| RMCAT WG | I. Johansson |
| Internet-Draft | Z. Sarker |
| Intended status: Experimental | Ericsson AB |
| Expires: April 29, 2018 | October 26, 2017 |
Self-Clocked Rate Adaptation for Multimedia
draft-ietf-rmcat-scream-cc-13
This memo describes a rate adaptation algorithm for conversational media services such as interactive 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.
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 April 29, 2018.
Copyright (c) 2017 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 Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License.
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 have to 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 can 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 [QoS-3GPP] 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 implements congestion control for RTP streams [RFC3550]. While SCReAM was originally devised for WebRTC (Web Real-Time Communication) [RFC7478], it can also be used for other applications where congestion control of RTP streams is necessary. SCReAM is based on the self-clocking principle of TCP and uses techniques similar to what is used in the LEDBAT based rate adaptation algorithm [RFC6817]. SCReAM is not entirely self-clocked as it augments self-clocking with pacing and a minimum send rate.
SCReAM can take advantage of ECN (Explicit Congestion Notification) in cases where ECN is supported by the network and the hosts. ECN is however not required for the basic congestion control functionality in SCReAM.
[I-D.ietf-rmcat-wireless-tests] 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 can 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 can 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 aims to 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 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 [RFC2119].
The core SCReAM algorithm has similarities to the concepts of self-clocking used in TFWC [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 [Packet-conservation].
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 [RFC6817]. 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. [LEDBAT-delay-impact] 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 bitrate. This gives two useful properties when LEDBAT is used with SCReAM that help to avoid the issues described in [LEDBAT-delay-impact]:
It is sufficient that any of the two conditions above is fulfilled to make the base delay update properly. Furthermore [LEDBAT-delay-impact] 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:
The above mentioned features will be described in more detail in sections Section 3.1 to Section 3.3. The full details are described in Section 4.
+---------------------------+
| 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 |
+------------+
Figure 1: SCReAM sender functional view
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. Figure 1 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 [RFC6817] 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. Failure to meet this requirement leads to malfunction in the SCReAM congestion control algorithm due to incorrect estimation of the network queue delay.
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. Figure 1 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 [SCReAM-CPP-implementation].
Media frames are encoded and forwarded to the RTP queue (1) in Figure 1. 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 [ ].
The values within () indicate initial values.
This section explains the network congestion control, it contains two main functions:
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.
When the sender receives RTCP feedback, the qdelay is calculated as outlined in [RFC6817]. 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'. As mentioned in Section 7 , calculation of the proper congestion window and media bitrate may benefit from additional optimizations for handling of very high and very low bitrates, and from additional damping to handle periodic packet bursts. Some such optimizations are implemented in [SCReAM-CPP-implementation], but they do not form part of the specification of SCReAM at this time.
<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)
# Compute the average of the values in qdelay_fraction_hist
avg_t = average(qdelay_fraction_hist)
# R is an autocorrelation function of qdelay_fraction_hist,
# with the mean (DC component) removed, at lag K
# The subtraction of the scalar avg_t from
# qdelay_fraction_hist is performed element-wise
a_t = R(qdelay_fraction_hist-avg_t,1)/
R(qdelay_fraction_hist-avg_t,0)
# Calculate qdelay trend
qdelay_trend = min(1.0,max(0.0,a_t*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_t’ has positive values if qdelay shows an increasing or decreasing trend, thus an indication of congestion is obtained before the qdelay target is reached. As a side effect, also the case that qdelay decreases is taken as a sign of congestion, experiments have however shown that this is beneficial as varying queue delay up or down is an indication that the transmit rate is very close to the path capacity.
n=N-k
R(x,k) = SUM x(n)*x(n+k)
n=1
The autocorrelation function 'R' is defined as follows. Let x be a vector constituting N values, the biased autocorrelation function for a given lag=k for the vector x is given by.
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 that could 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 Section 4.1.2.4. Once a loss event is detected, further detected lost RTP packets SHOULD be 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 [I-D.ietf-tcpm-alternativebackoff-ecn] 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 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 [RFC6817]. However, [RFC6817] 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, examples are large FTP flows using loss based congestion control. The worst condition occurs when the bottleneck queues are of tail-drop type with a large buffer size. SCReAM takes care of such situations by adjusting the qdelay_target when loss based flows are detected, as given by the pseudo code below.
<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
new_target_t = qdelay_norm_avg_t + sqrt(qdelay_norm_var_t)
new_target_t *= QDELAY_TARGET_LO
if (loss_event_rate > 0.002)
# Packet losses detected
qdelay_target = 1.5*new_target_t
else
if (qdelay_norm_var_t < 0.2)
# Reasonably safe to set target qdelay
qdelay_target = new_target_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 (new_target_t < QDELAY_TARGET_LO)
# Decrease target delay quickly as measured queueing
# delay is lower than target
qdelay_target = max(qdelay_target*0.5,new_target_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>
Two temporary variables are calculated. qdelay_norm_avg_t is the long term average queue delay, qdelay_norm_var_t is the long term variance of the queue delay. A high qdelay_norm_var_t indicates that the queue delay changes, this can be an indication of reduced bottleneck bandwidth or that a competing flow has just entered. Thus, it indicates that it is not safe to adjust the queue delay target.
A low qdelay_norm_var_t indicates that the queue delay is relatively stable, the reason can be that the queue delay is low but it can also be an indication that a competing flow is filling up the bottleneck to the limit where packet losses may start to occur, in which case the queue delay will stay relatively high for a longer time.
The queue delay target is allowed to be increased if, either the loss event rate is above a given threshold or that qdelay_norm_var_t is low. Both these conditions indicate that a competing flow may be present. In all other cases the queue delay target is decreased.
The function that adjusts the qdelay_target is simple and has a certain risk to produce both false positive and negatives, The case that self-inflicted congestion by the SCReAM algorithm may be falsely interpreted as the presence of competing loss based FTP flows is a false positive. The opposite case where the algorithm fails to detect the presence of a competing FTP flow is a false negative.
Extensive simulations have shown that the algorithm performs well in LTE test cases and that it also performs well in simple bandwidth limited bottleneck test cases with competing FTP flows. It can however not be completely ruled out that this algorithm can fail. Especially the false positives can be problematic as the end to end delay can increase dramatically if the target queue delay is increased by accident as a result of self-inflicted congestion.
If it is deemed unlikely that competing flows occur over the same bottleneck, the algorithm described in this section MAY be turned off. One such case can be QoS enabled bearers in 3GPP based access such as LTE. However, when sending over the Internet, often the network conditions are not known for sure and it is in general not possible to make safe assumptions on how a network is used and whether or not competing flows share the same bottleneck. Therefore turning this algorithm off must be considered with caution as that can lead to basically zero throughput if competing with other, loss based, traffic.
Lost packet detection is based on the received sequence number list. A reordering window SHOULD be applied to avoid that packet reordering triggers loss events.
The reordering window is specified as a time unit, similar to the ideas behind RACK (Recent ACKnowledgement) [I-D.ietf-tcpm-rack]. 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:
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. Packet pacing also mitigates possible issues with queue overflow due to key-frame generation in video coders. 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 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 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 [I-D.ietf-rmcat-coupled-cc] can be used. The implementation of SCReAM [SCReAM-CPP-implementation] 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 latter is also implemented in [SCReAM-CPP-implementation].
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 that the flow can become 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 that RTP packets become queued up when the network throughput is reduced. 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 :
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 can 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 [SCReAM-CPP-implementation]. A too short interval is shown to make the rate control loop in video coders 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.
The following feedback elements are REQUIRED for the basic functionality in SCReAM.
The basic feedback needed for SCReAM involves the use of the Loss RLE report block and the Packet Receipt Times block defined in Figure 2.
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 |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
Figure 2: Basic feedback message for SCReAM, based on RFC3611
In a typical use case, no more than four Loss RLE chunks are 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:
SCReAM benefits from a relatively frequent feedback. It is RECOMMENDED that a SCReAM implementation follows the guidelines below.
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, at very high bitrates, even shorter feedback intervals MAY be needed in order to keep the self-clocking in SCReAM working well. One piece of evidence of a too sparse feedback is that the SCReAM implementation cannot reach high bitrates, even in uncongested links. A more frequent feedback might solve this issue.
rate_fb = min(50,max(2.5,rate_media/10000))
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:
fb_int = 1.0/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:
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 can nonetheless be useful for e.g RTCP messages not directly related to SCReAM, such as those specified in [RFC4585]. It is RECOMMENDED to use reduced size RTCP [RFC5506] where regular full compound RTCP transmission is controlled by trr-int as described in [RFC4585].
media_rate = target_bitrate - rtp_plus_fec_overhead_bitrate
This section covers a few discussion points
[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 [RFC7942]. 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 [RFC7942], "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 [OpenWebRTC], an open source WebRTC implementation from Ericsson Research. This SCReAM implementation is usable with any WebRTC endpoint using OpenWebRTC.
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:
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, Robert Hedman for finding a few additional flaws in the running code, and Gustavo Garcia and 'miseri' for code contributions.
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: