Static Rate Management (SRM) for Low Latency, Low Loss, and Scalable Throughput (L4S)

Introduction The Internet's evolution has led to an increasing demand for Applications that require low latency and low loss, such as real-time communication, online gaming, and industrial control. Traditional TCP congestion control mechanisms, while robust, often introduce significant queuing delay under load, which can degrade the performance of these latency-sensitive applications. L4S (Low Latency, Low Loss, Scalable Throughput) is a set of mechanisms designed to address this challenge by enabling network elements to signal incipient congestion to L4S-capable transport protocols using the L4S mode of Explicit Congestion Notification (ECN) redefined in from the original Classic ECN in , specifically, the ECT(1) codepoint. made it possible to enable experiments in which ECT(1) is used differently. This allows L4S senders to react to congestion before queues build up, maintaining low latency and low loss while achieving high throughput. describes the overall L4S architecture and requirements. While many L4S solutions rely on Active Queue Management (AQM) mechanisms to detect and signal congestion, this document proposes an alternative: Static Rate Management (SRM). SRM is particularly suited for scenarios where a "safe" and fixed rate can be defined for L4S traffic on a given link, offering a simpler deployment model without the need for building and monitoring queues. SRM directly manages the aggregate rate of applications and represents an alternative to the Dual-Queue coupled AQM algorithm , which is still necessary for connections with variable rate. This document describes the SRM solution, its operational principles, and configuration guidelines.

Terminology The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 when, and only when, they appear in all capitals, as shown here. This document uses the following terms:

L4S: Low Latency, Low Loss, Scalable Throughput. A set of mechanisms for congestion control that aims to provide low latency and low loss for specific traffic.
ECN: Explicit Congestion Notification and . A mechanism where network devices can signal congestion to endpoints without dropping packets.
ECT(0): ECN Capable Transport (0). A codepoint set by the application or transport layer stack in the IP ECN field indicating that the transport is ECN-capable but uses Classic congestion control.
ECT(1): ECN Capable Transport (1). A codepoint set by the application or transport layer stack in the IP ECN field indicating that the transport is ECN-capable and uses L4S congestion control.
NotECT: Not ECN Capable Transport. The default codepoint in the IP ECN field indicating that the transport is not ECN-capable.
CE: Congestion Experienced. A codepoint in the IP ECN field set by the network to indicate that congestion has been experienced.
CIR: Committed Information Rate. The guaranteed rate for a traffic flow, below which packets are typically marked green (in- profile). In SRM, this is the rate below which L4S packets are not CE-marked. Above this rate markets are yellow, in SRM: CE-marked.
PIR: Peak Information Rate. The maximum rate allowed for a traffic flow, above which packets are typically marked red (out- of-profile) and dropped. In SRM, this is the rate above which L4S packets are dropped.
trTCM: Two-Rate, Three-Color Marker . A traffic policer that marks packets based on two rates (CIR and PIR) and two associated burst sizes, resulting in three possible colors (green, yellow, red).
RTT: Round-Trip Time. The time it takes for a signal to be sent and the acknowledgment of that signal to be received.
SRM: Static Rate Management. The solution described in this document.

Static Rate Management (SRM) for L4S

Overview The Static Rate Management (SRM) solution for L4S flows leverages equipment supporting by utilizing a standard policer with a Two-Rate, Three-Color Marker (trTCM). This approach serves as an alternative to AQM-based L4S solutions, particularly suitable for scenarios where a "safe" (non-blocking) and fixed rate can be defined on a fixed-rate link. This solution is applicable across a wide range of link speeds, from Mbps to Tbps. It is especially interesting for very high-speed and aggregated networks where queue size access and AQM algorithms might introduce complexity or be challenging to implement when packet processing capabilities are limited. The only packet processing required at the point of SRM application is for setting the CE marking bit, or a lower layer bit or codepoint that can later be moved into the IP ECN field by edge nodes.

Operation The SRM solution operates by configuring non-coupled dual queues and applying a trTCM policer to the L4S queue:

Dual Queue Configuration: Two parallel queues MUST be configured for the same traffic class:
- An L4S queue for packets marked with ECT(1) and CE.
- A Classic queue for packets marked with ECT(0) and NotECT.
Priority and Policer Application: The L4S queue MUST be given priority over the Classic queue. This priority SHOULD be strict or can be at a high enough weight to prevent latency for the mostly empty L4S queue. A Two-Rate, Three-Color Marker (trTCM) policer MUST be applied to the L4S queue.

Marking and Dropping Logic The trTCM policer applied to the L4S queue operates as follows:

CE Marking: The policer marks packets as CE if their rate exceeds the configured Committed Information Rate (CIR) (yellow state). This signals a rate limit to L4S-capable endpoints. Marks will be evenly spread over different flows in the aggregate, resulting in approximate equal rates for all flows in the aggregate.
Packet Dropping: Packets are dropped if their rate exceeds the Peak Information Rate (PIR) (red state). The PIR serves as a protection mechanism to prevent misuse by non-responsive traffic and to protect Classic flows from being starved by excessive L4S traffic.

Advantages

No Queuing Delay: This approach avoids generating additional queuing latency for L4S flows, as it does not rely on AQM thresholds that inherently relies on delay. "Congestion" is signalled via rate excess rather than queue build-up.
Predictable Performance: SRM provides predictable performance for L4S traffic within the defined rate limits (CIR and PIR). The rate is only limited by the number of applications that are active. This can be beneficial for applications requiring a low rate resulting in consistent quality of service.
Isolation between L4S and Classic: An SRM configuration provides isolation between misbehaving and broken congestion controls in one traffic class to the other. Non-behaving or overloaded L4S traffic is limited by the PIR rate, protecting the aggregate Classic left-over rate, and on the other hand, non-behaving or overloaded Classic traffic cannot harm the L4S traffic which is prioritized (but limited between CIR and PIR).
Scalability for Network Segments: A single strategic bottleneck or edge node where SRM trTCM is applied can ensure that other parts of a network with higher non-blocking (overprovisioned) capacities only need to implement the priority queue for L4S, without the need for additional marking policers. This simplifies configuration in larger network segments. The marking rate limit can be at any place in the NW to perform its rate limiting functionality. The dropping policer functionality is recommended to be performed at the ingress, to prevent that the excess PIR rate will cause latency and Classic starvation on previous network links and elements. Due to this property, it is also possible to have 2 separate single-rate Two Color Markers applied in series and on different elements.

Disadvantages

Rate Limitation: The rate for L4S flows is explicitly limited by the policer (CIR to PIR) and cannot utilize the full link capacity if Classic traffic is not utilizing the left-over capacity. This is less of a concern on large aggregation links or where sufficient bandwidth is provisioned. On the other hand, it means that a minimum rate can be guaranteed for Classic traffic (total_capacity - L4S_PIR), as L4S traffic cannot consume all available excess capacity beyond the PIR and Classic traffic can still use the full link capacity when no L4S traffic is active ("speed test safe").
Fixed Rate Requirement: This solution is only viable where a "safe" (non-blocking) and fixed rate can be reliably defined for the link or network segment. For links with large, unknown, or highly variable capacity (wireless or highly variable priority traffic), other solutions (e.g., AQM-based as described in ) are more appropriate.

Two-Rate, Three-Color Marker (trTCM) Configuration Proper configuration of the trTCM is crucial for the effective operation of SRM.

Burst Time A burst time between 1ms and 10ms is generally sufficient for both the PIR and CIR meters. A default of 4 ms is RECOMMENDED.

Links that carry a large aggregate of flows could use a lower- than-default value for burst time to ensure quicker reaction to rate changes and less jitter impact for the lower priorities.
Links that are immediately following a bursty network technology like Wi-Fi or 4G/5G might require a higher-than-default value to accommodate natural bursts without premature marking or dropping.

If a burst size is needed for configuration (e.g., in bytes), the following conversion SHOULD be used:

Peak Information Rate (PIR) Dimensioning To prevent excessive latency for Classic traffic and avoid Classic throughput starvation, the dropping PIR rate SHOULD be configured to occur before any schedulers or shapers block due to oversubscription. Typically, 30% to 50% of the total link capacity can be reserved for Classic traffic when L4S is under full load (just before dropping starts). Therefore, a PIR rate of 50% to 70% of the total link capacity is RECOMMENDED for L4S traffic. This ensures that Classic traffic always has a significant portion of the link capacity available, even if L4S traffic is attempting to consume its maximum allowed rate. If a network node handles multiple subscriptions with isolation between them, it can be considered to set the PIR closer to the aggregate subscriber rate or remove the dropping PIR completely. The higher level scheduler will guarantee rate fairness between subscribers, and misbehaving or overloaded subscribers will only cause harm on themselves.

Committed Information Rate (CIR) and Excess Marking The CIR marking policer acts as an "excess" marker. For example, if a 10 Gbps CIR is configured:

0.99% of packets will be marked CE if the aggregate L4S rate reaches 10.1 Gbps.
50% of packets will be marked CE if the aggregate L4S rate reaches 20 Gbps.

Thus, the marking probability 'p' for L4S traffic exceeding the CIR is given by: For steady-state Prague flows , the rule of thumb for the rate per flow (Rpf) depends on the marking probability 'p' as follows: Using the previous examples:

With 0.99% marking (p=0.0099), the Rpf would be approximately 100 Mbps.
With 50% marking (p=0.5), the Rpf would be 1 Mbps.

If 0.99% marking occurs on a 10 Gbps CIR rate, the aggregate arrival rate will be 10.1 Gbps, this implies that approximately 10.1 Gbps / 100 Mbps = 101 capacity-seeking flows are active. Similarly, to reach a 50% marking rate, 20 Gbps / 1 Mbps = 20,000 capacity- seeking flows would need to be active. When a single L4S flow is present, its rate will be slightly above the CIR. As the number of flows increases or the absolute CIR rate decreases, the aggregate rate will climb higher above the CIR. At some point, a very large amount of flows will cause the aggregate rate to reach the PIR, at which point dropping begins. The number of flows (N) at the point where the aggregate rate (N * Rpf) is equal to the PIR can be approximated by: where PIR and CIR are expressed in Mbps. If more than this number of capacity-seeking flows are active, the aggregate rate will exceed the PIR, and drops will begin.

PIR/CIR Ratio The marking CIR and dropping PIR rates MUST be sufficiently separated to allow a large number of flows to share the capacity and ensure L4S flows can converge effectively. The ratio between the dropping PIR and the marking CIR SHOULD be at least a factor of 2. This ratio allows for a significant portion of packets to be CE- marked before drops occur, providing a robust signal for L4S transports to reduce their rate. It also supports slow start without loss, as L4S slow start typically doubles the rate every RTT.

CIR Dimensioning The RECOMMENDED PIR/CIR ratio of 2 is generally sufficient when the number of flows is not expected to exceed the PIR expressed in Mbps units (e.g., 1000 flows for a 1 Gbps PIR).

The PIR/CIR ratio MAY be reduced below the recommended factor of 2 for links with higher capacity or less aggregation, where the impact of a smaller marking window is less critical.
The PIR/CIR ratio MAY need to be greater than 2 for constrained links that carry a very large number of flows, to provide a higher ECN marking probability before drops occur and to better accommodate the dynamics of many concurrent flows.

Security Considerations Similar as DualPI2 , also the SRM solution does not need to inspect beyond the ECN field. It is fully independent of higher layer protocols and tunnels. It poses no restrictions and traffic can be further fully encrypted over the available IP layer. The SRM solution relies on the proper classification and marking of L4S traffic. Misclassification, malicious marking of non-L4S traffic as ECT(1), or exploiting L4S for DoS attacks will have no different impact on other traffic as non-responsive traffic has on Classic-only networks. To the benefit of the SRM solution, due to the isolation, on-purpose overload attacks will need to generate a mix of L4S and Classic traffic to fully overload the network service, as the SRM solution does not couple congestion between the traffic classes.

Contributors Thanks to Greg White, and members of the TSVWG mailing list for their contributions to this document.

IANA Considerations This document has no IANA actions.