SCReAM

This project includes an implementation of SCReAM, a mobile optimised congestion control algorithm for realtime interactive media.

News

• 2026-03-04 :
  • Calculation of queueDelayDev changed.
  • Configurable estimatedJitter with function setEstimatedJitter
  • Build date for BW test tool sender changed
• 2026-01-21 :
  • Calculation of queueDelayDev modified to be less sensitive to scheduling jitter
  • Additional restrition of window headroom and rate increase applied when CWND/MSS is low
  • RTCP adjusted for slightly less frequent feedback
  • Build date for BW test tool sender/receiver changed
• 2025-11-10 :
  • virtual L4S backoff made faster
  • Build date for BW test tool sender changed
• 2025-11-04 - 06 :
  • l4sAlpha calculation use fast attack slow decay filter
  • Reference window headroom made adaptive based on queue delay variation
  • CWND increase restriction based on queue delay variation
  • Build date for BW test tool sender changed
• 2025-10-16 :
  • Delay and L4S based CC further modified for increased stability
  • Build date for BW test tool changed
• 2025-10-15 :
  • Delay based CC modified slightly
  • L4S based CC action after long non-congested period is modified
  • Build date for BW test tool changed
• 2025-10-14 :
  • Bug in congestion avoidance logic fixed
  • Delay based and L4S CC stablity improved for large RTTs
• 2025-09-27 - 28 :
  • Delay based congestion control (default enabled) can be selectable with function enableDelayBasedCongestionControl. With this change, delay based congestion congtrol (if enabled) runs fully in parallel with L4S.
  • Option -nodelaycc added BW test tool to disable delay based congestion control.
    • Build date for BW test tool changed
• 2025-08-08 :
  • Reordering time (packet reodering margin) is made configurable
• 2025-05-09 :
  • Added new user guide for the SCReAM BW test tool with examples
  • SCReAM BW test, added end of session summary
• 2025-04-17 :
  • Added option -txrxlog that logs time, sequence_number, tx_time, rx_time, rx_time-tx_time for each RTP packet.
• 2025-01-29 :
  • -relaxedpacing option added. Enables increased pacing rate when max rate reached
  • -postcongestiondelay option removed, replaced with a constant
  • -openwindow option replaced with windowheadroom option
  • More conservative CWND increase when max rate reached
  • Bytes inflight restriction to target rate enabled only when queue detected
  • Max feedback interval set to 10ms (was 5ms)
  • Release dates for SCReAM BW test changed

Older version history is found here https://github.com/EricssonResearch/scream/blob/master/version-history.md

What is SCReAM

SCReAM (Self-Clocked Rate Adaptation for Multimedia) is a congestion control algorithm devised mainly for Video. Congestion control for WebRTC media is currently being standardized in the IETF RMCAT WG, the scope of the working group was to define requirements for congestion control and also to standardize a few candidate solutions. SCReAM is a congestion control candidate solution for WebRTC developed at Ericsson Research and optimized for good performance in wireless access.

The algorithm is an IETF experimental standard [1], a Sigcomm paper [2] and [3] explains the rationale behind the design of the algorithm in more detail. Because SCReAM as most other congestion control algorithms are continously improved over time, the current implementation available here has deviated from what is described in the papers and IETF RFC. The most important new development is addition of L4S support. In addition the algorithm has been modified to become more stable.

As mentioned above, SCReAM was originally devised for WebRTC but is sofar not incorporated into that platform. Instead, SCReAM has found use as congestion control for remote controlled vehicles, cloud gaming demos and benchmarking of 5G networks with and without L4S support.

Since standardization in RFC8298, SCReAM has undergone changes and a new V2 is described in https://datatracker.ietf.org/doc/draft-johansson-ccwg-rfc8298bis-screamv2/

Test report(s) for SCReAM V2 is found here https://github.com/EricssonResearch/scream/blob/master/test-record.md

A test report at CableLabs L4S interop test in november 2024 shows that SCReAM V2 works fine when subject to competing flows over the same bottleneck. https://github.com/EricssonResearch/scream/blob/master/CableLabs-L4S-interop-nov-2024-Ericsson.pdf

What is L4S ?

L4S is short for Low Latency Low Loss Scalable thorughput, L4S is specified in [4]. A network node that is L4S capable can remark packets that have the ECT(1) code point set to CE. The marking threshold is set very low (milliseconds).

A sender that is L4S capable sets the ECT(1) code point on outgoing packets. If CE packets are detected, then the sender should reduce the transmission rate in proportion to the amount of packets that are marked. A document that highlights how L4S improves performance for low latency applications is found in https://github.com/EricssonResearch/scream/blob/master/L4S-Results.pdf

In steady state, 2 packets per RTT should be marked. The expected rate then becomes <br> rate = (2.0/pMark) * MSS * 8/RTT [bps]
How SCReAM (V2) manages this is illustrated in the figure below
Figure 1 : SCReAM V2 bitrate as function of packet marking probability. RTT = 25ms, MSS=1360B. Dotted is theoretical, blue is actual

The more nitty gritty details

Unlike many other congestion control algorithms that are rate based i.e. they estimate the network throughput and adjust the media bitrate accordingly, SCReAM is self-clocked which essentially means that the algorithm does not send in more data into a network than what actually exits the network.

To achieve this, SCReAM implements a feedback protocol over RTCP that acknowledges received RTP packets. A congestion window is determined from the feedback, this congestion window determines how many RTP packets that can be in flight i.e. transmitted by not yet acknowledged, an RTP queue is maintained at the sender side to temporarily store the RTP packets pending transmission, this RTP queue is mostly empty but can temporarily become larger when the link throughput decreases. The congestion window is frequently adjusted for minimal e2e delay while still maintaining as high link utilization as possible. The use of self-clocking in SCReAM which is also the main principle in TCP has proven to work particularly well in wireless scenarios where the link throughput may change rapidly. This enables a congestion control which is robust to channel jitter, introduced by e.g. radio resource scheduling while still being able to respond promptly to reduced link throughput. SCReAM is optimized in house in a state of the art LTE system simulator for optimal performance in deployments where the LTE radio conditions are limiting. In addition, SCReAM is also optimized for good performance in simple bottleneck case such as those given in home gateway deployments. SCReAM is verified in simulator and in a testbed to operate in a rate range from a couple of 100kbps up to well over 100Mbps. The fact that SCReAM maintains a RTP queue on the sender side opens up for further optimizations to congestion, for instance it is possible to discard the contents of the RTP queue and replace with an I frame in order to refresh the video quickly at congestion.

SCReAM performance and behavior

SCReAM has been evaluated in a number of experiments over the years. Some of these are exemplified below.

A short video exemplifies the use of SCReAM in a small vehicle, remote controlled over a public LTE network. [8] explains the rationale behind the use of SCReAM in remote controlled applications over LTE/5G.

ECN (Explicit Congestion Notification)

SCReAM supports "classic" ECN, i.e. that the sending rate is reduced as a result of one or more ECN marked RTP packets in one RTT, similar to the guidelines in RFC3168. .

In addition SCReAM also supports L4S, i.e that the sending rate is reduced proportional to the fraction of the RTP packets that are ECN-CE marked. This enables lower network queue delay.

Below is shown two simulation examples with a simple 50Mbps 25ms. The video trace is from a video encoder.

L4S gives a somewhat lower media rate, the reason is that a larger headroom is added to ensure the low delay, considering the varying output rate of the video encoder. This is self-adjusting by inherent design because the larger frames hit the L4S enabled queue more and thus causes more marking. The average bitrate would increase if the frame size variations are smaller.

Figure 2 : SCReAM V2 without L4S support

Figure 3 : SCReAM V2 with L4S support. L4S ramp-marker (Th_low=2ms, Th_high=10ms)

Another example with the SCreAM BW test tool over a 50Mbps, 25ms bottleneck with DualPi2 AQM Figure 4 : SCReAM V2 without L4S support, 50Mbps, 25ms bottleneck with DualPi2 AQM

Figure 5 : SCReAM V2 with L4S support, 50Mbps, 25ms bottleneck with DualPi2 AQM

SCReAM is also implemented in a remote controlled car prototype. The two vide