Self-Clocked Rate Adaptation for Multimedia

1.1. Wireless (LTE and 5G) Access Properties

[RFC8869] describes the complications that can be observed in wireless environments. Wireless access such as LTE and 5G typically cannot 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 5G can support bitrates well above 100 Mbps, 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 (e.g., at 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, a 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.¶

1.2. Why is it a self-clocked algorithm?

Self-clocked congestion control algorithms provide a benefit over their 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; it operates on a shorter time scale (1 RTT). The benefits of self-clocking are also elaborated upon in [TFWC]. The self-clocking however acts more like an emergency break as bytes in flight can exceed the congesion window only to a certain degree. The rationale is to be able to transmit large video frames and avoid that they are unnecessarily queued up on the sender side, but still prevent a bootleneck queue¶

A rate-based congestion control algorithm typically adjusts the rate based on delay and loss. The congestion detection needs to be done with a certain time lag to avoid overreaction to spurious congestion events such as delay spikes. Despite the fact that there are two or more congestion indications, the outcome is 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.¶

1.3. Comparison with LEDBAT and TFWC in TCP

The core SCReAM algorithm, which is still similar in SCReAMv2, has similarities to the concepts of self-clocking used in TCP-friendly window-based congestion control [TFWC] and follows the packet conservation principle. The packet conservation principle is described as a key factor behind the protection of networks from congestion [Packet-conservation].¶

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 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 facts that the source is rate limited and that the RTP queue must be 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 yields two useful properties when LEDBAT is used with SCReAM; they help to avoid the issues described in [LEDBAT-delay-impact]:¶

1. There is always a certain probability that SCReAM is short of data to transmit; this means that the network queue will become empty every once in a while.¶

2. The max video bitrate can be lower than the link capacity. If the max video bitrate is 5 Mbps and the capacity is 10 Mbps, then the network queue will become empty.¶

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. In SCReAM, these short-lived flows will cause the self-clocking to slow down, thereby building up the RTP queue; in turn, this results in a reduced media video bitrate. Thus, SCReAM slows the bitrate more when there are competing short-lived flows than the traditional use of LEDBAT does. The basic functionality in the use of LEDBAT in SCReAM is quite simple; however, there are a few steps in order to make the concept work with conversational media:¶

• Addition of a media rate control function.¶

• Congestion window validation techniques. The congestion window is used as a basis for the target bitrate calculation. For that reason, various actions are taken to avoid that the congestion window grows too much beyond the bytes in flight. Additional contraints are applied when in congested state and when the max target bitrate is reached.¶

• Use of inflection points in the congestion window calculation to achieve reduced delay jitter (when L4S is not active).¶

• 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 Section Section 4.¶

The SCReAM/SCReAMv2 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 the sender and receiver in order to compute the qdelay. However, it is 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/SCReAMv2 congestion control algorithm due to incorrect estimation of the network queue delay. Use of [RFC8888] as feedback ensures that the same time base is used in sender and receiver.¶

3.1. Network Congestion Control

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 sender calculates the congestion window based on the feedback from the RTP receiver. The feedback is timestamp and ECN echo for individual RTP packets.¶

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 however not an absolute limit as a slack is given to efficiently transmit large video frames.¶

The congestion window seeks to increase by at least one segment per RTT and this increase regardless congestion occurs or not, the congestion window increase is restriced or relaxed based on the current value of the congestion window and the time elapsed since last congestion event. The congestion window update is increased by one MSS (maximum known RTP packet size) per RTT with some variation based on congestion window size and time elapsed since the last congestion event. Multiplicative increase allows the congestion to increase by a fraction of cwnd when congestion has not occured for a while. The congestion window is thus an adaptive multiplicative increase that is mainly additive increase when steady state is reached but allows a faster convergence to a higher link speed.¶

Congestion window reduction is triggered by:¶

• Packet loss or Classic ECN marking is detected : The congestion window is reduced by a predetermined fraction.¶

• Estimated queue delay exceeds a given threshold : The congestion window is reduced given by how much the delay exceeds the threshold.¶

• L4S ECN marking detected : The congestion window is reduced in proportion to the fraction of packets that are marked (scalable congestion control).¶

3.2. Sender Transmission Control

The sender transmission control limits the output of data, given by the relation between the number of bytes in flight and the congestion window. The congestion window is however not a hard limit, additional slack is given to avoid that RTP packets are queued up unnecessarily on the sender side. This means that the algoritm prefers to build up a queue in the network rather than on the sender side. Additional congestion that this causes will reflect back and cause a reduction of 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; rather, they are transmitted in intervals given by the packet size and the estimated link throughput. Packets are generally paced at a higher rate than the target bitrate, this makes it possible to transmit occasionally larger video frames in a timely manner.¶

3.3. Media Rate Control

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 rate is calculated based on the congestion window and RTT. For the case that multiple streams are enabled, the media rate among the streams is distrubuted according to relative priorities.¶

In cases where the sender's frame queues increase rapidly, such as in the case of a Radio Access Type (RAT) handover, the SCReAMv2 sender MAY 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 is outside the scope of this algorithm description.¶

4.1. Constants and variables

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

4.2. Network Congestion Control

This section explains the network congestion control, which performs two main functions:¶

• Computation of congestion window at the sender: This 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.¶

SCReAMv2 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 an 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, the 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.¶

bytes_in_flight_ratio is calculated as the ratio between bytes_flight and cwnd. This value should be computed at the beginning of the ACK processing. cwnd_ratio is computed as the relation between MSS and cwnd.¶

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_ce is, similar to bytes_newly_acked, a counter of bytes newly acked with the extra condition that they are ECN-CE marked. The bytes_newly_acked and bytes_newly_acked_ce are 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. With an indication if packets are ECN-CE marked.¶

• The wall-clock timestamp corresponding to the received RTP packet with the highest sequence number.¶

It is recommended to use RFC8888 [RFC8888] for the feedback as it supports the feedback elements described above.¶

When the sender receives RTCP feedback, the qdelay is calculated as outlined in [RFC6817]. A qdelay sample is obtained for each received acknowledgement. A number of variables are updated as illustrated by the pseudocode below; temporary variables are appended with '_t'. Division operation is always floating point unless otherwise noted. l4s_alpha is calculated based in number of packets delivered (and marked). This makes calculation of L4S alpha more accurate at very low bitrates, given that the tail RTP packet in a video frame is often smaller than MSS.¶

    packets_delivered_this_rtt += packets_acked
    packets_marked_this_rtt += packets_acked_ce
    if (now - last_update_l4s_alpha_time >= s_rtt)
      # l4s_alpha is calculated from packets marked istf bytes marked
      fraction_marked_t = packets_marked_this_rtt/
                          packets_delivered_this_rtt
      l4s_alpha = L4S_AVG_G*fraction_marked_t + (1.0-L4S_AVG_G)*l4S_alpha

      last_update_l4s_alpha_time = now
      packets_delivered_this_rtt = 0
      packets_marked_this_rtt = 0
    end

    if (now - last_update_qdelay_avg_time >= s_rtt)
      # qdelay_avg is updated with a slow attack, fast decay EWMA filter
      if (qdelay < qdelay_avg)
        qdelay_avg = qdelay
      else
        qdelay_avg = QDELAY_AVG_G*qdelay + (1.0-QDELAY_AVG_G)*qdelay_avg
      end
      last_update_qdelay_avg_time = now
    end

      if (now - last_congestion_detected_time >= s_rtt)
        if (loss detected)
          is_loss_t = true
        end
        if (packets marked)
          is_ce_t = true
        end
        if (qdelay > qdelay_target/2)
          # It is expected that l4s_alpha is below a given value,
          l4_alpha_lim_t = 2 / target_bitrate * MSS * 8 / s_rtt
          if (l4s_alpha < l4_alpha_lim_t || !l4s_active)
            # L4S does not seem to be active
            l4s_alpha_v_t = min(1.0, max(0.0,
               (qdelay_avg - qdelay_target / 2) /
               (qdelay_target / 2)));
            is_virtual_ce_t = true
          end
        end
      end

      if (is_loss_t || is_ce_t || is_virtual_ce_t)
        if (now - last_cwnd_i_update_time > 0.25)
          last_cwnd_i_update_time = now
          cwnd_i = cwnd
        end
      end


      # Scale factor for cwnd update
      cwnd_scale_factor_t =
        (LOW_CWND_SCALE_FACTOR + (MUL_INCREASE_FACTOR  * cwnd) / MSS)

      # Either loss, ECN mark or increased qdelay is detected
      if (is_loss_t)
        # Loss is detected
        cwnd = cwnd * BETA_LOSS
      end
      if (is_ce_t)
        # ECN-CE detected
        if (IS_L4S)
          # L4S mode
          backoff_t = l4s_alpha / 2
          # Increase stability for very small cwnd
          backOff_t *= min(1.0, cwnd_scale_factor_t)
          backOff_t *= max(0.8, 1.0 - cwnd_ratio * 2)

          if (now - last_congestion_detected_time > 5)
            # A long time since last congested because link throughput
            # exceeds max video bitrate.
            # There is a certain risk that CWND has increased way above
            # bytes in flight, so we reduce it here to get it better on
            # track and thus the congestion episode is shortened
            cwnd = min(cwnd, max_bytes_in_flight_prev)

            # Also, we back off a little extra if needed
            # because alpha is quite likely very low
            # This can in some cases be an over-reaction
            # but as this function should kick in relatively seldom
            # it should not be to too big concern
            backoff_t = max(backoff_t, 0.25)

            # In addition, bump up l4sAlpha to a more credible value
            # This may over react but it is better than
            # excessive queue delay
            l4sAlpha = 0.25
          end
          cwnd = (1.0 - backoff_t) * cwnd
        else
          # Classic ECN mode
          cwnd = cwnd * BETA_ECN
        end
      end
      if (is_virtual_ce_t)
        backoff_t = l4s_alpha_v_t / 2
        cwnd = (1.0 - backoff_t) * cwnd
      end
      cwnd = max(MIN_CWND, cwnd)

      if (is_loss_t || is_ce_t || is_virtual_ce_t)
        last_congestion_detected_time = now
      end

      # Additional factor for cwnd update
      post_congestion_scale_t = max(0.0, min(1.0,
        (now - last_congestion_detected_time) / POST_CONGESTION_DELAY))

      # Calculate bytes acked that are not CE marked
      bytes_newly_acked_minus_ce_t = bytes_newly_acked-
                                     bytes_newly_acked_ce

      increment_t = bytes_newly_acked_minus_ce_t*cwnd_ratio

      # Reduce increment for small RTTs
      tmp_t = min(1.0, s_rtt / VIRTUAL_RTT)
      increment *= tmp_t * tmp_t

      if (!is_l4s_active)
        # The increment is scaled down for more cautious
        # ramp-up around the last known congestion window
        # when congestion last occured.
        # This is only applied when L4S is inactive
        scl_t = 1.0
        scl_t = (cwnd - cwnd_i) / cwnd_i
        scl_t *= 4
        scl_t = scl_t * scl_t
        scl_t = max(0.1, min(1.0, scl_t))
        increment_t *= scl_t
      end

      # Slow down CWND increase when CWND is only a few MSS
      # This goes hand in hand with that the down scaling is also
      # slowed down then. Cwnd increase can be as slow as
      # LOW_CWND_SCALE_FACTOR*MSS per RTT
      float tmp_t = cwnd_scale_factor_t

      # Further limit multiplicative increase when congestion occured
      # recently.
      if (tmp_t > 1.0)
        tmp_t = 1.0 + ((tmp_t - 1.0) * post_congestion_scale_t);
      end
      increment *= tmp_t

      # Increase CWND only if bytes in flight is large enough
      # Quite a lot of slack is allowed here to avoid that bitrate
      # locks to low values.
      max_allowed_t = MSS + max(max_bytes_in_flight,
        max_bytes_in_flight_prev) * BYTES_IN_FLIGHT_HEAD_ROOM
      int cwnd_t = cwnd + increment_t
      if (cwnd_t <= max_allowed_t)
        cwnd = cwnd_t
      end

       send_wnd = cwnd * CWND_OVERHEAD * rel_framesize_high -
         bytes_in_flight

      pace_bitrate = max(RATE_PACE_MIN, target_bitrate) *
        PACKET_PACING_HEADROOM
      t_pace = rtp_size * 8 / pace_bitrate

4.3. Media Rate Control

The media rate control algorithm is executed whenever the congestion window is updated and updates the target bitrate. The target bitrate is essentiatlly based on the congestion window and the (smoothed) RTT according to¶

         target_bitrate = 8 * cwnd / s_rtt

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. Because the congestion window is updated based on loss, ECN-CE and delay, so does the target rate also update.¶

The code above however needs some modifications to work fine in a number of scenarios¶

• L4S is inactive, i.e L4S is either not enabled or congested bottlenecks do not L4S mark packets¶

• Frame sizes vary¶

• cwnd is very small, just a few MSS or smaller¶

The complete pseudo code for adjustment of the target bitrate is shown below¶

        tmp_t = 1.0

        # limit bitrate if bytes in flight exceeds is close to or
        # exceeds cwnd. This helps to avoid large rate fluctiations and
        # variations in RTT
        # Only applied when L4S is inactive
        if (!l4s_active && bytes_in_flight_ratio > BYTES_IN_FLIGHT_LIMIT)
          tmp_t /= min(BYTES_IN_FLIGHT_LIMIT_COMPENSATION,
            bytesInFlightRatio / BYTES_IN_FLIGHT_LIMIT)
        end

        # Scale down rate slighty when the congestion window is very
        # small compared to MSS
        tmp_t *= 1.0 - min(0.8, max(0.0, cwnd_ratio - 0.1))

        # Scale down rate when frame sizes vary much
        tmp_t /= rel_framesize_high

        # Calculate target bitrate and limit to min and max allowed
        # values
        target_bitrate = tmp_t * 8 * cwnd / s_rtt
        target_bitrate = min(TARGET_BITRATE_MAX,
          max(TARGET_BITRATE_MIN,target_bitrate))

The variable rel_framesize_high is based on calculation of the high percentile of the frame sizes. The calculation is based on a histogram of the frame sizes relative to the expected frame size given the target bitrate and frame period. The calculation of rel_framesize_high is done for every new video frame and is outlined roughly with the pseudo code below. For more detailed code, see [SCReAM-CPP-implementation].¶

        # frame_size is that frame size for the last encoded frame
        tmp_t = frame_size / (target_bitrate * frame_period / 8)

        if (tmp_t > 1.0)
          # Insert sample into histogram
          insert_into_histogram(tmp_t)
          # Get high percentile
          rel_framesize_high = get_histogram_high_percentile()
        end

A 75%-ile is used in [SCReAM-CPP-implementation], the histogram can be made leaky such that old samples are gradually forgotten.¶

4.4. Competing Flows Compensation

It is likely that a flow will have to share congested bottlenecks with other flows that use a more aggressive congestion control algorithm (for example, 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. SCReAMv2 takes care of such situations by adjusting the qdelay_target when loss-based flows are detected, as shown in the pseudocode below.¶

        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 prevent
              # the target delay from being locked to high values forever
              if (new_target_t < QDELAY_TARGET_LO)
                # Decrease target delay quickly, as measured queuing
                # 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)

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 that bottleneck bandwidth is reduced 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 could be that the queue delay is low, but it could also be that a competing flow is causing the bottleneck to reach the point that packet losses 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 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 could produce false positives and false negatives. The case that self-inflicted congestion by the SCReAMv2 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. However, the potential failure of the algorithm cannot be completely ruled out. A false positive (i.e., when self-inflicted congestion is mistakenly identified as competing flows) is especially problematic when it leads to increasing the target queue delay, which can cause the end-to-end delay to increase dramatically.¶

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 is 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, so in general it is 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 it can lead to basically zero throughput if competing with loss-based traffic.¶

4.5. Handling of systematic errors in video coders

Some video encoders are prone to systematically generate an output bitrate that is systematically larger or smaller than the target bitrate. SCReAMv2 can handle some deviation inherently but for larger devation it becomes necessary to compensate for this. The algorithm for this is detailed in [SCReAM-CPP-implementation].¶

ToDo: A future draft version will describe this in more detail as it has been fully integrated into SCReAMv2.¶