Academia.eduAcademia.edu

Figure 7 – uploaded by Santiago Vargas

See full PDFdownloadDownload figure
Figure 9: CDF of BBR’s bandwidth estimates under shallow buffer (100KB) and deep buffer (10MB). Under deep buffers, BBR severely overestimates bandwidth. The available bandwidth is 25Mbps shown by the black vertical line.

Figure 9 CDF of BBR’s bandwidth estimates under shallow buffer (100KB) and deep buffer (10MB). Under deep buffers, BBR severely overestimates bandwidth. The available bandwidth is 25Mbps shown by the black vertical line.

Related Figures (15)

Figure 1: The state diagram for BBR’s congestion control algorithm showing the four modes of operation.
Figure 1: The state diagram for BBR’s congestion control algorithm showing the four modes of operation.
Figure 5: Average Stall Rate across bottleneck buffer sizes for Cubic, BBR, and BBR2. Similarly to average quality (Figure 4), under deep buffers, the stall rate increases. Figure 4: Average Segment Quality across bottleneck buffer sizes for Cubic, BBR, and BBR2. As the bottleneck buffer size increases, the video quality reduces even though BBR and BBR2 are explicitly designed to work well under bufferbloat.
Figure 5: Average Stall Rate across bottleneck buffer sizes for Cubic, BBR, and BBR2. Similarly to average quality (Figure 4), under deep buffers, the stall rate increases. Figure 4: Average Segment Quality across bottleneck buffer sizes for Cubic, BBR, and BBR2. As the bottleneck buffer size increases, the video quality reduces even though BBR and BBR2 are explicitly designed to work well under bufferbloat.
Table 1: Average quality and stall rate for HTTP/1.1 video runs un- der 10ms and 100ms RTTs, shallow and deep buffers, and BBR and BBR2. QoE is low under deep buffers regardless of HTTP version or RTT.
Table 1: Average quality and stall rate for HTTP/1.1 video runs un- der 10ms and 100ms RTTs, shallow and deep buffers, and BBR and BBR2. QoE is low under deep buffers regardless of HTTP version or RTT.
Figure 6: Video QoE under deep and shallow buffers in a WAN. The client and server are located at two different continents with roughly 200ms RTT. As before, we see the video QoE in terms of quality is poor under deep buffer conditions compared to shallow buffer condition.
Figure 6: Video QoE under deep and shallow buffers in a WAN. The client and server are located at two different continents with roughly 200ms RTT. As before, we see the video QoE in terms of quality is poor under deep buffer conditions compared to shallow buffer condition.
Figure 7: Video QoE deteriorates under deep buffers across ABR al- gorithms both in terms of video quality and stall rate.
Figure 7: Video QoE deteriorates under deep buffers across ABR al- gorithms both in terms of video quality and stall rate.
Figure 8: CDF of the RTTs estimated by BBR under shallow buffer (100KB) and deep buffer (10MB). Under deep buffers, RTTs are much higher than the propagation delay (100ms), indicating that packets experience large queueing delays. This is not the case in shallow buffers.
Figure 8: CDF of the RTTs estimated by BBR under shallow buffer (100KB) and deep buffer (10MB). Under deep buffers, RTTs are much higher than the propagation delay (100ms), indicating that packets experience large queueing delays. This is not the case in shallow buffers.
Figure 10: Average segment quality is lower for runs in STARTUP than runs in PROBE_BW. The quality difference is larger for BBR2 than BBR.
Figure 10: Average segment quality is lower for runs in STARTUP than runs in PROBE_BW. The quality difference is larger for BBR2 than BBR.
Figure 11: Average stall rate is higher for STARTUP runs than PROBE_BW runs. Both BBR and BBR2 have stall rates more than 3x higher for STARTUP than PROBE_BW.
Figure 11: Average stall rate is higher for STARTUP runs than PROBE_BW runs. Both BBR and BBR2 have stall rates more than 3x higher for STARTUP than PROBE_BW.
able 2: Average percentage of video run spent in each BBR state. Runs spend more time in STARTUP for the deep 10MB buffer than the 1allow 100KB buffer. The difference is more dramatic for BBR2. Santiago Vargas”, Rebecca Drucker’, Aiswarya Renganathan, Aruna Balasubramanian, and Anshul Gandhi
able 2: Average percentage of video run spent in each BBR state. Runs spend more time in STARTUP for the deep 10MB buffer than the 1allow 100KB buffer. The difference is more dramatic for BBR2. Santiago Vargas”, Rebecca Drucker’, Aiswarya Renganathan, Aruna Balasubramanian, and Anshul Gandhi
Figure 12: Under deep buffers and large burst, BBR and ABR both operate according to the control loop shown here. BBR’s bandwidth overestimation and resulting excessive sending causes high RTTs due to queuing delay, resulting in poor throughput for the appli- cation. The application then sends lower-quality segments, which allows the router queue to drain and RTTs to return to normal. The cycle repeats when the application sends higher-quality segments at an unsustainable rate, filling the buffer again.
Figure 12: Under deep buffers and large burst, BBR and ABR both operate according to the control loop shown here. BBR’s bandwidth overestimation and resulting excessive sending causes high RTTs due to queuing delay, resulting in poor throughput for the appli- cation. The application then sends lower-quality segments, which allows the router queue to drain and RTTs to return to normal. The cycle repeats when the application sends higher-quality segments at an unsustainable rate, filling the buffer again.
Figure 13: Illustration of how RTT, segment quality, and ABR esti- mated throughput cycle in response to one another over the course of a video run under a 10MB buffer and 1MB burst. Red dashed lines on the TCP calculated RTT plot and the ABR estimated through- put plot indicate the propagation delay (100ms) and set bandwidth (25Mbps) respectively.
Figure 13: Illustration of how RTT, segment quality, and ABR esti- mated throughput cycle in response to one another over the course of a video run under a 10MB buffer and 1MB burst. Red dashed lines on the TCP calculated RTT plot and the ABR estimated through- put plot indicate the propagation delay (100ms) and set bandwidth (25Mbps) respectively.
Figure 15: Average video segment quality for BBR and BBR2 under four burst sizes. BBR’s average quality is 60% under a 1MB burst than an 8KB burst, while BBR2’s quality is 30% lower under a 1MB burst than an 8KB burst.
Figure 15: Average video segment quality for BBR and BBR2 under four burst sizes. BBR’s average quality is 60% under a 1MB burst than an 8KB burst, while BBR2’s quality is 30% lower under a 1MB burst than an 8KB burst.
Figure 14: CDF of throughput samples obtained by BBR and BBR2 to estimate bandwidth. The bandwidth is estimated based on historical samples. For BBR, the samples exceed the set bandwidth of 25 Mbps in more than 30% of the cases; for BBR2, this happens in about 20% of the cases.
Figure 14: CDF of throughput samples obtained by BBR and BBR2 to estimate bandwidth. The bandwidth is estimated based on historical samples. For BBR, the samples exceed the set bandwidth of 25 Mbps in more than 30% of the cases; for BBR2, this happens in about 20% of the cases.
Related topics:
Computer Science
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved