| Internet-Draft | BBR-Realtime | July 2024 |
| Huitema, et al. | Expires 5 January 2025 | [Page] |
We are studying the transmission of real-time Media over QUIC. There are two priorities: maintain low transmission delays by avoiding building queues, and use the available network capacity to provide the best possible experience. We found through experiments that while the BBR algorithm generally allow us to correctly and timely assess network capacity, we still see "glitches" in specific conditions, in particular when using wireless networks. We analyze these issues and propose small changes in the BBR algorithm that could improve the quality of experience delivered by the real-time media application.¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
This Internet-Draft will expire on 5 January 2025.¶
Copyright (c) 2024 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
When designing real-time communication applications over QUIC, we want to maintaining low transmission delays but provide the best possible experience that the current transmission capacity allows. We elected to use the version 3 of the BBR congestion control algorithm (see [I-D.cardwell-iccrg-bbr-congestion-control] and [BBRv3-Slides]). We found through experiments that while the BBR algorithm generally allow us to correctly and timely assess network capacity, we still see "glitches" in specific conditions, during which the BBR algorithm either allows the building of unnecessary queues or unnecessarily restrict the transmission rate of the application.¶
The real-time application carries media over QUIC in a series of QUIC streams [RFC9000]. Each of these streams is marked with a scheduling priority. For example, in a "simulcast" service, audio packets may be sent as QUIC datagrams scheduled at a high priority, then a low definition version of the video stream in a QUIC stream marked at the next highest priority, then medium definition video in another stream, then high definition video in the lowest priority stream. At any given time, the sender would schedule data according to the connection's capacity as evaluated by the congestion control algorithm. If the path has a high capacity the receiver will receive all QUIC streams and enjoy a high definition experience. If the capacity is lower, the higher definition streams will be delayed but the receiver will reliably obtain the medium or low definition version of the media.¶
This real-time retransmission strategy relies on timely assessment of the path capacity by the congestion control algorithm. If the assessment is delayed, the scheduling algorithm will make wrong decisions, such as wrongly believing that the path does not have the capacity to send high definition media, or in contrast sending high definition media and causing queues and maybe packet losses because the lowering of the path capacity has not yet been detected.¶
Other real-time applications may use different categories of traffic than low or high definition video, but they will follow the general principle of trying to schedule just the right amount of transmission to obtain a good experience without creating queues.¶
Real-time transmissions have a variable data rate. Different media codecs have different behavior, but a common pattern is to periodically send "fully encoded" frames, followed by series of "differentially encoded" frame. The fully encoded frames serve as possible synchronization points at which a recipient can start rendering data, while the differentially encoded frames can only be processed if the previous frames have been received. This structure create periodic peaks of traffic, during which the connection might experience some for of congestion, followed by long periods of relatively low bandwidth demand, which the congestion control algorithm may treat as "application limited". Of course, if the bandwidth is too low, even the differentially encode frame may create some congestion. In the simulcast structure, the application would react to congestion by delaying the higher definition video channels.¶
In our experience, we see that BBR generally works well for these applications. However, we see problems in the early stage of the connection, when the path capacity is not yet assessed, and during sudden transitions, such as experienced in Wi-Fi networks. There are also details of the algorithm that work poorly when the traffic alternate between periodic congestion and application limited periods.¶
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP┬á14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.¶
With BBRv3, we see less than ideal "real-time" performance during the initial phase of the connection, during Wi-Fi suspension events, or when the quality of the wireless transmission changes suddenly. We also see anomalous behavior that may be tied to the "application limited" nature of the application.¶
These issues are developed in the following sections.¶
At the beginning of the connection, BBR is initially in Startup state. In Startup BBR sets BBR.pacing_gain to BBRStartupPacingGain (2.77) and BBR.cwnd_gain to BBRStartupCwndGain (2). BBR exits the end of the startup phase when the estimated bandwidth does grow significantly (by at least 25%) for 3 consecutive rounds, or if packet losses are detected. This results in an algorithm that robustly discovers the bottleneck bandwidth. Inconveniently for us, is also results in building significant queues.¶
Towards the end of the Startup phase, we reach a point when both the bottleneck bandwidth and the min RTT have been correctly estimated. After that point, per specification, the pacing rate will be set to 2.77 times the bottleneck bandwidth, and the CWND to twice the product of estimated bandwidth and min RTT. Since the pacing rate is significantly larger than the bottleneck capacity, packets will be queued at the bottleneck. If the bottleneck has sufficient buffers, the queue size will increase until the amount of bytes in transit matches the CWND value. The RTT will increase to twice the min RTT, and will remain at that level for 3 roundtrips, which means 6 times the minimum RTT.¶
The extra buffers and the extra delays do affect the performance of real time application. The application will end up sending more "low priority" data than necessary, resulting in a form of "priority inversion" as high priority data are queued and low priority data are delivered during the "draining" phase that follows startup.¶
After Startup and Drain, BBR will enter the ProbeBW state. BBRv2 organizes those as series of epochs, each composed of a ProbeBW-Down rounds, a variable number of ProbeBW-Cruise rounds, a ProbeBW-Refill round, and finally a series of ProbeBW-Up phases. The number of ProbeBW-Cruise rounds is computed so that BBR competes fairly with other congestion control algorithms like Reno or Cubic. In some conditions, there can be up to 60 rounds of ProbeBW-Cruise, during which the pacing rate will never exceed the bandwidth estimated previously.¶
If BBR exited the Startup state too soon for any reason, the large number of ProbeBW-Cruise causes the connection to retain a very low pacing rate for a long time. In real time applications, this could mean using a much lower video definition than actual bandwidth permits.¶
We observed that Wi-Fi networks often go into "suspension" for brief intervals (see [Wi-Fi-Suspension-Blog], typically of 100 to 200ms. During the transmission, the Wi-Fi radio appears to be turned off, or maybe switched to a different channel. Packets sent by the application during the interval are kept in a queue for the Wi-Fi driver. Packets sent by remote nodes are at the Wi-Fi router. After the suspension, queued packets are sent. There is generally no packet loss, merely an added delay. We have observed this behavior on multiple Wi-Fi networks, using different brands of Wi-Fi equipment and different operating systems. We can speculate about the cause of the suspension, such as exploring alternative Wi-Fi channels or saving energy.¶
When Wi-Fi suspension happens, the sender, using BBR, keeps sending at the computed pacing rate as long as the congestion window allows -- these packets will be queued. The suspension often lasts longer than 1 PTO, at which point no new data will be sent but some queued packets maybe repeated. Then, at the end of the suspension, the sender starts receiving ACK from the peer that were queued at the router. BBR refreshes the congestion parameters, and start polling the application. New data will be sent in order of priority, ending with the 1080p frames.¶
We can describe this process as a succession of phases:¶
Normal transmission, at the application "nominal" rate when all quality levels are sent, with no queue.¶
Undetected suspension, during which queues are building up with a mix of data of various priority levels¶
Detected suspension, during which no more data is sent and new audio or video frames are kept in application queues.¶
End of suspension, during which the Wi-Fi driver sent the queued packets,¶
Resumption, during which the sender using BBR progressively ramps up the sending rate and dequeues the frames stuck in application queues, in order of priority¶
and back to normal state, when the application queues are emptied rapidly.¶
The driver queue fills during the "undetected suspension" state, which currently lasts one PTO, i.e., a bit more than one RTT. So we get at least one RTT worth of audio, 360p, 720p and 1080p frames in the pipe-line, typically queued in front of the Wi-Fi driver. These frames will be delivered by the network before any other frame sent during the "resumption". We will observe some kind of "priority inversion".¶
If new suspension happens shortly after the first one interval, it may catch the system during the "resumption" state. During that state, the sender may have ramped up the data rate to match the underlying network rate (maybe 100 Mbps) and empty the application queues quickly. The transmission rate may be several times the normal rate. The amount of frames queued in front of the Wi-Fi driver will be several times more than during the first suspension. The "priority inversion" effect will be much larger. Depending on random events, we may see the 360p video freezing while the 1080p video is still animating.¶
Wi-Fi bandwidth can drop suddenly due to small changes in the position or orientation of the device, or if a mobile obstacle like a human body moves into the path of transmission. We see small events causing a drastic reduction in bandwidth, and sometimes a marked increase in packet loss rates. We expect that the congestion control algorithm will quickly detect the bandwidth reduction, and inform the application, but in practice the detection always lag the event.¶
As in the Wi-Fi suspension scenario, the delay in detecting the bandwidth reduction causes priority inversions. The application will continue sending data at the previously allowed rate for some time, which will cause less important packets (say 1080p) to be enqueued for transmission. These packets will likely be queued in front of the Wi-Fi driver, and will be transmitted before any more urgent data can be sent.¶
BBRv3 will reduce the CWND quickly if packet losses are detected, but these losses will typically only be detected after a couple of RTT -- or maybe more if the packet queues are allowed to grow to large sizes. The effect is amplified if the pacing rate or the CWND was over estimated before the event, maybe because the connection was application limited.¶
The bandwidth will be restored when the position or orientation of the device changes again, or if the obstacles are removed. When that happens, we expect that the congestion control algorithm will find out rapidly and let the application resume full transmission, but this will only happen after when BBR reaches the ProbeBW UP state, which may take a large number of RTTs with BBR v3.¶
In addition to suspension, we also observe that the congestion control API tends to gradually drop the packing rate over the long run.¶
It seems that the application never fully uses the available rate, and that successive queuing events cause the bandwidth to go down progressively. Maybe we have something systematic here? The application always stays within the limits posed by congestion control, which means that the measured rate will always be lower than the allowed rate. BBR sometimes move to a "probe BW UP" state, but the application does not appear to push much more transmission during that state, so the measured rate does not really increase.¶
The hypothesis is that the application never sends faster than pacing permits, so there is some kind of negative feedback loop.¶
The downward spiral may be compounded by the practice of resetting streams that have fallen behind. A new stream will be restarted for the next block of frames, starting with the next I-Frame, using the "stream per group" approach defined in [I-D.ietf-moq-transport]. The application will try to send more data at that point, but the probing rate will stay low until until the next "probe BW UP" phase, which may be a few seconds away. When BBR probes for more data, the high speed stream may have already been reset, and the offered bandwidth will not really "push up" the data rate.¶
The offered rate of real-time application alternates between rare peaks during the transmission of I-Frames, and a lower data rate between these frames. In a typical setup, BBR will operate in "application limited" mode most of the time, except maybe during the transmission of I-Frames, or if the network becomes congested.¶
If BBR reaches the Probe BW UP state while application limited, it will only exit that state when the next I-Frames are sent, which may be seconds away, or if congestion happens. When congestion or Wi-Fi suspension happens in that state, both the pacing rate and the congestion window will be larger than the "cruise" value. In the case of congestion, the connection will remain in Probe BW UP state during three rounds. This will exacerbate the building of queues and the occurrence of priority inversion during congestion events.¶
We implemented several improvements to BBR for handling real time applications:¶
implement an "early exit" from startup.¶
add an option for rapid start of ProbeBW-Up,¶
add explicit handling of "suspension" to BBR,¶
add detection of feedback loss to minimize building of queues during suspension,¶
exiting Probe BW UP on delay increase,¶
entering Probe BW UP after new streams are started¶
BBR exits start up if packet losses are observed, or if the estimated bottleneck bandwidth does not increase for 3 rounds. We added a third exit condition, exit the startup state if the RTT increases too much.¶
We defined too much as "the RTT measurement is at least 25% larger than the min RTT. However, there can be significant jitter in RTT measurements, and we do not want to exit startup based solely on an event caused by random circumstances. Instead, we exit Startup only if 7 consecutive measurements of the RTT are significantly larger than the Min RTT.¶
Even with the requirement of 7 measurements, there is still a risk that spurious events cause early exit of startup while the bandwidth is not properly assessed. We mitigate these risks by requiring an early entry in ProbeBW-Up state during the first ProbeBW epoch after StartUp.¶
We add to the BBR state a flag "bw_probe_quickly", to signal a desire to enter the ProbeBW-Up state at the first opportunity. When that flag is set, instead of moving to the ProbeBW-Cruise state at the end of the ProbeBW-Down round, BBR moves directly to ProbeBW-Refill and then ProbeBW-UP. The flag is cleared after starting ProbeBW-UP.¶
During each round of ProbeBW-Up, BBR will increase the pacing rate to 25% more than the measured value. The combination of exiting Startup after a delay test and then starting ProbeBW-UP quickly is very similar to the two phases of Hystart++ [RFC9406].¶
During Wi-Fi suspension, the BBR endpoint will not receive any acknowledgement. This will cause the PTO and then RTO timers to expire. We handle that issue by adding a "PTO recovery" flag to the BBR state. The flag is set if a PTO event occurs, at which point BBR enters "packet conservation" mode, in which only a single packet is sent after each PTO or RTO event.¶
The first acknowledgement receiver after setting the PTO recovery flag causes the flag to be cleared, and BBR to reenter the "Startup" state.¶
Detecting suspension by waiting for a PTO works, but still allows transmission of a full CWND worth of packets between the start of the suspension and the PTO timer. We mitigate that by adding a "feedback lost" event.¶
In normal operation, successive acknowledgements are received at short intervals. These intervals can be predicted if the QUIC implementation supports the "ACK Frequency" extension [I-D.ietf-quic-ack-frequency], which directs the peer to acknowledge packets before a maximum delay. The "feedback lost" event is set if expected packet acknowledgements do not arrive after twice the predicted interval.¶
Our modified BBR code reacts to this event by resetting the CWND to the current "in-flight" value, which effectively prevents adding new packets until a next acknowledgement is received. The "feedback lost" event typically arrives sooner than the PTO event, and thus helps avoiding queue building during the "undetected" phase of suspension defined in Section 3.3.¶
We added a test of delay increase when in ProbeBW-Up state. This helps avoiding build excessive queues in ProbeBW-up state, and it also avoids lingering in ProbeBW-up state (see Section 3.6).¶
The bandwidth drift issue happens largely because there is no synchronization between the increase of bandwidth demand and the probing cycle of BBR. We are considering adding a mechanism to synchronize bandwidth probing with period of higher bandwidth demand, characterized for example with the beginning of new streams.¶
At the time of this writing, this synchronization mechanism is not yet implemented.¶
In Section 4, we provide a list of improvements to BBR to better serve real-time applications. We validated these improvements with a mix of simulations and real-life testing. We also considered other improvements, which appeared logical at first sight but were not validated by simulations or experience: pushing the bandwidth limit by injecting fake traffic; and,¶
We were concerned that BBR did not exploit the full bandwidth of the connection. Real-time applications often appear "bandwidth limited", and the bandwidth discovered by BBR is never greater than the highest bandwidth previously used by the application. This can slow down the transmission of data during peaks, for example when the video codec decides to produce a "self referenced" frame (sometimes called I-Frame). To solve that, we tried an option to add "filler" traffic during the "probe BW UP" phases, instructing the QUIC code to schedule redundant copies of the previously transmitted packets in order to "fill the pipe".¶
This did not quite pan out. The various simulation did not show any particular improvement, and in fact some test cases show worse results. Pushing the bandwidth limits increases the risk of filling bottleneck buffers and causing queues or losses, and discovering a high bandwidth did not really compensate for that.¶
The theoretical case for injecting fake traffic during just a few segments of the connection is also weak. Suppose that multiple connections competing for the same bottleneck do that. If they probe the bandwidth at different times, they may all find that there is a lot of capacity available. But if all the connection try to use that bandwidth, they will experience congestion. Similarly, if a real time connection competes with an elastic connection, the elastic connection may back off a little when the real-time connection is probing, but it will quickly discover and consume the available bandwidth after that. If the application wanted to "reserve" a higher bandwidth, it would probably need to inject filler traffic all the time, to "defend" its share of bandwidth. We were not sure that this would be a good idea.¶
We tested adding fewer packets in transit for low priority flows so there will be fewer "priority inversions" during events like Wifi Suspension. We implemented that using QUIC's flow control functions. The results did not really confirm our hypothesis. The change appears to create two competing effects:¶
having fewer packets in transit does limit the priority inversions, and some tests to show a (very small) improvement,¶
but when the low priority flows are rate limited, their load gets spread overtime, instead of being dealt with immediately when plenty of bandwidth is available. This means more packets lingering over time and thus more priority inversions during network events.¶
The measurements also show that any random change in scheduling does create random changes in performance, which means that observing small effects can be very misleading. Given the lack of compelling results, we did not push the idea further.¶
We do not believe that changes in the BBR implementation introduce new security issues.¶
This document has no IANA actions.¶
TODO acknowledge.¶