BBR: Congestion-Based Congestion Control
Queue, 2016 · paper · Computer science
Argues that loss-based TCP congestion control causes bufferbloat when buffers are large and low throughput when they are small. It motivates an alternative based on understanding where network congestion originates, by measuring bottleneck bandwidth and round-trip propagation time.
0 findings
None that passed a second check.
6 minor (typos and the like; most checked by the judge only)
Wording "Network-link bandwidths span a range of 10^12—from a few bits to 100 gigabits per second" (also "12 orders of magnitude" on p.13): the stated endpoints span at most about 10^11.
- p. 32 Appendix, Startup behavior
bandwidths span a range of 10
- p. 13 Single BBR flow startup behavior
Internet link bandwidths spanning 12 orders of magnitude
Only the authors can say what is right here.
Judge's note: Checked p.32 (superscript 12 is visible) and p.13. 100 Gb/s divided by 1 b/s is 1e11, and divided by a few b/s it is about 3e10. Either the range or the endpoints need adjusting; only the authors can say which.
Second check: another model agreed
Report a problem with this finding
Wording Text says "Figure 7 shows ... BBR's throughput is consistently 2 to 25 times greater than CUBIC's", but the ratios plotted in Figure 7 run from about 2.4 to about 19 and the y-axis stops at 20.
- p. 18 Google B4 WAN Deployment Experience, last paragraph
2 to 25 times greater than CUBIC’s
- p. 19 Figure 7, y-axis 'BBR throughput / CUBIC throughput'
BBR throughput / CUBIC throughput
- p. 18 Google B4 WAN deployment, para 1
BBR’s throughput is consistently 2 to 25 times greater than CUBIC’s
- p. 25 Conclusion
improving throughput by orders of magnitude compared with
Only the authors can say what is right here.
Judge's note: Checked p.18 and Figure 7 on p.19. The highest plotted ratio is about 19 (at CUBIC ≈5 Mbps) and the lowest is about 2.4. No plotted point reaches 25. The figure may show binned values, so individual connections could reach 25x; only the authors can say where the 25 comes from. Reviewer note 3's point about 'orders of magnitude' is handled under C010.
Second check: another model agreed
Report a problem with this finding
Wording Conclusion: "BBR is deployed on Google's B4 backbone, improving throughput by orders of magnitude compared with CUBIC", while the deployment results give 2 to 25 times (Figure 7 shows about 2.4–19x).
- p. 25 Conclusion, para 2
improving throughput by orders of magnitude compared with CUBIC
- p. 18 Google B4 WAN deployment
consistently 2 to 25 times greater than CUBIC’s
Judge's note: Checked p.25 and pp.18–19. The only result of two orders of magnitude is the single US-Europe path after the receive buffer was raised by hand (133x). The fleet-wide figures are about one order of magnitude at most. This wording point also covers note 3 of C006.
Second check: another model agreed
Report a problem with this finding
Wording The example "< one loss per 30 million packets for a 1-Gbps/100-ms path" is given as the inverse square of the BDP, but with 1500-byte packets that BDP is about 8,333 packets, giving about one loss per 69 million packets.
- p. 20 Google B4 section, second paragraph
< one loss per 30 million packets for a 1-Gbps/100-ms
- p. 20 Google B4 WAN deployment, para 2
< one loss per 30 million packets for a 1-Gbps/100-ms path
Only the authors can say what is right here.
Judge's note: Checked p.20. 1 Gbps × 100 ms = 1e8 bits = 12.5 MB, or about 8,333 packets of 1500 B; squared, that is about 6.9e7. Including the Mathis constant 1.22 gives about 4.7e7. Getting 3e7 would need a BDP of about 5,477 packets, i.e. packets of about 2.3 kB. The paper does not state a packet size, so the authors should say how the 30 million was derived.
Second check: another model agreed
Report a problem with this finding
Typo The defined term "RTprop" also appears as "RTProp" (e.g. p.17 "To learn the true RTProp", p.31 "RTProp window"), so the capitalisation is inconsistent.
- p. 17 Multiple BBR flows section
To learn the true RTProp, a flow moves to the left of
- p. 31 Appendix
if the RTProp estimate
- p. 2 Definition
RTprop (round-trip propagation time)
- p. 17 Multiple BBR flows sharing a bottleneck
To learn the true RTProp
Judge's note: The p.17 crop shows 'RTprop' and 'RTProp' in consecutive paragraphs. The p.31/32 crop shows 'RTProp window'. The term is defined on p.2 as 'RTprop'. Cosmetic.
Second check: none (judge only)
Report a problem with this finding
Typo onAck pseudocode prints "app_limited_until - = packet.size" with a space inside the compound operator, which should be "-=".
- p. 9 onAck pseudocode, last line
app_limited_until - = packet.size
Judge's note: The p.9 crop shows '- =' with a visible gap in the monospace font. 'delivered += packet.size' elsewhere in the code is printed without a gap. The meaning is clear, so this is cosmetic.
Second check: none (judge only)