]> Ericsson AB

Laboratoriegränd 11 Luleaa 977 53 Sweden +46 730783289 ingemar.s.johansson@ericsson.com

Ericsson AB

Laboratoriegränd 11 Luleaa 977 53 Sweden +46 761153743 zaheduzzaman.sarker@ericsson.com

TSV RMCAT WG Cellular Network Congestion Control RTP This memo describes a rate adaptation algorithm for conversational media services such as video. The solution conforms to the packet conservation principle and uses a hybrid loss and delay based congestion control algorithm. The algorithm is evaluated over both simulated Internet bottleneck scenarios as well as in a Long Term Evolution (LTE) system simulator and is shown to achieve both low latency and high video throughput in these scenarios.

Congestion in the Internet occurs when the transmitted bitrate is higher than the available capacity over a given transmission path. Applications that are deployed in the Internet MUST employ congestion control, to achieve robust performance and to avoid congestion collapse in the Internet. Interactive realtime communication imposes a lot of requirements on the transport, therefore a robust, efficient rate adaptation for all access types is an important part of interactive realtime communications as the transmission channel bandwidth MAY vary over time. Wireless access such as LTE, which is an integral part of the current Internet, increases the importance of rate adaptation as the channel bandwidth of a default LTE bearer can change considerably in a very short time frame. Thus a rate adaptation solution for interactive realtime media, such as WebRTC, SHOULD be both quick and be able to operate over a large range in channel capacity. This memo describes SCReAM (Self-Clocked Rate Adaptation for Multimedia), a solution that is based on the self-clocking principle of TCP and uses techniques similar to what is used in the LEDBAT based rate adaptation algorithm . SCReAM is not entirely self-clocked as it augments self-clocking with pacing and a minimum send rate.

describes the complications that can be observed in wireless environments. Wireless access such as LTE can typically not guarantee a given bandwidth, this is true especially for default bearers. The network throughput MAY vary considerably for instance in cases where the wireless terminal is moving around. Even though LTE can support bitrates well above 100Mbps, there are cases when the available bitrate can be much lower, examples are situations with high network load and poor coverage. An additional complication is that the network throughput MAY drop for short time intervals at e.g. handover, these short glitches are initially very difficult to distinguish from more permanent reductions in throughput. Unlike wireline bottlenecks with large statistical multiplexing it is not possible to try to maintain a given bitrate when congestion is detected with the hope that other flows will yield, this is because there are generally few other flows competing for the same bottleneck. Each user gets its own variable throughput bottleneck, where the throughput depends on factors like channel quality, network load and historical throughput. The bottom line is, if the throughput drops, the sender has no other option than to reduce the bitrate. Once the radio scheduler has reduced the resource allocation for a bearer, an RMCAT flow in that bearer SHOULD reduce the sending rate quite quickly (within one RTT) in order to avoid excessive queuing delay or packet loss.

Self-clocked congestion control algorithms provide a benefit over the rate based counterparts in that the former consists of two adaptation mechanisms: A congestion window computation that evolves over a longer timescale (several RTTs) especially when the congestion window evolution is dictated by estimated delay (to minimize vulnerability to e.g. short term delay variations). A fine grained congestion control given by the self-clocking which operates on a shorter time scale (1 RTT). The benefits of self-clocking are also elaborated upon in . A rate based congestion control typically adjusts the rate based on delay and loss. The congestion detection needs to be done with a certain time lag to avoid over-reaction to spurious congestion events such as delay spikes. Despite the fact that there are two or more congestion indications, the outcome is still that there is still only one mechanism to adjust the sending rate. This makes it difficult to reach the goals of high throughput and prompt reaction to congestion.

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 .

The core SCReAM algorithm has similarities to the concepts of self-clocking used in TFWC and follows the packet conservation principle. The packet conservation principle is described as an important key-factor behind the protection of networks from congestion . In SCReAM, the receiver of the media echoes a list of received RTP packets and the timestamp of the RTP packet with the highest sequence number back to the sender in feedback packets. The sender keeps a list of transmitted packets, their respective sizes and the time they were transmitted. This information is used to determine the number of bytes that can be transmitted at any given time instant. A congestion window puts an upper limit on how many bytes can be in flight, i.e. transmitted but not yet acknowledged. The congestion window is determined in a way similar to LEDBAT . LEDBAT is a congestion control algorithm that uses send and receive timestamps to estimate the queuing delay (from now on denoted qdelay) along the transmission path. This information is used to adjust the congestion window. The use of LEDBAT ensures that the end-to-end latency is kept low. shows that LEDBAT has certain inherent issues that makes it counteract its purpose to achieve low delay. The general problem described in the paper is that the base delay is offset by LEDBAT´s own queue buildup. The big difference with using LEDBAT in the SCReAM context lies in the fact that the source is rate limited and that it is required that the RTP queue is kept short (preferably empty). In addition the output from a video encoder is rarely constant bitrate, static content (talking heads) for instance gives almost zero video rate. This gives two useful properties when LEDBAT is used with SCReAM that help to avoid the issues described in : There is always a certain probability that SCReAM is short of data to transmit, which means that the network queue will run empty every once in a while. The max video bitrate can be lower than the link capacity. If the max video bitrate is 5Mbps and the capacity is 10Mbps then the network queue will run empty. It is sufficient that any of the two conditions above is fulfilled to make the base delay update properly. Furthermore describes an issue with short lived competing flows, the case in SCReAM is that these short lived flows will cause the self-clocking in SCReAM to slow down with the result that the RTP queue is built up, which will in turn result in a reduced media video bitrate. SCReAM will thus yield more to competing short lived flows than what is the case with traditional use of LEDBAT. The basic functionality in the use of LEDBAT in SCReAM is quite simple, there are however a few steps to take to make the concept work with conversational media: Congestion window validation techniques. These are similar in action as the method described in . Congestion window validation ensures that the congestion window is limited by the actual number bytes in flight, this is important especially in the context of rate limited sources such as video. Lack of congestion window validation would lead to a slow reaction to congestion as the congestion window does not properly reflect the congestion state in the network. The allowed idle period in this memo is shorter than in , this to avoid excessive delays in the cases where e.g. wireless throughput has decreased during a period where the output bitrate from the media coder has been low, for instance due to inactivity. Furthermore, this memo allows for more relaxed rules for when the congestion window is allowed to grow, this is necessary as the variable output bitrate generally means that the congestion window is often under-utilized. Fast increase makes the bitrate increase faster when no congestion is detected. It makes the media bitrate ramp-up within 5 to 10 seconds. The behavior is similar to TCP slowstart. The fast increase is exited when congestion is detected. The fast increase state can however resume if the congestion level is low, this enables a reasonably quick rate increase in case link throughput increases. A qdelay trend is computed for earlier detection of incipient congestion and as a result it reduces jitter. Addition of a media rate control function. Use of inflection points in the media rate calculation to achieve reduced jitter. Adjustment of qdelay target for better performance when competing with other loss based congestion controlled flows. The above mentioned features will be described in more detail in sections to . The full details are described in .

+---------------------------+ | Media encoder | +---------------------------+ ^ | | |(1) |(3) RTP | V | +-----------+ +---------+ | | | Media | (2) | Queue | | rate |<------| | | control | |RTP packets| +---------+ | | +-----------+ | |(4) RTP | v +------------+ +--------------+ | Network | (7) | Sender | +-->| congestion |------>| Transmission | | | control | | Control | | +------------+ +--------------+ | | |-------------RTCP----------| |(5) (6) | RTP | v +------------+ | UDP | | socket | +------------+

The SCReAM algorithm consists of three main parts: network congestion control, sender transmission control and media rate control. All of these three parts reside at the sender side. shows the functional overview of a SCReAM sender. The receiver side algorithm is very simple in comparison as it only generates feedback containing acknowledgements of received RTP packets and an ECN count.

The network congestion control sets an upper limit on how much data can be in the network (bytes in flight); this limit is called CWND (congestion window) and is used in the sender transmission control. The SCReAM congestion control method, uses techniques similar to LEDBAT to measure the qdelay. As is the case with LEDBAT, it is not necessary to use synchronized clocks in sender and receiver in order to compute the qdelay. It is however necessary that they use the same clock frequency, or that the clock frequency at the receiver can be inferred reliably by the sender. The SCReAM sender calculates the congestion window based on the feedback from the SCReAM receiver. The congestion window is allowed to increase if the qdelay is below a predefined qdelay target, otherwise the congestion window decreases. The qdelay target is typically set to 50-100ms. This ensures that the queuing delay is kept low. The reaction to loss or ECN events leads to an instant reduction of CWND. Note that the source rate limited nature of real time media such as video, typically means that the queuing delay will mostly be below the given delay target, this is contrary to the case where large files are transmitted using LEDBAT congestion control, in which case the queuing delay will stay close to the delay target.

The sender transmission control limits the output of data, given by the relation between the number of bytes in flight and the congestion window. Packet pacing is used to mitigate issues with ACK compression that MAY cause increased jitter and/or packet loss in the media traffic. Packet pacing limits the packet transmission rate given by the estimated link throughput. Even if the send window allows for the transmission of a number of packets, these packets are not transmitted immediately, but rather they are transmitted in intervals given by the packet size and the estimated link throughput.

The media rate control serves to adjust the media bitrate to ramp-up quickly enough to get a fair share of the system resources when link throughput increases. The reaction to reduced throughput MUST be prompt in order to avoid getting too much data queued in the RTP packet queue(s) in the sender. The media bitrate is decreased if the RTP queue size exceeds a threshold. In cases where the sender frame queues increase rapidly such as in the case of a RAT (Radio Access Type) handover it MAY be necessary to implement additional actions, such as discarding of encoded media frames or frame skipping in order to ensure that the RTP queues are drained quickly. Frame skipping results in the frame rate being temporarily reduced. Which method to use is a design choice and outside the scope of this algorithm description.

This section describes the sender side algorithm in more detail. It is split between the network congestion control, sender transmission control and the media rate control. A SCReAM sender implements media rate control and an RTP queue for each media type or source, where RTP packets containing encoded media frames are temporarily stored for transmission. shows the details when a single media source (or stream) is used. A transmission scheduler (not shown in the figure) is added to support multiple streams. The transmission scheduler can enforce differing priorities between the streams and act like a coupled congestion controller for multiple flows. Support for multiple streams is implemented in . Media frames are encoded and forwarded to the RTP queue (1) in . The media rate adaptation adapts to the size of the RTP queue (2) and provides a target rate for the media encoder (3). The RTP packets are picked from the RTP queue (for multiple flows from each RTP queue based on some defined priority order or simply in a round robin fashion) (4) by the sender transmission controller. The sender transmission controller (in case of multiple flows a transmission scheduler) sends the RTP packets to the UDP socket (5). In the general case all media SHOULD go through the sender transmission controller and is limited so that the number of bytes in flight is less than the congestion window. RTCP packets are received (6) and the information about bytes in flight and congestion window is exchanged between the network congestion control and the sender transmission control (7).

Constants and state variables are listed in this section. Temporary variables are not listed, instead they are appended with '_t' in the pseudo code to indicate their local scope.

The RECOMMENDED values, within (), for the constants are deduced from experiments. The units are enclosed in square brackets [ ]. Target value for the minimum qdelay. Target value for the maximum qdelay. This parameter provides an upper limit to how much the target qdelay (qdelay_target) can be increased in order to cope with competing loss based flows. The target qdelay MUST NOT be initialized to this high value however as it would increase e2e delay and also make the rate control and congestion control loop sluggish. Averaging factor for qdelay_fraction_avg. Averaging factor for qdelay_fraction_avg. Headroom for the limitation of CWND. Gain factor for congestion window adjustment. CWND scale factor due to loss event. CWND scale factor due to ECN event. Target rate scale factor due to loss event. Maximum segment size = Max RTP packet size. Interval between media bitrate adjustments. Min target bitrate [bps], bps is bits per second. Max target bitrate [bps]. Maximum allowed rate increase speed. Guard factor against early congestion onset. A higher value gives less jitter, possibly at the expense of a lower link utilization. This value MAY be subject to tuning depending on e.g media coder characteristics, experiments with H264 and VP8 indicate that 0.1 is a suitable value. See and for evaluation of a real implementation. Guard factor against RTP queue buildup. This value MAY be subject to tuning depending on e.g media coder characteristics, experiments with H264 and VP8 indicate that 1.0 is a suitable value. See and for evaluation of a real implementation. RTP queue delay threshold for a target rate reduction. Target rate scale when RTP qdelay threshold exceeds. Threshold value for qdelay_trend. Time span until fast increase can be resumed, given that the qdelay_trend is below QDELAY_TREND_LO.

The values within () indicate initial values. qdelay target, a variable qdelay target is introduced to manage cases where e.g. FTP competes for the bandwidth over the same bottleneck, a fixed qdelay target would otherwise starve the RMCAT flow under such circumstances. The qdelay target is allowed to vary between QDELAY_TARGET_LO and QDELAY_TARGET_HI. EWMA filtered fractional qdelay. Vector of the last 20 fractional qdelay samples. qdelay trend, indicates incipient congestion. Low pass filtered version of qdelay_trend. Vector of the last 100 normalized qdelay samples. Minimum congestion window. True if in fast increase state. Congestion window. The number of bytes that was acknowledged with the last received acknowledgement i.e. bytes acknowledged since the last CWND update. Upper limit to how many bytes that can currently be transmitted. Updated when cwnd is updated and when RTP packet is transmitted. Media target bitrate. Media target bitrate inflection point i.e. the last known highest target_bitrate. Used to limit bitrate increase speed close to the last known congestion point. Measured transmit bitrate. Measured throughput based on received acknowledgements. Measured bitrate from the media encoder. Median value of rate_media, computed over more than 10s. Smoothed RTT [s], computed with a similar method to that described in . Size of RTP packets in queue. Size of the last transmitted RTP packet. The estimated fraction of RTTs with lost packets detected.

This section explains the network congestion control, it contains two main functions: Computation of congestion window at the sender: Gives an upper limit to the number of bytes in flight. Calculation of send window at the sender: RTP packets are transmitted if allowed by the relation between the number of bytes in flight and the congestion window. This is controlled by the send window. SCReAM is a window based and byte oriented congestion control protocol, where the number of bytes transmitted is inferred from the size of the transmitted RTP packets. Thus a list of transmitted RTP packets and their respective transmission times (wall-clock time) MUST be kept for further calculation. The number of bytes in flight (bytes_in_flight) is computed as the sum of the sizes of the RTP packets ranging from the RTP packet most recently transmitted down to but not including the acknowledged packet with the highest sequence number. This can be translated to the difference between the highest transmitted byte sequence number and the highest acknowledged byte sequence number. As an example: If RTP packet with sequence number SN is transmitted and the last acknowledgement indicates SN-5 as the highest received sequence number then bytes in flight is computed as the sum of the size of RTP packets with sequence number SN-4, SN-3, SN-2, SN-1 and SN, it does not matter if for instance packet with sequence number SN-3 was lost, the size of RTP packet with sequence number SN-3 will still be considered in the computation of bytes_in_flight. Furthermore, a variable bytes_newly_acked is incremented with a value corresponding to how much the highest sequence number has increased since the last feedback. As an example: If the previous acknowledgement indicated the highest sequence number N and the new acknowledgement indicated N+3, then bytes_newly_acked is incremented by a value equal to the sum of the sizes of RTP packets with sequence number N+1, N+2 and N+3. Packets that are lost are also included, which means that even though e.g packet N+2 was lost, its size is still included in the update of bytes_newly_acked. The bytes_newly_acked variable is reset to zero after a CWND update. The feedback from the receiver is assumed to consist of the following elements. A list of received RTP packets' sequence numbers. The wall clock timestamp corresponding to the received RTP packet with the highest sequence number. Accumulated number of ECN-CE marked packets (n_ECN). When the sender receives RTCP feedback, the qdelay is calculated as outlined in . A qdelay sample is obtained for each received acknowledgement. No smoothing of the qdelay samples occur, however some smoothing occurs anyway as the computation of the CWND is a low pass filter function. A number of variables are updated as illustrated by the pseudo code below, temporary variables are appended with '_t'. Note that the pseudo code does not show all details for reasons of readability, the reader is encouraged to look into the C++ code in for the details.

<CODE BEGINS> update_variables(qdelay): qdelay_fraction_t = qdelay/qdelay_target #calculate moving average qdelay_fraction_avg = (1-QDELAY_WEIGHT)*qdelay_fraction_avg+ QDELAY_WEIGHT*qdelay_fraction_t update_qdelay_fraction_hist(qdelay_fraction_t) # R is an autocorrelation function of qdelay_fraction_hist # at lag K a = R(qdelay_fraction_hist,1)/R(qdelay_fraction_hist,0) #calculate qdelay trend qdelay_trend = min(1.0,max(0.0,a*qdelay_fraction_avg)) #calculate a 'peak-hold' qdelay_trend, this gives a memory # of congestion in the past qdelay_trend_mem = max(0.99*qdelay_trend_mem, qdelay_trend) <CODE ENDS>

The qdelay fraction is sampled every 50ms and the last 20 samples are stored in a vector (qdelay_fraction_hist). This vector is used in the computation of an qdelay trend that gives a value between 0.0 and 1.0 depending on the estimated congestion level. The prediction coefficient 'a' has positive values if qdelay shows an increasing trend, thus an indication of congestion is obtained before the qdelay target is reached. The autocorrelation function 'R' is defined in . The prediction coefficient is further multiplied with qdelay_fraction_avg to reduce sensitivity to increasing qdelay when it is very small. The 50ms sampling is a simplification and MAY have the effect that the same qdelay is sampled several times, this does however not pose any problem as the vector is only used to determine if the qdelay is increasing or decreasing. The qdelay_trend is utilized in the media rate control to indicate incipient congestion and to determine when to exit from fast increase mode. qdelay_trend_mem is used to enforce a less aggressive rate increase after congestion events. The function update_qdelay_fraction_hist(..) removes the oldest element and adds the latest qdelay_fraction element to the qdelay_fraction_hist vector. A loss event is indicated if one or more RTP packets are declared missing. The loss detection is described in . Once a loss event is detected, further detected lost RTP packets are ignored for a full smoothed round trip time, the intention of this is to limit the congestion window decrease to at most once per round trip. The congestion window back off due to loss events is deliberately a bit less than is the case with e.g. TCP Reno. The reason is that TCP is generally used to transmit whole files, which can be translated to an infinite source bitrate. SCReAM on the other hand has a source whose rate is limited to a value close to the available transmit rate and often below that value, the effect of this is that SCReAM has less opportunity to grab free capacity than a TCP based file transfer. To compensate for this it is RECOMMENDED to let SCReAM reduce the congestion window less than what is the case with TCP when loss events occur. An ECN event is detected if the n_ECN counter in the feedback report has increased since the previous received feedback. Once an ECN event is detected, the n_ECN counter is ignored for a full smoothed round trip time, the intention of this is to limit the congestion window decrease to at most once per round trip. The congestion window back off due to an ECN event MAY be smaller than if a loss event occurs. This is in line with the idea outlined in to enable ECN marking thresholds lower than the corresponding packet drop thresholds. The update of the congestion window depends on whether loss or ECN-marking or neither occurs. The pseudo code below describes actions taken in case of the different events.

<CODE BEGINS> on congestion event(qdelay): # Either loss or ECN mark is detected in_fast_increase = false if (is loss) # loss is detected cwnd = max(min_cwnd,cwnd*BETA_LOSS) else # No loss, so it is then an ECN mark cwnd = max(min_cwnd,cwnd*BETA_ECN) end adjust_qdelay_target(qdelay) #compensating for competing flows calculate_send_window(qdelay,qdelay_target) # when no congestion event on acknowledgement(qdelay): update_bytes_newly_acked() update_cwnd(bytes_newly_acked) adjust_qdelay_target(qdelay) #compensating for competing flows calculate_send_window(qdelay, qdelay_target) check_to_resume_fast_increase() <CODE ENDS>

The methods are further described in detail below.

The congestion window update is based on qdelay, except for the occurrence of loss events (one or more lost RTP packets in one RTT), or ECN events, which was described earlier. Pseudo code for the update of the congestion window is found below.

<CODE BEGINS> update_cwnd(bytes_newly_acked): # in fast increase ? if (in_fast_increase) if (qdelay_trend >= QDELAY_TREND_TH) # incipient congestion detected, exit fast increase in_fast_increase = false else # no congestion yet, increase cwnd if it # is sufficiently used # an additional slack of bytes_newly_acked is # added to ensure that CWND growth occurs # even when feedback is sparse if (bytes_in_flight*1.5+bytes_newly_acked > cwnd) cwnd = cwnd+bytes_newly_acked end return end end # not in fast increase phase # off_target calculated as with LEDBAT off_target_t = (qdelay_target - qdelay) / qdelay_target gain_t = GAIN # adjust congestion window cwnd_delta_t = gain_t * off_target_t * bytes_newly_acked * MSS / cwnd if (off_target_t > 0 && bytes_in_flight*1.25+bytes_newly_acked <= cwnd) # no cwnd increase if window is underutilized # an additional slack of bytes_newly_acked is # added to ensure that CWND growth occurs # even when feedback is sparse cwnd_delta_t = 0; end # apply delta cwnd += cwnd_delta_t # limit cwnd to the maximum number of bytes in flight cwnd = min(cwnd, max_bytes_in_flight*MAX_BYTES_IN_FLIGHT_HEAD_ROOM) cwnd = max(cwnd, MIN_CWND) <CODE ENDS>

CWND is updated differently depending on whether the congestion control is in fast increase state or not, as controlled by the variable in_fast_increase. When in fast increase state, the congestion window is increased with the number of newly acknowledged bytes as long as the window is sufficiently used. Sparse feedback can potentially limit congestion window growth, an additional slack is therefore added, given by the number of newly acknowledged bytes. The congestion window growth when in_fast_increase is false is dictated by the relation between qdelay and qdelay_target, congestion window growth is limited if the window is not used sufficiently. SCReAM calculates the GAIN in a similar way to what is specified in . There are however a few differences. specifies a constant GAIN, this specification however limits the gain when CWND is increased dependent on near congestion state and the relation to the last known max CWND value. specifies that the CWND increase is limited by an additional function controlled by a constant ALLOWED_INCREASE. This additional limitation is removed in this specification. Further the CWND is limited by max_bytes_in_flight and min_cwnd. The limitation of the congestion window by the maximum number of bytes in flight over the last 5 seconds (max_bytes_in_flight) avoids possible over-estimation of the throughput after for example, idle periods. An additional MAX_BYTES_IN_FLIGHT_HEAD_ROOM allows for a slack, to allow for a certain amount of media coder output rate variability.

It is likely that a flow using SCReAM algorithm will have to share congested bottlenecks with other flows that use a more aggressive congestion control algorithm. SCReAM takes care of such situations by adjusting the qdelay_target.

<CODE BEGINS> adjust_qdelay_target(qdelay) qdelay_norm_t = qdelay / QDELAY_TARGET_LOW update_qdelay_norm_history(qdelay_norm_t) # Compute variance qdelay_norm_var_t = VARIANCE(qdelay_norm_history(200)) # Compensation for competing traffic # Compute average qdelay_norm_avg_t = AVERAGE(qdelay_norm_history(50)) # Compute upper limit to target delay oh_t = qdelay_norm_avg_t + sqrt(qdelay_norm_var_t) oh_t *= QDELAY_TARGET_LO if (loss_event_rate > 0.002) # Packet losses detected qdelay_target = 1.5*oh_t else if (qdelay_norm_var_t < 0.2) # Reasonably safe to set target qdelay qdelay_target = oh_t else # Check if target delay can be reduced, this helps to avoid # that the target delay is locked to high values for ever if (oh_t < QDELAY_TARGET_LO) # Decrease target delay quickly as measured queueing # delay is lower than target qdelay_target = max(qdelay_target*0.5,oh_t) else # Decrease target delay slowly qdelay_target *= 0.9 end end end # Apply limits qdelay_target = min(QDELAY_TARGET_HI, qdelay_target) qdelay_target = max(QDELAY_TARGET_LO, qdelay_target) <CODE ENDS>

The qdelay_target is adjusted differently, depending on if qdelay_norm_var_t is above or below a given value.A low qdelay_norm_avg_t value indicates that the qdelay does not change rapidly. It is desired to avoid the case that the qdelay target is increased due to self-congestion, indicated by a changing qdelay and consequently an increased qdelay_norm_var_t. Still it SHOULD be possible to increase the qdelay target if the qdelay continues to be high. This is a simple function with a certain risk of both false positives and negatives. In the simulated LTE test cases it manages competing FTP flows reasonably well at the same time as generally avoiding accidental increases in the qdelay target. The algorithm can however accidentally increase the qdelay target and cause self-inflicted congestion in certain cases. It is therefore RECOMMENDED that the algorithm described in this section is turned off it is deemed unlikely that competing flows occur over the same bottleneck

Lost packet detection is based on the received sequence number list. A reordering window SHOULD be applied to avoid packet reordering triggering loss events. The reordering window is specified as a time unit, similar to the ideas behind RACK (Recent ACKnowledgement) . The computation of the reordering window is made possible by means of a lost flag in the list of transmitted RTP packets. This flag is set if the received sequence number list indicates that the given RTP packet is missing. If a later feedback indicates that a previously lost marked packet was indeed received, then the reordering window is updated to reflect the reordering delay. The reordering window is given by the difference in time between the event that the packet was marked as lost and the event that it was indicated as successfully received.Loss is detected if a given RTP packet is not acknowledged within a time window (indicated by the reordering window) after an RTP packet with higher sequence number was acknowledged.

The basic design principle behind packet transmission in SCReAM is to allow transmission only if the number of bytes in flight is less than the congestion window. There are however two reasons why this strict rule will not work optimally: Bitrate variations: The media frame size is always varying to a larger or smaller extent. A strict rule can lead to that the media bitrate will have difficulties to increase as the congestion window puts a too hard restriction on the media frame size variation. This can lead to occasional queuing of RTP packets in the RTP packet queue that will prevent bitrate increase. Reverse (feedback) path congestion: Especially in transport over buffer-bloated networks, the one way delay in the reverse direction MAY jump due to congestion. The effect of this is that the acknowledgements are delayed with the result that the self-clocking is temporarily halted, even though the forward path is not congested. The send window is adjusted depending on qdelay and its relation to the qdelay target and the relation between the congestion window and the number of bytes in flight. A strict rule is applied when qdelay is higher than qdelay_target, to avoid further queue buildup in the network. For cases when qdelay is lower than the qdelay_target, a more relaxed rule is applied. This allows the bitrate to increase quickly when no congestion is detected while still being able to give a stable behavior in congested situations. The send window is given by the relation between the adjusted congestion window and the amount of bytes in flight according to the pseudo code below.

<CODE BEGINS> calculate_send_window(qdelay, qdelay_target) # send window is computed differently depending on congestion level if (qdelay <= qdelay_target) send_wnd = cwnd+MSS-bytes_in_flight else send_wnd = cwnd-bytes_in_flight end <CODE ENDS>

The send window is updated whenever an RTP packet is transmitted or an RTCP feedback messaged is received.

Packet pacing is used in order to mitigate coalescing i.e. that packets are transmitted in bursts, with the increased risk of more jitter and potentially increased packet loss. The time interval between consecutive packet transmissions is enforced to be equal to or higher than t_pace where t_pace is given by the equations below :

<CODE BEGINS> pace_bitrate = max (RATE_PACE_MIN, cwnd* 8 / s_rtt) t_pace = rtp_size * 8 / pace_bitrate <CODE ENDS>

rtp_size is the size of the last transmitted RTP packet, s_rtt is the smoothed round trip time. RATE_PACE_MIN=50000 is the minimum pacing rate.

Fast increase can resume in order to speed up the bitrate increase in case congestion abates. The condition to resume fast increase (in_fast_increase = true) is that qdelay_trend is less than QDELAY_TREND_LO for T_RESUME_FAST_INCREASE seconds or more.

The media rate control algorithm is executed at regular intervals RATE_ADJUSTMENT_INTERVAL, with the exception of a prompt reaction to loss events. The media rate control operates based on the size of the RTP packet send queue and observed loss events. In addition, qdelay_trend is also considered in the media rate control to reduce the amount of induced network jitter. The role of the media rate control is to strike a reasonable balance between a low amount of queuing in the RTP queue(s) and a sufficient amount of data to send in order to keep the data path busy. A too cautious setting leads to possible under-utilization of network capacity leading to the flow being starved out by other more opportunistic traffic. On the other hand, a too aggressive setting leads to increased jitter. The target_bitrate is adjusted depending on the congestion state. The target bitrate can vary between a minimum value (TARGET_BITRATE_MIN) and a maximum value (TARGET_BITRATE_MAX). TARGET_BITRATE_MIN SHOULD be chosen to a low enough value to avoid RTP packets being queued up when the network throughput becomes low. The sender SHOULD also be equipped with a mechanism that discards RTP packets in cases where the network throughput becomes very low and RTP packets are excessively delayed. For the overall bitrate adjustment, two network throughput estimates are computed : rate_transmit: The measured transmit bitrate. rate_ack: The ACKed bitrate, i.e. the volume of ACKed bits per second. Both estimates are updated every 200ms. The current throughput, current_rate, is computed as the maximum value of rate_transmit and rate_ack. The rationale behind the use of rate_ack in addition to rate_transmit is that rate_transmit is affected also by the amount of data that is available to transmit, thus a lack of data to transmit can be seen as reduced throughput that MAY itself cause an unnecessary rate reduction. To overcome this shortcoming; rate_ack is used as well. This gives a more stable throughput estimate. The rate change behavior depends on whether a loss or ECN event has occurred and if the congestion control is in fast increase or not.

<CODE BEGINS> # The target_bitrate is updated at a regular interval according # to RATE_ADJUST_INTERVAL on loss: # Loss event detected target_bitrate = max(BETA_R* target_bitrate, TARGET_BITRATE_MIN) exit on ecn_mark: # ECN event detected target_bitrate = max(BETA_ECN* target_bitrate, TARGET_BITRATE_MIN) exit ramp_up_speed_t = min(RAMP_UP_SPEED, target_bitrate/2.0) scale_t = (target_bitrate - target_bitrate_last_max)/ target_bitrate_last_max scale_t = max(0.2, min(1.0, (scale_t*4)^2)) # min scale_t value 0.2 as the bitrate should be allowed to # increase at least slowly --> avoid locking the rate to # target_bitrate_last_max if (in_fast_increase = true) increment_t = ramp_up_speed_t*RATE_ADJUST_INTERVAL increment_t *= scale_t target_bitrate += increment_t else current_rate_t = max(rate_transmit, rate_ack) # compute a bitrate change delta_rate_t = current_rate_t*(1.0-PRE_CONGESTION_GUARD* queue_delay_trend)-TX_QUEUE_SIZE_FACTOR *rtp_queue_size # limit a positive increase if close to target_bitrate_last_max if (delta_rate_t > 0) delta_rate_t *= scale_t delta_rate_t = min(delta_rate_t,ramp_up_speed_t*RATE_ADJUST_INTERVAL) end target_bitrate += delta_rate_t # force a slight reduction in bitrate if RTP queue # builds up rtp_queue_delay_t = rtp_queue_size/current_rate_t if (rtp_queue_delay_t > RTP_QDELAY_TH) target_bitrate *= TARGET_RATE_SCALE_RTP_QDELAY end end rate_media_limit_t = max(current_rate_t, max(rate_media,rtp_rate_median)) rate_media_limit_t *= (2.0-qdelay_trend_mem) target_bitrate = min(target_bitrate, rate_media_limit_t) target_bitrate = min(TARGET_BITRATE_MAX, max(TARGET_BITRATE_MIN,target_bitrate)) <CODE ENDS>

In case of a loss event the target_bitrate is updated and the rate change procedure is exited. Otherwise the rate change procedure continues. The rationale behind the rate reduction due to loss is that a congestion window reduction will take effect, a rate reduction pro actively avoids RTP packets being queued up when the transmit rate decreases due to the reduced congestion window. A similar rate reduction happens when ECN events are detected. The rate update frequency is limited by RATE_ADJUST_INTERVAL, unless a loss event occurs. The value is based on experimentation with real life limitations in video coders taken into account . A too short interval is shown to make the video coder internal rate control loop more unstable, a too long interval makes the overall congestion control sluggish. When in fast increase state (in_fast_increase=true), the bitrate increase is given by the desired ramp-up speed (RAMP_UP_SPEED) . The ramp-up speed is limited when the target bitrate is low to avoid rate oscillation at low bottleneck bitrates. The setting of RAMP_UP_SPEED depends on preferences, a high setting such as 1000kbps/s makes it possible to quickly get high quality media, this is however at the expense of a increased jitter, which can manifest itself as e.g. choppy video rendering. When in_fast_increase is false, the bitrate increase is given by the current bitrate and is also controlled by the estimated RTP queue and the qdelay trend, thus it is sufficient that an increased congestion level is sensed by the network congestion control to limit the bitrate. The target_bitrate_last_max is updated when congestion is detected. Finally the target_bitrate is enforced to be within the defined min and max values. The aware reader MAY notice the dependency on the qdelay in the computation of the target bitrate, this manifests itself in the use of the qdelay_trend. As these parameters are used also in the network congestion control one MAY suspect some odd interaction between the media rate control and the network congestion control, this is in fact the case if the parameter PRE_CONGESTION_GUARD is set to a high value. The use of qdelay_trend in the media rate control is solely to reduce jitter, the dependency can be removed by setting PRE_CONGESTION_GUARD=0, the effect is a somewhat faster rate increase after congestion, at the expense of increased jitter in congested situations.

The simple task of the SCReAM receiver is to feedback acknowledgements of received packets and total ECN count to the SCReAM sender, in addition, the receive time of the RTP packet with the highest sequence number is echoed back. Upon reception of each RTP packet the receiver MUST maintain enough information to send the aforementioned values to the SCReAM sender via a RTCP transport layer feedback message. The frequency of the feedback message depends on the available RTCP bandwidth. The requirements on the feedback elements and the feedback interval is described.

SCReAM requires the following elements for its basic functionality, i.e. only including features that are strictly necessary in order to make SCReAM function. ECN is not included as basic functionality as it regarded as an additional feature that is not strictly necessary even though it can improve quality of experience quite considerably. A list of received RTP packets. This list SHOULD be sufficiently long to cover all received RTP packets. This list can be realized with the Loss RLE report block in . A wall clock timestamp corresponding to the received RTP packet with the highest sequence number is required in order to compute the qdelay. This can be realized by means of the Packet Receipt Times Report Block in . begin_seq MUST be set to the highest received (possibly wrapped around) sequence number, end_seq MUST be set to begin_seq+1 % 65536. The timestamp clock MAY be set according to i.e. equal to the RTP timestamp clock. Detailed individual packet receive times is not necessary as SCReAM does currently not describe how this can be used. The basic feedback needed for SCReAM involves the use of the Loss RLE report block and the Packet Receipt Times block defined in .

0 1 2 3 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ |V=2|P|reserved | PT=XR=207 | length | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | SSRC | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | BT=2 | rsvd. | T=0 | block length | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | SSRC of source | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | begin_seq | end_seq | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | chunk 1 | chunk 2 | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ : ... : +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | chunk n-1 | chunk n | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | BT=3 | rsvd. | T=0 | block length | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | SSRC of source | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | begin_seq | end_seq | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | Receipt time of packet begin_seq | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

In a typical use case, no more than four Loss RLE chunks SHOULD be needed, thus the feedback message will be 44bytes. It is obvious from the figure that there is a lot of redundant information in the feedback message. A more optimized feedback format, including the additional feedback elements listed below, could reduce the feedback message size a bit. Additional feedback elements that can improve the performance of SCReAM are: Accumulated number of ECN-CE marked packets (n_ECN). This can for instance be realized with the ECN Feedback Report Format in . The given feedback report format is actually a slight overkill as SCReAM would do quite well with only a counter that increments by one for each received packet with the ECN-CE code point set. The more bulky format MAY be nevertheless be useful for e.g ECN black-hole detection.

SCReAM benefits from a relatively frequent feedback. The feedback interval depends on the media bitrate. At low bitrates it is sufficient with a feedback interval of 100 to 400ms, while at high bitrates a feedback interval of roughly 20ms is to prefer. The numbers above can be formulated as feedback interval function that can be useful for the computation of the desired RTCP bandwidth. The following equation expresses the feedback rate:

rate_fb = min(50,max(2.5,rate_media/10000))

rate_media is the RTP media bitrate expressed in [bits/s], rate_fb is the feedback rate expressed in [packets/s]. Converted to feedback interval we get:

fb_int = 1.0/min(50,max(2.5,rate_media/10000))

The transmission interval is not critical, this means that in the case of multi-stream handling between two hosts, the feedback for two or more SSRCs can be bundled to save UDP/IP overhead, the final realized feedback interval SHOULD however not exceed 2*fb_int in such cases meaning that a scheduled feedback transmission event should not be delayed more that fb_int. SCReAM works with AVPF regular mode, immediate or early mode is not REQUIRED by SCReAM but MAY nonetheless be useful for e.g RTCP messages not directly related to SCReAM, such as those specified in . It is RECOMMENDED to use reduced size RTCP where regular full compound RTCP transmission is controlled by trr-int as described in .

This section covers a few discussion points Clock drift: SCReAM can suffer from the same issues with clock drift as is the case with LEDBAT . Section A.2 in however describes ways to mitigate issues with clock drift. Support for alternate ECN semantics: This specification adopts the proposal in to reduce the congestion window less when ECN based congestion events are detected. Future work on Low Loss Low Latency for Scalable throughput (L4S) MAY lead to updates in a future RFC that describes SCReAM support for L4S. A new RFC4585 transport layer feedback message MAY to be standardized if the use of the already existing RTCP extensions as described in is not deemed sufficient. The target bitrate given by SCReAM depicts the bitrate including RTP and FEC overhead. The media encoder SHOULD take this overhead into account when the media bitrate is set. This means that the media coder bitrate SHOULD be computed as

media_rate = target_bitrate - rtp_plus_fec_overhead_bitrate

It is not strictly necessary to make a 100% perfect compensation for the overhead as the SCReAM algorithm will inherently compensate for moderate errors. Under-compensation of the overhead has the effect of increasing jitter while overcompensation will have the effect of causing the bottleneck link to become under-utilized.

[Editor's note: Please remove the whole section before publication, as well reference to RFC 7942] This section records the status of known implementations of the protocol defined by this specification at the time of posting of this Internet-Draft, and is based on a proposal described in . The description of implementations in this section is intended to assist the IETF in its decision processes in progressing drafts to RFCs. Please note that the listing of any individual implementation here does not imply endorsement by the IETF. Furthermore, no effort has been spent to verify the information presented here that was supplied by IETF contributors. This is not intended as, and MUST NOT be construed to be, a catalog of available implementations or their features. Readers are advised to note that other implementations MAY exist. According to , "this will allow reviewers and working groups to assign due consideration to documents that have the benefit of running code, which may serve as evidence of valuable experimentation and feedback that have made the implemented protocols more mature. It is up to the individual working groups to use this information as they see it".

The SCReAM algorithm has been implemented in the OpenWebRTC project , an open source WebRTC implementation from Ericsson Research. This SCReAM implementation is usable with any WebRTC endpoint using OpenWebRTC. Organization : Ericsson Research, Ericsson. Name : OpenWebRTC gst plug-in. Implementation link : The GStreamer plug-in code for SCReAM can be found at github repository The wiki (https://github.com/EricssonResearch/openwebrtc/wiki) contains required information for building and using OpenWebRTC. Coverage : The code implements the specification in this memo. The current implementation has been tuned and tested to adapt a video stream and does not adapt the audio streams. Implementation experience : The implementation of the algorithm in the OpenWebRTC has given great insight into the algorithm itself and its interaction with other involved modules such as encoder, RTP queue etc. In fact it proves the usability of a self-clocked rate adaptation algorithm in the real WebRTC system. The implementation experience has led to various algorithm improvements both in terms of stability and design. The current implementation use an n_loss counter for lost packets indication, this is subject to change in later versions to a list of received RTP packets. Contact : irc://chat.freenode.net/openwebrtc

Organization : Ericsson Research, Ericsson. Name : SCReAM. Implementation link : A C++ implementation of SCReAM is available at. The code includes full support for congestion control, rate control and multi stream handling, it can be integrated in web clients given the addition of extra code to implement the RTCP feedback and RTP queue(s). The code also includes a rudimentary implementation of a simulator that allows for some initial experiments. An additional experiment with SCReAM in a remote control arrangement is also documented. Coverage : The code implements the specification in this memo. Contact : ingemar.s.johansson@ericsson.com

SCReAM has been evaluated in a number of different ways, most of the evaluation has been in simulator. The OpenWebRTC implementation work involved extensive testing with artificial bottlenecks with varying bandwidths and using two different video coders (OpenH264 and VP9), the experience of this lead to further improvements of the media rate control logic. Further experiments are preferably done by means of implementation in real clients and web browsers. RECOMMENDED experiments are: Trials with various access technologies: EDGE/3G/4G, WiFi, DSL. Some experiments have already been carried out with LTE access, see e.g. and Trials with different kinds of media: Audio, Video, slide show content. Evaluation of multi stream handling in SCReAM. Evaluation of functionality of competing flows compensation mechanism: Evaluate how SCReAM performs with competing TCP like traffic and to what extent the competing flows compensation causes self-inflicted congestion. Determine proper parameters: A set of default parameters are given that makes SCReAM work over a reasonably large operation range, however for instance for very low or very high bitrates it MAY be necessary to use different values for instance for the RAMP_UP_SPEED.

We would like to thank the following persons for their comments, questions and support during the work that led to this memo: Markus Andersson, Bo Burman, Tomas Frankkila, Frederic Gabin, Laurits Hamm, Hans Hannu, Nikolas Hermanns, Stefan Haakansson, Erlendur Karlsson, Daniel Lindström, Mats Nordberg, Jonathan Samuelsson, Rickard Sjöberg, Robert Swain, Magnus Westerlund, Stefan Aalund. Many additional thanks to RMCAT chairs Karen E. E. Nielsen and Mirja Kühlewind for patiently reading, suggesting improvements and also for asking all the difficult but necessary questions. Thanks to Stefan Holmer, Xiaoqing Zhu, Safiqul Islam and David Hayes for the additional review of this document. Thanks to Ralf Globisch for taking time to try out SCReAM in his challenging low bitrate use cases.

There is currently no request to IANA

The feedback can be vulnerable to attacks similar to those that can affect TCP. It is therefore RECOMMENDED that the RTCP feedback is at least integrity protected. Furthermore, as SCReAM is self-clocked, a malicious middlebox can drop RTCP feedback packets and thus cause the self-clocking in SCReAM to stall. This attack is however mitigated by the minimum send rate maintained by SCReAM when no feedback is received.

A list of changes: WG-08 to WG-09: Updated based shepherd review by Martin Stiemerling, Q-bit semantics are removed as this is superfluous for the moment. Pacing and RTCP considerations are moved up from the appendix, FEC discussion moved to discussion section. WG-07 to WG-08: Avoid draft expiry WG-06 to WG-07: Updated based on WGLC review by David Hayes and Safiqul Islam WG-05 to WG-06: Added list of suggested experiments WG-04 to WG-05: Congestion control and rate control simplified somewhat WG-03 to WG-04: Editorial fixes WG-02 to WG-03: Review comments from Stefan Holmer and Xiaoqing Zhu addressed, owd changed to qdelay for clarity. Added appendix section with RTCP feedback requirements, including a suggested basic feedback format based Loss RLE report block and the Packet Receipt Times blocks in . Loss detection added as a section. Transmission scheduling and packet pacing explained in appendix. Source quench semantics added to appendix. WG-01 to WG-02: Complete restructuring of the document. Moved feedback message to a separate draft. WG-00 to WG-01 : Changed the Source code section to Implementation status section. -05 to WG-00 : First version of WG doc, moved additional features section to Appendix. Added description of prioritization in SCReAM. Added description of additional cap on target bitrate -04 to -05 : ACK vector is replaced by a loss counter, PT is removed from feedback, references to source code added -03 to -04 : Extensive changes due to review comments, code somewhat modified, frame skipping made optional -02 to -03 : Added algorithm description with equations, removed pseudo code and simulation results -01 to -02 : Updated GCC simulation results -00 to -01 : Fixed a few bugs in example code

University College London

The SCReAM algorithm makes a good distinction between network congestion control and the media rate control. This is easily extended to many streams, in which case RTP packets from two or more RTP queues are scheduled at the rate permitted by the network congestion control. The scheduling can be done by means of a few different scheduling regimes. For example the method applied in can be used. The implementation of SCReAM use credit based scheduling. In credit based scheduling, credit is accumulated by queues as they wait for service and are spent while the queues are being serviced. For instance, if one queue is allowed to transmit 1000bytes, then a credit of 1000bytes is allocated to the other unscheduled queues. This principle can be extended to weighted scheduling in which case the credit allocated to unscheduled queues depends on the relative weights.

The autocorrelation function is computed over a vector of values. Let x be a vector constituting N values, the biased autocorrelation function for a given lag=k for the vector x is given by .

n=N-k R(x,k) = SUM x(n)*x(n+k) n=1