals do not overlap. However, for the scenarios where Zoom competes with TCP flows, the results are statistically insignificant and allow no statement. Similar, all audio quality and interactivity votes allow no statistically significant statement.

Flow-Queuing AQM Induced QoE Changes. Hence, interpreting these results is complex. What can be said is that CoDel’s positive effect of reducing the queuing delay was not perceived by the users. On the other hand, also the reduction in bandwidth did not yield any statistically significant quality reduction. However, a trend against using FQ_CoDel is visible, but it cannot be statistically reasoned. Only following the trend, it might be not worth using FQ_CoDel due to its potentially worse QoE. Otherwise, only few users considered the connection unacceptable (c.f. Fig. 8b), surprisingly uncorrelated to whether FQ_CoDel was used or whether a concurrent flow was actually started. I.e., some users considered our scenarios generally as unacceptable regardless of FQ_CoDel.

Influence of Concurrent Downloads on Acceptability. Surprisingly, users also consider the quality unacceptable when imagining a concurrent download of documents in business or private conversations. We expected that users accept deteriorations, as they would not pay attention to the video conference, but want their download to complete. However, specifically in the business case, our users did not. Also quality deteriorations induced by other users downloading movies or documents were not seen more disturbing. I.e., independent of self-inflicted or not, some users do not accept quality deteriorations at all, while others do.

Takeaway. Unfortunately, our study did not yield statistically conclusive results with respect to how participants perceive the difference in Zoom quality between using a tail-drop queue and FQ_CoDel when a flow competes. Also regarding acceptance, users did not see strong differences and either disliked the quality regardless of possible concurrent downloads as reasons or just accepted it, disagreeing on a generally applicable statement. Looking at the general trend of our study, FQ_CoDel could decrease QoE.

Reprints and permissions

© 2021 Springer Nature Switzerland AG

Policies and ethics