A Google Congestion Control
Algorithm for Real-Time CommunicationGoogleKungsbron 2Stockholm11122SwedenGoogleKungsbron 2Stockholm11122Swedenholmer@google.comGoogleKungsbron 2Stockholm11122Swedenharald@alvestrand.noThis document describes two methods of congestion control when using
real-time communications on the World Wide Web (RTCWEB); one
sender-based and one receiver-based.It is published as an input document to the RMCAT working group on
congestion control for media streams. The mailing list of that WG is
rmcat@ietf.org.The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
"SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this
document are to be interpreted as described in RFC 2119.Congestion control is a requirement for all applications that wish to
share the Internet .The problem of doing congestion control for real-time media is made
difficult for a number of reasons:The media is usually encoded in forms that cannot be quickly
changed to accommodate varying bandwidth, and bandwidth requirements
can often be changed only in discrete, rather large stepsThe participants may have certain specific wishes on how to
respond - which may not be reducing the bandwidth required by the
flow on which congestion is discoveredThe encodings are usually sensitive to packet loss, while the
real time requirement precludes the repair of packet loss by
retransmissionThis memo describes two congestion control algorithms that
together are seen to give reasonable performance and reasonable (not
perfect) bandwidth sharing with other conferences and with TCP-using
applications that share the same links.The signalling used consists of standard RTP timestamps possibly augmented with RTP transmission time
offsets , standard RTCP feedback reports
and Temporary Maximum Media Stream Bit Rate Requests (TMMBR) as defined
in section 3.5.4, or by using the REMB
feedback report defined in The mathematics of this document have been transcribed from a more
formula-friendly format.The following notational conventions are used:The variable X, where X is a vector -
conventionally marked by a bar on top of the variable name.An estimate of the true value of variable X -
conventionally marked by a circumflex accent on top of the
variable name.The "i"th value of X - conventionally marked by
a subscript i.A row vector consisting of elements x, y and
z.The transpose of vector X_bar.The expected value of the stochastic variable
XThe following elements are in the system:RTP packet - an RTP packet containing media data.Frame - a set of RTP packets transmitted from the sender at the
same time instant. This could be a video frame, an audio frame, or a
mix of audio and video packets. A frame can be defined by the RTP
packet send time (RTP timestamp + transmission time offset), or by
the RTP timestamp if the transmission time offset field is not
present.Incoming media streams - a stream of frames consisting of RTP
packets.Media codec - has a bandwidth control, and encodes the incoming
media stream into an RTP stream.RTP sender - sends the RTP stream over the network to the RTP
receiver. Generates the RTP timestamp.RTP receiver - receives the RTP stream, notes the time of
arrival. Regenerates the media stream for the recipient.RTCP sender at RTP sender - sends sender reports with mappings
between RTP timestamps and NTP time.RTCP sender at RTP receiver - sends receiver reports and
TMMBR/REMB messages.RTCP receiver at RTP sender - receives receiver reports and
TMMBR/REMB messages, reports these to sender side control.RTCP receiver at RTP receiver.Sender side control - takes loss rate info, round trip time info,
and TMMBR/REMB messages and computes a sending bitrate.Receiver side control - takes the packet arrival info at the RTP
receiver and decides when to send TMMBR/REMB messages.Together, sender side control and receiver side control
implement the congestion control algorithm.The receive-side algorithm can be further decomposed into four parts:
an RTP timestamp to NTP time conversion, arrival-time filter, an
over-use detector, and a remote rate-control.It is common that multiple RTP streams are sent from the sender to
the receiver. In such a situation the RTP timestamps of incoming can
first be converted to a common time base using the RTP timestamp and
NTP time pairs in RTCP SR reports. The
converted timestamps can then be used instead of RTP timestamps in the
arrival-time filtering, and since all streams from the same sender
have timestamps in the same time base they can all be processed by the
same filter. This has the advantage of quicker reactions and reduces
problems of noisy measurements due to self-inflicted
cross-traffic.In the time interval from the start of the call until a stream from
the same sender has received an RTCP SR report, the receiver-side
control operates in single-stream mode. In that mode only one RTP
stream can be processed by the over-use detector. As soon as a stream
has received one or more RTCP SR reports the receiver-side control can
change to a multi-stream mode, where all RTP streams from the same
sender which have received one or more RTCP SR reports can be
processed by the over-use detector. When switching to the multi-stream
mode the state of the over-use detector must be modified to avoid a
time base mismatch. This can either be done by resetting the stored
RTP timestamp values or by converting them using the newly received
RTCP SR report.This section describes an adaptive filter that continuously updates
estimates of network parameters based on the timing of the received
frames.At the receiving side we are observing groups of incoming packets,
where each group of packets corresponding to the same frame having
timestamp T(i).Each frame is assigned a receive time t(i), which corresponds to
the time at which the whole frame has been received (ignoring any
packet losses). A frame is delayed relative to its predecessor if
t(i)-t(i-1)>T(i)-T(i-1), i.e., if the arrival time difference is
larger than the timestamp difference.We define the (relative) inter-arrival time, d(i) asSince the time ts to send a frame of size L over a path with a
capacity of C is roughlywe can model the inter-arrival time asHere, w(i) is a sample from a stochastic process W, which is a
function of the capacity C, the current cross traffic X(i), and the
current send bit rate R(i). We model W as a white Gaussian process. If
we are over-using the channel we expect w(i) to increase, and if a
queue on the network path is being emptied, w(i) will decrease;
otherwise the mean of w(i) will be zero.Breaking out the mean m(i) from w(i) to make the process zero mean,
we getThis is our fundamental model, where we take into account that a
large frame needs more time to traverse the link than a small frame,
thus arriving with higher relative delay. The noise term represents
network jitter and other delay effects not captured by the model.When graphing the values for d(i) versus dL(i) on a scatterplot, we
find that most samples cluster around the center, and the outliers are
clustered along a line with average slope 1/C and zero offset.For instance, when using a regular video codec, most frames are
roughly the same size after encoding (the central
“cloud”); the exceptions are I-frames (or key frames)
which are typically much larger than the average causing positive
outliers (the I-frame itself) and negative outliers (the frame after
an I-frame) on the dL axis. Audio frames on the other hand often
consist of single packets of equal size, and an audio-only media
stream would have its frames scattered at dL = 0.The parameters d(i) and dL(i) are readily available for each frame
i > 1, and we want to estimate C(i) and m(i) and use those
estimates to detect whether or not we are over-using the bandwidth
currently available. These parameters are easily estimated by any
adaptive filter – we are using the Kalman filter.Letand call it the state of time i. We model the state evolution from
time i to time i+1 aswhere u_bar(i) is the zero mean white Gaussian process noise with
covarianceGiven equation 5 we getwhere v(i) is zero mean white Gaussian measurement noise with
variance var_v = sigma(v,i)^2The Kalman filter recursively updates our estimateasI is the 2-by-2 identity matrix.The variance var_v = sigma(v,i)^2 is estimated using an exponential
averaging filter, modified for variable sampling ratewhere f_max = max {1/(T(j) - T(j-1))} for j in i-K+1...i is the
highest rate at which frames have been captured by the camera the last
K frames and alpha is a filter coefficient typically chosen as a
number in the interval [0.1, 0.001]. Since our assumption that v(i)
should be zero mean WGN is less accurate in some cases, we have
introduced an additional outlier filter around the updates of
var_v_hat. If z(i) > 3 var_v_hat the filter is updated with 3
sqrt(var_v_hat) rather than z(i). For instance v(i) will not be white
in situations where packets are sent at a higher rate than the channel
capacity, in which case they will be queued behind each other. In a
similar way, Q(i) is chosen as a diagonal matrix with main diagonal
elements given byIt is necessary to scale these filter parameters with the frame
rate to make the detector respond as quickly at low frame rates as at
high frame rates.The offset estimate m(i) is compared with a threshold gamma_1. An
estimate above the threshold is considered as an indication of
over-use. Such an indication is not enough for the detector to signal
over-use to the rate control subsystem. Not until over-use has been
detected for at least gamma_2 milliseconds and at least gamma_3
frames, a definitive over-use will be signaled. However, if the offset
estimate m(i) was decreased in the last update, over-use will not be
signaled even if all the above conditions are met. Similarly, the
opposite state, under-use, is detected when m(i) < -gamma_1. If
neither over-use nor under-use is detected, the detector will be in
the normal state.The rate control at the receiving side is designed to increase the
receive-side estimate of the available bandwidth A_hat as long as the
detected state is normal. Doing that assures that we, sooner or later,
will reach the available bandwidth of the channel and detect an
over-use.As soon as over-use has been detected the receive-side estimate of
the available bandwidth is decreased. In this way we get a recursive
and adaptive estimate of the available bandwidth.In this document we make the assumption that the rate control
subsystem is executed periodically and that this period is
constant.The rate control subsystem has 3 states: Increase, Decrease and
Hold. "Increase" is the state when no congestion is detected;
"Decrease" is the state where congestion is detected, and "Hold" is a
state that waits until built-up queues have drained before going to
"increase" state.The state transitions (with blank fields meaning "remain in state")
are:The subsystem starts in the increase state, where it will stay
until over-use or under-use has been detected by the detector
subsystem. On every update the receive-side estimate of the available
bandwidth is increased with a factor which is a function of the global
system response time and the estimated measurement noise variance
var_v_hat. The global system response time is the time from an
increase that causes over-use until that over-use can be detected by
the over-use detector. The variance var_v_hat affects how responsive
the Kalman filter is, and is thus used as an indicator of the delay
inflicted by the Kalman filter.Here, B, b, d, c1 and c2 are design parameters.Since the system depends on over-using the channel to verify the
current available bandwidth estimate, we must make sure that our
estimate doesn't diverge from the rate at which the sender is actually
sending. Thus, if the sender is unable to produce a bit stream with
the bit rate the receiver is asking for, the available bandwidth
estimate must stay within a given bound. Therefore we introduce a
thresholdwhere R_hat(i) is the incoming bit rate measured over a T seconds
window:N(i) is the number of frames received the past T seconds and L(j)
is the payload size of frame j. Ideally T should be chosen to match
the rate controller at the sender. A window between 0.5 and 1 second
is recommended.When an over-use is detected the system transitions to the decrease
state, where the receive-side available bandwidth estimate is
decreased to a factor times the currently incoming bit rate.alpha is typically chosen to be in the interval [0.8, 0.95].When the detector signals under-use to the rate control subsystem,
we know that queues in the network path are being emptied, indicating
that our available bandwidth estimate is lower than the actual
available bandwidth. Upon that signal the rate control subsystem will
enter the hold state, where the receive-side available bandwidth
estimate will be held constant while waiting for the queues to
stabilize at a lower level – a way of keeping the delay as low
as possible. This decrease of delay is wanted, and expected,
immediately after the estimate has been reduced due to over-use, but
can also happen if the cross traffic over some links is reduced. In
either case we want to measure the highest incoming rate during the
under-use interval:where K is the number of frames of under-use before returning to
the normal state. R_max is a measure of the actual bandwidth available
and is a good guess of what bit rate the sender should be able to
transmit at. Therefore the receive-side available bandwidth estimate
will be set to R_max when we transition from the hold state to the
increase state.One design decision is when to send rate control messages. The time
from a change in congestion to the sending of the feedback message is
a limitation on how fast the sender can react. Sending too many
messages giving no new information is a waste of bandwidth - but in
the case of severe congestion, feedback messages can be lost,
resulting in a failure to react in a timely manner.The conclusion is that feedback messages should be sent on a
"heartbeat" schedule, allowing the sender side control to react to
missing feedback messages by reducing its send rate, but they should
also be sent whenever the estimated bandwidth value has changed
significantly, without waiting for the heartbeat time, up to some
limiting upper bound on the send rate.The minimum interval is named t_min_fb_interval.The maximum interval is named t_max_fb_interval.The permissible values of these intervals will be bounded by the
RTP session's RTCP bandwidth and its rtcp_frr setting.[TODO: Get some example values for these timers]An additional congestion controller resides at the sending side. It
bases its decisions on the round-trip time, packet loss and available
bandwidth estimates transmitted from the receiving side.The available bandwidth estimates produced by the receiving side are
only reliable when the size of the queues along the channel are large
enough. If the queues are very short, over-use will only be visible
through packet losses, which aren't used by the receiving side
algorithm.This algorithm is run every time a receive report arrives at the
sender, which will happen no more often than t_min_fb_interval, and no
less often than t_max_fb_interval. If no receive report is received
within 2x t_max_fb_interval (indicating at least 2 lost feedback
reports), the algorithm will take action as if all packets in the
interval have been lost, resulting in a halving of the send rate.If 2-10% of the packets have been lost since the previous report
from the receiver, the sender available bandwidth estimate As(i) (As
denotes ‘sender available bandwidth’) will be kept
unchanged.If more than 10% of the packets have been lost a new estimate is
calculated as As(i)=As(i-1)(1-0.5p), where p is the loss ratio.As long as less than 2% of the packets have been lost As(i) will
be increased as As(i)=1.05(As(i-1)+1000)The new send-side estimate is limited by the TCP Friendly Rate
Control formula and the receive-side
estimate of the available bandwidth A(i):where b is the number of packets acknowledged by a single TCP
acknowledgement (set to 1 per TFRC recommendations), t_RTO is the TCP
retransmission timeout value in seconds (set to 4*R) and s is the
average packet size in bytes. R is the round-trip time in seconds.(The multiplication by 8 comes because TFRC is computing bandwidth in
bytes, while this document computes bandwidth in bits.)In words: The sender-side estimate will never be larger than the
receiver-side estimate, and will never be lower than the estimate from
the TFRC formula.We motivate the packet loss thresholds by noting that if the
transmission channel has a small amount of packet loss due to over-use,
that amount will soon increase if the sender does not adjust his bit
rate. Therefore we will soon enough reach above the 10 % threshold and
adjust As(i). However if the packet loss rate does not increase, the
losses are probably not related to self-induced channel over-use and
therefore we should not react on them.There are three scenarios of interest, and one included for
referenceBoth parties implement the algorithms described hereSender implements the algorithm described in section , recipient does not implement Recipient implements the algorithm in section , sender does not implement .In the case where both parties implement the algorithms, we
expect to see most of the congestion control response to slowly varying
conditions happen by TMMBR/REMB messages from recipient to sender. At
most times, the sender will send less than the congestion-inducing
bandwidth limit C, and when he sends more, congestion will be detected
before packets are lost.If sudden changes happen, packets will be lost, and the sender side
control will trigger, limiting traffic until the congestion becomes low
enough that the system switches back to the receiver-controlled
state.In the case where sender only implements, we expect to see somewhat
higher loss rates and delays, but the system will still be overall TCP
friendly and self-adjusting; the governing term in the calculation will
be the TFRC formula.In the case where recipient implements this algorithm and sender does
not, congestion will be avoided for slow changes as long as the sender
understands and obeys TMMBR/REMB; there will be no backoff for
packet-loss-inducing changes in capacity. Given that some kind of
congestion control is mandatory for the sender according to the TMMBR
spec, this case has to be reevaluated against the specific congestion
control implemented by the sender.This algorithm has been implemented in the open-source WebRTC
project.This draft is offered as input to the congestion control
discussion.Work that can be done on this basis includes:Consideration of timing info: It may be sensible to use the
proposed TFRC RTP header extensions to carry per-packet
timing information, which would both give more data points and a
timestamp applied closer to the network interface. This draft
includes consideration of using the transmission time offset defined
in Considerations of cross-channel calculation: If all packets in
multiple streams follow the same path over the network, congestion
or queueing information should be considered across all packets
between two parties, not just per media stream. A feedback message
(REMB) that may be suitable for such a purpose is given in .Considerations of cross-channel balancing: The decision to slow
down sending in a situation with multiple media streams should be
taken across all media streams, not per stream.Considerations of additional input: How and where packet loss
detected at the recipient can be added to the algorithm.Considerations of locus of control: Whether the sender or the
recipient is in the best position to figure out which media streams
it makes sense to slow down, and therefore whether one should use
TMMBR to slow down one channel, signal an overall bandwidth change
and let the sender make the decision, or signal the (possibly
processed) delay info and let the sender run the algorithm.Considerations of over-bandwidth estimation: Whether we can use
the estimate of how much we're over bandwidth in section 3 to
influence how much we reduce the bandwidth, rather than using a
fixed factor.Startup considerations. It's unreasonable to assume that just
starting at full rate is always the best strategy.Dealing with sender traffic shaping, which delays sending of
packets. Using send-time timestamps rather than RTP timestamps may
be useful here, but as long as the sender's traffic shaping does not
spread out packets more than the bottleneck link, it should not
matter.Stability considerations. It is not clear how to show that the
algorithm cannot provide an oscillating state, either alone or when
competing with other algorithms / flows.These are matters for further work; since some of them involve
extensions that have not yet been standardized, this could take some
time.This document makes no request of IANA.Note to RFC Editor: this section may be removed on publication as an
RFC.An attacker with the ability to insert or remove messages on the
connection will, of course, have the ability to mess up rate control,
causing people to send either too fast or too slow, and causing
congestion.In this case, the control information is carried inside RTP, and can
be protected against modification or message insertion using SRTP, just
as for the media. Given that timestamps are carried in the RTP header,
which is not encrypted, this is not protected against disclosure, but it
seems hard to mount an attack based on timing information only.Thanks to Randell Jesup, Magnus Westerlund, Varun Singh, Tim Panton,
Soo-Hyun Choo, Jim Gettys, Ingemar Johansson, Michael Welzl and others
for providing valuable feedback on earlier versions of this draft.Added change logAdded appendix outlining new extensionsAdded a section on when to send feedback to the end of section
3.3 "Rate control", and defined min/max FB intervals.Added size of over-bandwidth estimate usage to "further work"
section.Added startup considerations to "further work" section.Added sender-delay considerations to "further work"
section.Filled in acknowledgements section from mailing list
discussion.Defined the term "frame", incorporating the transmission time
offset into its definition, and removed references to "video
frame".Referred to "m(i)" from the text to make the derivation
clearer.Made it clearer that we modify our estimates of available
bandwidth, and not the true available bandwidth.Removed the appendixes outlining new extensions, added pointers
to REMB draft and RFC 5450.Added a section on how to process multiple streams in a single
estimator using RTP timestamps to NTP time conversion.Stated in introduction that the draft is aimed at the RMCAT
working group.Renamed draft to link the draft name to the rmcat WG.