1.1. Outline of the Problem

Latency is becoming the critical performance factor for many (perhaps most) applications on the public Internet, e.g., interactive web, web services, voice, conversational video, interactive video, interactive remote presence, instant messaging, online gaming, remote desktop, cloud-based applications, and video-assisted remote control of machinery and industrial processes. Once access network bitrates reach levels now common in the developed world, further increases offer diminishing returns unless latency is also addressed [Dukkipati06]. In the last decade or so, much has been done to reduce propagation time by placing caches or servers closer to users. However, queuing remains a major intermittent component of latency.¶

Previously, very low latency has only been available for a few selected low-rate applications, that confine their sending rate within a specially carved-off portion of capacity, which is prioritized over other traffic, e.g., Diffserv Expedited Forwarding (EF) [RFC3246]. Up to now, it has not been possible to allow any number of low-latency, high throughput applications to seek to fully utilize available capacity, because the capacity-seeking process itself causes too much queuing delay.¶

To reduce this queuing delay caused by the capacity-seeking process, changes either to the network alone or to end systems alone are in progress. L4S involves a recognition that both approaches are yielding diminishing returns:¶

• Recent state-of-the-art AQM in the network, e.g., Flow Queue CoDel [RFC8290], Proportional Integral controller Enhanced (PIE) [RFC8033], and Adaptive Random Early Detection (ARED) [ARED01]), has reduced queuing delay for all traffic, not just a select few applications. However, no matter how good the AQM, the capacity-seeking (sawtoothing) rate of TCP-like congestion controls represents a lower limit that will cause either the queuing delay to vary or the link to be underutilized. These AQMs are tuned to allow a typical capacity-seeking TCP-Reno-friendly flow to induce an average queue that roughly doubles the base round-trip time (RTT), adding 5-15 ms of queuing on average for a mix of long-running flows and web traffic (cf. 500 microseconds with L4S for the same traffic mix [L4Seval22]). However, for many applications, low delay is not useful unless it is consistently low. With these AQMs, 99th percentile queuing delay is 20-30 ms (cf. 2 ms with the same traffic over L4S).¶
• Similarly, recent research into using end-to-end congestion control without needing an AQM in the network (e.g., Bottleneck Bandwidth and Round-trip propagation time (BBR) [BBR-CC]) seems to have hit a similar queuing delay floor of about 20 ms on average, but there are also regular 25 ms delay spikes due to bandwidth probes and 60 ms spikes due to flow-starts.¶

L4S learns from the experience of Data Center TCP (DCTCP) [RFC8257], which shows the power of complementary changes both in the network and on end systems. DCTCP teaches us that two small but radical changes to congestion control are needed to cut the two major outstanding causes of queuing delay variability:¶

1. Far smaller rate variations (sawteeth) than Reno-friendly congestion controls.¶
2. A shift of smoothing and hence smoothing delay from network to sender.¶

Without the former, a 'Classic' (e.g., Reno-friendly) flow's RTT varies between roughly 1 and 2 times the base RTT between the machines in question. Without the latter, a 'Classic' flow's response to changing events is delayed by a worst-case (transcontinental) RTT, which could be hundreds of times the actual smoothing delay needed for the RTT of typical traffic from localized Content Delivery Networks (CDNs).¶

These changes are the two main features of the family of so-called 'Scalable' congestion controls (which include DCTCP, Prague, and Self-Clocked Rate Adaptation for Multimedia (SCReAM)). Both of these changes only reduce delay in combination with a complementary change in the network, and they are both only feasible with ECN, not drop, for the signalling:¶

1. The smaller sawteeth allow an extremely shallow ECN packet-marking threshold in the queue.¶
2. No smoothing in the network means that every fluctuation of the queue is signalled immediately.¶

Without ECN, either of these would lead to very high loss levels. In contrast, with ECN, the resulting high marking levels are just signals, not impairments. (Note that BBRv2 [BBRv2] combines the best of both worlds -- it works as a Scalable congestion control when ECN is available, but it also aims to minimize delay when ECN is absent.)¶

However, until now, Scalable congestion controls (like DCTCP) did not coexist well in a shared ECN-capable queue with existing Classic (e.g., Reno [RFC5681] or CUBIC [RFC8312]) congestion controls -- Scalable controls are so aggressive that these 'Classic' algorithms would drive themselves to a small capacity share. Therefore, until now, L4S controls could only be deployed where a clean-slate environment could be arranged, such as in private data centres (hence the name DCTCP).¶

One way to solve the problem of coexistence between Scalable and Classic flows is to use a per-flow-queuing (FQ) approach such as FQ-CoDel [RFC8290]. It classifies packets by flow identifier into separate queues in order to isolate sparse flows from the higher latency in the queues assigned to heavier flows. However, if a Classic flow needs both low delay and high throughput, having a queue to itself does not isolate it from the harm it causes to itself. Also FQ approaches need to inspect flow identifiers, which is not always practical.¶

In summary, Scalable congestion controls address the root cause of the latency, loss and scaling problems with Classic congestion controls. Both FQ and DualQ AQMs can be enablers for this smooth low-latency scalable behaviour. The DualQ approach is particularly useful because identifying flows is sometimes not practical or desirable.¶

1.2. Context, Scope, and Applicability

L4S involves complementary changes in the network and on end systems:¶

Network:

A DualQ Coupled AQM (defined in the present document) or a modification to flow queue AQMs (described in paragraph "b" in Section 4.2 of the L4S architecture [RFC9330]).¶

End system:

A Scalable congestion control (defined in Section 4 of the L4S ECN protocol spec [RFC9331]).¶

Packet identifier:

The network and end-system parts of L4S can be deployed incrementally, because they both identify L4S packets using the experimentally assigned ECN codepoints in the IP header: ECT(1) and CE [RFC8311] [RFC9331].¶

DCTCP [RFC8257] is an example of a Scalable congestion control for controlled environments that has been deployed for some time in Linux, Windows, and FreeBSD operating systems. During the progress of this document through the IETF, a number of other Scalable congestion controls were implemented, e.g., Prague over TCP and QUIC [PRAGUE-CC] [PragueLinux], BBRv2 [BBRv2] [BBR-CC], and the L4S variant of SCReAM for real-time media [SCReAM-L4S] [RFC8298].¶

The focus of this specification is to enable deployment of the network part of the L4S service. Then, without any management intervention, applications can exploit this new network capability as the applications or their operating systems migrate to Scalable congestion controls, which can then evolve while their benefits are being enjoyed by everyone on the Internet.¶

The DualQ Coupled AQM framework can incorporate any AQM designed for a single queue that generates a statistical or deterministic mark/drop probability driven by the queue dynamics. Pseudocode examples of two different DualQ Coupled AQMs are given in the appendices. In many cases the framework simplifies the basic control algorithm and requires little extra processing. Therefore, it is believed the Coupled AQM would be applicable and easy to deploy in all types of buffers such as buffers in cost-reduced mass-market residential equipment; buffers in end-system stacks; buffers in carrier-scale equipment including remote access servers, routers, firewalls, and Ethernet switches; buffers in network interface cards; buffers in virtualized network appliances, hypervisors; and so on.¶

For the public Internet, nearly all the benefit will typically be achieved by deploying the Coupled AQM into either end of the access link between a 'site' and the Internet, which is invariably the bottleneck (see Section 6.4 of [RFC9330] about deployment, which also defines the term 'site' to mean a home, an office, a campus, or mobile user equipment).¶

Latency is not the only concern of L4S:¶

• The 'Low Loss' part of the name denotes that L4S generally achieves zero congestion loss (which would otherwise cause retransmission delays), due to its use of ECN.¶
• The 'Scalable throughput' part of the name denotes that the per-flow throughput of Scalable congestion controls should scale indefinitely, avoiding the imminent scaling problems with 'TCP-Friendly' congestion control algorithms [RFC3649].¶

The former is clearly in scope of this AQM document. However, the latter is an outcome of the end-system behaviour and is therefore outside the scope of this AQM document, even though the AQM is an enabler.¶

The overall L4S architecture [RFC9330] gives more detail, including on wider deployment aspects such as backwards compatibility of Scalable congestion controls in bottlenecks where a DualQ Coupled AQM has not been deployed. The supporting papers [L4Seval22], [DualPI2Linux], [PI2], and [PI2param] give the full rationale for the AQM design, both discursively and in more precise mathematical form, as well as the results of performance evaluations. The main results have been validated independently when using the Prague congestion control [Boru20] (experiments are run using Prague and DCTCP, but only the former is relevant for validation, because Prague fixes a number of problems with the Linux DCTCP code that make it unsuitable for the public Internet).¶

1.3. 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 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.¶

The DualQ Coupled AQM uses two queues for two services:¶

Classic Service/Queue:

The Classic service is intended for all the congestion control behaviours that coexist with Reno [RFC5681] (e.g., Reno itself, CUBIC [RFC8312], and TFRC [RFC5348]). The term 'Classic queue' means a queue providing the Classic service.¶

Low Latency, Low Loss, and Scalable throughput (L4S) Service/Queue:

The 'L4S' service is intended for traffic from Scalable congestion control algorithms, such as the Prague congestion control [PRAGUE-CC], which was derived from Data Center TCP [RFC8257]. The L4S service is for more general traffic than just Prague -- it allows the set of congestion controls with similar scaling properties to Prague to evolve, such as the examples listed below (Relentless, SCReAM, etc.). The term 'L4S queue' means a queue providing the L4S service.¶

Classic Congestion Control:

A congestion control behaviour that can coexist with standard Reno [RFC5681] without causing significantly negative impact on its flow rate [RFC5033]. With Classic congestion controls, such as Reno or CUBIC, because flow rate has scaled since TCP congestion control was first designed in 1988, it now takes hundreds of round trips (and growing) to recover after a congestion signal (whether a loss or an ECN mark) as shown in the examples in Section 5.1 of the L4S architecture [RFC9330] and in [RFC3649]. Therefore, control of queuing and utilization becomes very slack, and the slightest disturbances (e.g., from new flows starting) prevent a high rate from being attained.¶

Scalable Congestion Control:

A congestion control where the average time from one congestion signal to the next (the recovery time) remains invariant as flow rate scales, all other factors being equal. This maintains the same degree of control over queuing and utilization whatever the flow rate, as well as ensuring that high throughput is robust to disturbances. For instance, DCTCP averages 2 congestion signals per round trip, whatever the flow rate, as do other recently developed Scalable congestion controls, e.g., Relentless TCP [RELENTLESS], Prague [PRAGUE-CC] [PragueLinux], BBRv2 [BBRv2] [BBR-CC], and the L4S variant of SCReAM for real-time media [SCReAM-L4S] [RFC8298]. For the public Internet, a Scalable transport has to comply with the requirements in Section 4 of [RFC9331] (a.k.a. the 'Prague L4S requirements').¶

C:

Abbreviation for Classic, e.g., when used as a subscript.¶

L:

Abbreviation for L4S, e.g., when used as a subscript.¶

The terms Classic or L4S can also qualify other nouns, such as 'codepoint', 'identifier', 'classification', 'packet', and 'flow'. For example, an L4S packet means a packet with an L4S identifier sent from an L4S congestion control.¶

Both Classic and L4S services can cope with a proportion of unresponsive or less-responsive traffic as well but, in the L4S case, its rate has to be smooth enough or low enough to not build a queue (e.g., DNS, Voice over IP (VoIP), game sync datagrams, etc.). The DualQ Coupled AQM behaviour is defined to be similar to a single First-In, First-Out (FIFO) queue with respect to unresponsive and overload traffic.¶

Reno-friendly:

The subset of Classic traffic that is friendly to the standard Reno congestion control defined for TCP in [RFC5681]. The TFRC spec [RFC5348] indirectly implies that 'friendly' is defined as "generally within a factor of two of the sending rate of a TCP flow under the same conditions". 'Reno-friendly' is used here in place of 'TCP-friendly', given the latter has become imprecise, because the TCP protocol is now used with so many different congestion control behaviours, and Reno is used in non-TCP transports, such as QUIC [RFC9000].¶

DualQ or DualQ AQM:

Used loosely as shorthand for a Dual-Queue Coupled AQM, where the context makes 'Coupled AQM' obvious.¶

Classic ECN:

The original Explicit Congestion Notification (ECN) protocol [RFC3168] that requires ECN signals to be treated as equivalent to drops, both when generated in the network and when responded to by the sender.¶

For L4S, the names used for the four codepoints of the 2-bit IP-ECN field are unchanged from those defined in the ECN spec [RFC3168], i.e., Not-ECT, ECT(0), ECT(1), and CE, where ECT stands for ECN-Capable Transport and CE stands for Congestion Experienced. A packet marked with the CE codepoint is termed 'ECN-marked' or sometimes just 'marked' where the context makes ECN obvious.¶

2.1. Coupled AQM

In the 1990s, the 'TCP formula' was derived for the relationship between the steady-state congestion window, cwnd, and the drop probability, p of standard Reno congestion control [RFC5681]. To a first-order approximation, the steady-state cwnd of Reno is inversely proportional to the square root of p.¶

The design focuses on Reno as the worst case, because if it does no harm to Reno, it will not harm CUBIC or any traffic designed to be friendly to Reno. TCP CUBIC implements a Reno-friendly mode, which is relevant for typical RTTs under 20 ms as long as the throughput of a single flow is less than about 350 Mb/s. In such cases, it can be assumed that CUBIC traffic behaves similarly to Reno. The term 'Classic' will be used for the collection of Reno-friendly traffic including CUBIC and potentially other experimental congestion controls intended not to significantly impact the flow rate of Reno.¶

A supporting paper [PI2] includes the derivation of the equivalent rate equation for DCTCP, for which cwnd is inversely proportional to p (not the square root), where in this case p is the ECN-marking probability. DCTCP is not the only congestion control that behaves like this, so the term 'Scalable' will be used for all similar congestion control behaviours (see examples in Section 1.2). The term 'L4S' is used for traffic driven by a Scalable congestion control that also complies with the additional 'Prague L4S requirements' [RFC9331].¶

For safe coexistence, under stationary conditions, a Scalable flow has to run at roughly the same rate as a Reno TCP flow (all other factors being equal). So the drop or marking probability for Classic traffic, p_C, has to be distinct from the marking probability for L4S traffic, p_L. The original ECN spec [RFC3168] required these probabilities to be the same, but [RFC8311] updates [RFC3168] to enable experiments in which these probabilities are different.¶

Also, to remain stable, Classic sources need the network to smooth p_C so it changes relatively slowly. It is hard for a network node to know the RTTs of all the flows, so a Classic AQM adds a worst-case RTT of smoothing delay (about 100-200 ms). In contrast, L4S shifts responsibility for smoothing ECN feedback to the sender, which only delays its response by its own RTT, as well as allowing a more immediate response if necessary.¶

The Coupled AQM achieves safe coexistence by making the Classic drop probability p_C proportional to the square of the coupled L4S probability p_CL. p_CL is an input to the instantaneous L4S marking probability p_L, but it changes as slowly as p_C. This makes the Reno flow rate roughly equal the DCTCP flow rate, because the squaring of p_CL counterbalances the square root of p_C in the 'TCP formula' of Classic Reno congestion control.¶

Stating this as a formula, the relation between Classic drop probability, p_C, and the coupled L4S probability p_CL needs to take the following form:¶

    p_C = ( p_CL / k )^2,                 (1)

where k is the constant of proportionality, which is termed the 'coupling factor'.¶

2.2. Dual Queue

Classic traffic needs to build a large queue to prevent underutilization. Therefore, a separate queue is provided for L4S traffic, and it is scheduled with priority over the Classic queue. Priority is conditional to prevent starvation of Classic traffic in certain conditions (see Section 2.4).¶

Nonetheless, coupled marking ensures that giving priority to L4S traffic still leaves the right amount of spare scheduling time for Classic flows to each get equivalent throughput to DCTCP flows (all other factors, such as RTT, being equal).¶

2.3. Traffic Classification

Both the Coupled AQM and DualQ mechanisms need an identifier to distinguish L4S (L) and Classic (C) packets. Then the coupling algorithm can achieve coexistence without having to inspect flow identifiers, because it can apply the appropriate marking or dropping probability to all flows of each type. A separate specification [RFC9331] requires the network to treat the ECT(1) and CE codepoints of the ECN field as this identifier. An additional process document has proved necessary to make the ECT(1) codepoint available for experimentation [RFC8311].¶

For policy reasons, an operator might choose to steer certain packets (e.g., from certain flows or with certain addresses) out of the L queue, even though they identify themselves as L4S by their ECN codepoints. In such cases, the L4S ECN protocol [RFC9331] states that the device "MUST NOT alter the end-to-end L4S ECN identifier" so that it is preserved end to end. The aim is that each operator can choose how it treats L4S traffic locally, but an individual operator does not alter the identification of L4S packets, which would prevent other operators downstream from making their own choices on how to treat L4S traffic.¶

In addition, an operator could use other identifiers to classify certain additional packet types into the L queue that it deems will not risk harm to the L4S service, for instance, addresses of specific applications or hosts; specific Diffserv codepoints such as EF, Voice-Admit, or the Non-Queue-Building (NQB) per-hop behaviour; or certain protocols (e.g., ARP and DNS) (see Section 5.4.1 of [RFC9331]. Note that [RFC9331] states that "a network node MUST NOT change Not-ECT or ECT(0) in the IP-ECN field into an L4S identifier." Thus, the L queue is not solely an L4S queue; it can be considered more generally as a low-latency queue.¶

2.4. Overall DualQ Coupled AQM Structure

Figure 1 shows the overall structure that any DualQ Coupled AQM is likely to have. This schematic is intended to aid understanding of the current designs of DualQ Coupled AQMs. However, it is not intended to preclude other innovative ways of satisfying the normative requirements in Section 2.5 that minimally define a DualQ Coupled AQM. Also, the schematic only illustrates operation under normally expected circumstances; behaviour under overload or with operator-specific classifiers is deferred to Section 2.5.1.1.¶

The classifier on the left separates incoming traffic between the two queues (L and C). Each queue has its own AQM that determines the likelihood of marking or dropping (p_L and p_C). In [PI2], it has been proved that it is preferable to control load with a linear controller, then square the output before applying it as a drop probability to Reno-friendly traffic (because Reno congestion control decreases its load proportional to the square root of the increase in drop). So, the AQM for Classic traffic needs to be implemented in two stages: i) a base stage that outputs an internal probability p' (pronounced 'p-prime') and ii) a squaring stage that outputs p_C, where¶

    p_C = (p')^2.                         (2)

Substituting for p_C in equation (1) gives¶

    p' = p_CL / k.

So the slow-moving input to ECN marking in the L queue (the coupled L4S probability) is¶

    p_CL = k*p'.                          (3)

The actual ECN-marking probability p_L that is applied to the L queue needs to track the immediate L queue delay under L-only congestion conditions, as well as track p_CL under coupled congestion conditions. So the L queue uses a 'Native AQM' that calculates a probability p'_L as a function of the instantaneous L queue delay. And given the L queue has conditional priority over the C queue, whenever the L queue grows, the AQM ought to apply marking probability p'_L, but p_L ought to not fall below p_CL. This suggests¶

    p_L = max(p'_L, p_CL),                (4)

which has also been found to work very well in practice.¶

The two transformations of p' in equations (2) and (3) implement the required coupling given in equation (1) earlier.¶

The constant of proportionality or coupling factor, k, in equation (1) determines the ratio between the congestion probabilities (loss or marking) experienced by L4S and Classic traffic. Thus, k indirectly determines the ratio between L4S and Classic flow rates, because flows (assuming they are responsive) adjust their rate in response to congestion probability. Appendix C.2 gives guidance on the choice of k and its effect on relative flow rates.¶

                        _________
                               | |    ,------.
                 L4S (L) queue | |===>| ECN  |
                    ,'| _______|_|    |marker|\
                  <'  |         |     `------'\\
                   //`'         v        ^ p_L \\
                  //       ,-------.     |      \\
                 //        |Native |p'_L |       \\,.
                //         |  L4S  |--->(MAX)    <  |   ___
   ,----------.//          |  AQM  |     ^ p_CL   `\|.'Cond-`.
   |  IP-ECN  |/           `-------'     |          / itional \
==>|Classifier|            ,-------.   (k*p')       [ priority]==>
   |          |\           |  Base |     |          \scheduler/
   `----------'\\          |  AQM  |---->:        ,'|`-.___.-'
                \\         |       |p'   |      <'  |
                 \\        `-------'   (p'^2)    //`'
                  \\            ^        |      //
                   \\,.         |        v p_C //
                   <  | _________     .------.//
                    `\|   |      |    | Drop |/
              Classic (C) |queue |===>|/mark |
                        __|______|    `------'

Legend: ===> traffic flow
        ---> control dependency

After the AQMs have applied their dropping or marking, the scheduler forwards their packets to the link. Even though the scheduler gives priority to the L queue, it is not as strong as the coupling from the C queue. This is because, as the C queue grows, the 'Base AQM' applies more congestion signals to L traffic (as well as to C). As L flows reduce their rate in response, they use less than the scheduling share for L traffic. So, because the scheduler is work preserving, it schedules any C traffic in the gaps.¶

Giving priority to the L queue has the benefit of very low L queue delay, because the L queue is kept empty whenever L traffic is controlled by the coupling. Also, there only has to be a coupling in one direction -- from Classic to L4S. Priority has to be conditional in some way to prevent the C queue from being starved in the short term (see Section 4.2.2) to give C traffic a means to push in, as explained next. With normal responsive L traffic, the coupled ECN marking gives C traffic the ability to push back against even strict priority, by congestion marking the L traffic to make it yield some space. However, if there is just a small finite set of C packets (e.g., a DNS request or an initial window of data), some Classic AQMs will not induce enough ECN marking in the L queue, no matter how long the small set of C packets waits. Then, if the L queue happens to remain busy, the C traffic would never get a scheduling opportunity from a strict priority scheduler. Ideally, the Classic AQM would be designed to increase the coupled marking the longer that C packets have been waiting, but this is not always practical -- hence the need for L priority to be conditional. Giving a small weight or limited waiting time for C traffic improves response times for short Classic messages, such as DNS requests, and improves Classic flow startup because immediate capacity is available.¶

Example DualQ Coupled AQM algorithms called 'DualPI2' and 'Curvy RED' are given in Appendices A and B. Either example AQM can be used to couple packet marking and dropping across a DualQ:¶

• DualPI2 uses a Proportional Integral (PI) controller as the Base AQM. Indeed, this Base AQM with just the squared output and no L4S queue can be used as a drop-in replacement for PIE [RFC8033], in which case it is just called PI2 [PI2]. PI2 is a principled simplification of PIE that is both more responsive and more stable in the face of dynamically varying load.¶

• Curvy RED is derived from RED [RED], except its configuration parameters are delay-based to make them insensitive to link rate, and it requires fewer operations per packet than RED. However, DualPI2 is more responsive and stable over a wider range of RTTs than Curvy RED. As a consequence, at the time of writing, DualPI2 has attracted more development and evaluation attention than Curvy RED, leaving the Curvy RED design not so fully evaluated.¶

Both AQMs regulate their queue against targets configured in units of time rather than bytes. As already explained, this ensures configuration can be invariant for different drain rates. With AQMs in a DualQ structure this is particularly important because the drain rate of each queue can vary rapidly as flows for the two queues arrive and depart, even if the combined link rate is constant.¶

It would be possible to control the queues with other alternative AQMs, as long as the normative requirements (those expressed in capitals) in Section 2.5 are observed.¶

The two queues could optionally be part of a larger queuing hierarchy, such as the initial example ideas in [L4S-DIFFSERV].¶

2.5. Normative Requirements for a DualQ Coupled AQM

The following requirements are intended to capture only the essential aspects of a DualQ Coupled AQM. They are intended to be independent of the particular AQMs implemented for each queue but to still define the DualQ framework built around those AQMs.¶

4.2. Handling Unresponsive Flows and Overload

In the absence of any per-flow control, it is important that the basic DualQ Coupled AQM gives unresponsive flows no more throughput advantage than a single-queue AQM would, and that it at least handles overload situations. Overload means that incoming load significantly or persistently exceeds output capacity, but it is not intended to be a precise term -- significant and persistent are matters of degree.¶

A trade-off needs to be made between complexity and the risk of either traffic class harming the other. In overloaded conditions, the higher priority L4S service will have to sacrifice some aspect of its performance. Depending on the degree of overload, alternative solutions may relax a different factor: for example, throughput, delay, or drop. These choices need to be made either by the developer or by operator policy, rather than by the IETF. Subsequent subsections discuss handling different degrees of overload:¶

• Unresponsive flows (L and/or C) but not overloaded, i.e., the sum of unresponsive load before adding any responsive traffic is below capacity.¶

  • This case is handled by the regular Coupled DualQ (Section 2.1) but not discussed there. So below, Section 4.2.1 explains the design goal and how it is achieved in practice.¶

• Unresponsive flows (L and/or C) causing persistent overload, i.e., the sum of unresponsive load even before adding any responsive traffic persistently exceeds capacity.¶

  • This case is not covered by the regular Coupled DualQ mechanism (Section 2.1), but the last paragraph in Section 2.5.1.1 sets out a requirement to handle the case where ECN-capable traffic could starve non-ECN-capable traffic. Section 4.2.3 below discusses the general options and gives specific examples.¶

• Short-term overload that lies between the 'not overloaded' and 'persistently overloaded' cases.¶

  • For the period before overload is deemed persistent, Section 4.2.2 discusses options for more immediate mechanisms at the scheduler timescale. These prevent short-term starvation of the C queue by making the priority of the L queue conditional, as required in Section 2.5.1.¶

A.1. Pass #1: Core Concepts

The pseudocode manipulates three main structures of variables: the packet (pkt), the L4S queue (lq), and the Classic queue (cq). The pseudocode consists of the following six functions:¶

• The initialization function dualpi2_params_init(...) (Figure 2) that sets parameter defaults (the API for setting non-default values is omitted for brevity).¶
• The enqueue function dualpi2_enqueue(lq, cq, pkt) (Figure 3).¶
• The dequeue function dualpi2_dequeue(lq, cq, pkt) (Figure 4).¶
• The recurrence function recur(q, likelihood) for de-randomized ECN marking (shown at the end of Figure 4).¶
• The L4S AQM function laqm(qdelay) (Figure 5) used to calculate the ECN-marking probability for the L4S queue.¶
• The Base AQM function that implements the PI algorithm dualpi2_update(lq, cq) (Figure 6) used to regularly update the base probability (p'), which is squared for the Classic AQM as well as being coupled across to the L4S queue.¶

It also uses the following functions that are not shown in full here:¶

• scheduler(), which selects between the head packets of the two queues. The choice of scheduler technology is discussed later.¶
• cq.byt() or lq.byt() returns the current length (a.k.a. backlog) of the relevant queue in bytes.¶
• cq.len() or lq.len() returns the current length of the relevant queue in packets.¶
• cq.time() or lq.time() returns the current queuing delay of the relevant queue in units of time (see Note a below).¶
• mark(pkt) and drop(pkt) for ECN marking and dropping a packet.¶

In experiments so far (building on experiments with PIE) on broadband access links ranging from 4 Mb/s to 200 Mb/s with base RTTs from 5 ms to 100 ms, DualPI2 achieves good results with the default parameters in Figure 2. The parameters are categorised by whether they relate to the PI2 AQM, the L4S AQM, or the framework coupling them together. Constants and variables derived from these parameters are also included at the end of each category. Each parameter is explained as it is encountered in the walk-through of the pseudocode below, and the rationale for the chosen defaults are given so that sensible values can be used in scenarios other than the regular public Internet.¶

1:  dualpi2_params_init(...) {         % Set input parameter defaults
2:    % DualQ Coupled framework parameters
5:    limit = MAX_LINK_RATE * 250 ms               % Dual buffer size
3:    k = 2                                         % Coupling factor
4:    % NOT SHOWN % scheduler-dependent weight or equival't parameter
6:
7:    % PI2 Classic AQM parameters
8:    target = 15 ms                             % Queue delay target
9:    RTT_max = 100 ms                      % Worst case RTT expected
10:   % PI2 constants derived from above PI2 parameters
11:   p_Cmax = min(1/k^2, 1)             % Max Classic drop/mark prob
12:   Tupdate = min(target, RTT_max/3)        % PI sampling interval
13:   alpha = 0.1 * Tupdate / RTT_max^2      % PI integral gain in Hz
14:   beta = 0.3 / RTT_max               % PI proportional gain in Hz
15:
16:   % L4S ramp AQM parameters
17:   minTh = 800 us        % L4S min marking threshold in time units
18:   range = 400 us                % Range of L4S ramp in time units
19:   Th_len = 1 pkt           % Min L4S marking threshold in packets
20:   % L4S constants
21:   p_Lmax = 1                               % Max L4S marking prob
22: }

The overall goal of the code is to apply the marking and dropping probabilities for L4S and Classic traffic (p_L and p_C). These are derived from the underlying base probabilities p'_L and p' driven, respectively, by the traffic in the L and C queues. The marking probability for the L queue (p_L) depends on both the base probability in its own queue (p'_L) and a probability called p_CL, which is coupled across from p' in the C queue (see Section 2.4 for the derivation of the specific equations and dependencies).¶

The probabilities p_CL and p_C are derived in lines 4 and 5 of the dualpi2_update() function (Figure 6) then used in the dualpi2_dequeue() function where p_L is also derived from p_CL at line 6 (Figure 4). The code walk-through below builds up to explaining that part of the code eventually, but it starts from packet arrival.¶

1:  dualpi2_enqueue(lq, cq, pkt) { % Test limit and classify lq or cq
2:    if ( lq.byt() + cq.byt() + MTU > limit)
3:      drop(pkt)                     % drop packet if buffer is full
4:    timestamp(pkt)     % only needed if using the sojourn technique
5:    % Packet classifier
6:    if ( ecn(pkt) modulo 2 == 1 )         % ECN bits = ECT(1) or CE
7:      lq.enqueue(pkt)
8:    else                             % ECN bits = not-ECT or ECT(0)
9:      cq.enqueue(pkt)
10: }

1:  dualpi2_dequeue(lq, cq, pkt) {     % Couples L4S & Classic queues
2:    while ( lq.byt() + cq.byt() > 0 ) {
3:      if ( scheduler() == lq ) {
4:        lq.dequeue(pkt)                      % Scheduler chooses lq
5:        p'_L = laqm(lq.time())                        % Native LAQM
6:        p_L = max(p'_L, p_CL)                  % Combining function
7:        if ( recur(lq, p_L) )                      % Linear marking
8:          mark(pkt)
9:      } else {
10:       cq.dequeue(pkt)                      % Scheduler chooses cq
11:       if ( recur(cq, p_C) ) {            % probability p_C = p'^2
12:         if ( ecn(pkt) == 0 ) {           % if ECN field = not-ECT
13:           drop(pkt)                                % squared drop
14:           continue        % continue to the top of the while loop
15:         }
16:         mark(pkt)                                  % squared mark
17:       }
18:     }
19:     return(pkt)                      % return the packet and stop
20:   }
21:   return(NULL)                             % no packet to dequeue
22: }

23: recur(q, likelihood) {   % Returns TRUE with a certain likelihood
24:   q.count += likelihood
25:   if (q.count > 1) {
26:     q.count -= 1
27:     return TRUE
28:   }
29:   return FALSE
30: }

When packets arrive, a common queue limit is checked first as shown in line 2 of the enqueuing pseudocode in Figure 3. This assumes a shared buffer for the two queues (Note b discusses the merits of separate buffers). In order to avoid any bias against larger packets, 1 MTU of space is always allowed, and the limit is deliberately tested before enqueue.¶

If limit is not exceeded, the packet is timestamped in line 4 (only if the sojourn time technique is being used to measure queue delay; see Note a below for alternatives).¶

At lines 5-9, the packet is classified and enqueued to the Classic or L4S queue dependent on the least significant bit (LSB) of the ECN field in the IP header (line 6). Packets with a codepoint having an LSB of 0 (Not-ECT and ECT(0)) will be enqueued in the Classic queue. Otherwise, ECT(1) and CE packets will be enqueued in the L4S queue. Optional additional packet classification flexibility is omitted for brevity (see the L4S ECN protocol [RFC9331]).¶

The dequeue pseudocode (Figure 4) is repeatedly called whenever the lower layer is ready to forward a packet. It schedules one packet for dequeuing (or zero if the queue is empty) then returns control to the caller so that it does not block while that packet is being forwarded. While making this dequeue decision, it also makes the necessary AQM decisions on dropping or marking. The alternative of applying the AQMs at enqueue would shift some processing from the critical time when each packet is dequeued. However, it would also add a whole queue of delay to the control signals, making the control loop sloppier (for a typical RTT, it would double the Classic queue's feedback delay).¶

All the dequeue code is contained within a large while loop so that if it decides to drop a packet, it will continue until it selects a packet to schedule. Line 3 of the dequeue pseudocode is where the scheduler chooses between the L4S queue (lq) and the Classic queue (cq). Detailed implementation of the scheduler is not shown (see discussion later).¶

• If an L4S packet is scheduled, in lines 7 and 8 the packet is ECN-marked with likelihood p_L. The recur() function at the end of Figure 4 is used, which is preferred over random marking because it avoids delay due to randomization when interpreting congestion signals, but it still desynchronizes the sawteeth of the flows. Line 6 calculates p_L as the maximum of the coupled L4S probability p_CL and the probability from the Native L4S AQM p'_L. This implements the max() function shown in Figure 1 to couple the outputs of the two AQMs together. Of the two probabilities input to p_L in line 6:¶

  • p'_L is calculated per packet in line 5 by the laqm() function (see Figure 5), whereas¶
  • p_CL is maintained by the dualpi2_update() function, which runs every Tupdate (Tupdate is set in line 12 of Figure 2).¶
• If a Classic packet is scheduled, lines 10 to 17 drop or mark the packet with probability p_C.¶

The Native L4S AQM algorithm (Figure 5) is a ramp function, similar to the RED algorithm, but simplified as follows:¶

• The extent of the ramp is defined in units of queuing delay, not bytes, so that configuration remains invariant as the queue departure rate varies.¶
• It uses instantaneous queuing delay, which avoids the complexity of smoothing, but also avoids embedding a worst-case RTT of smoothing delay in the network (see Section 2.1).¶
• The ramp rises linearly directly from 0 to 1, not to an intermediate value of p'_L as RED would, because there is no need to keep ECN-marking probability low.¶
• Marking does not have to be randomized. Determinism is used instead of randomness to reduce the delay necessary to smooth out the noise of randomness from the signal.¶

The ramp function requires two configuration parameters, the minimum threshold (minTh) and the width of the ramp (range), both in units of queuing time, as shown in lines 17 and 18 of the initialization function in Figure 2. The ramp function can be configured as a step (see Note c).¶

Although the DCTCP paper [Alizadeh-stability] recommends an ECN-marking threshold of 0.17*RTT_typ, it also shows that the threshold can be much shallower with hardly any worse underutilization of the link (because the amplitude of DCTCP's sawteeth is so small). Based on extensive experiments, for the public Internet the default minimum ECN-marking threshold (target) in Figure 2 is considered a good compromise, even though it is a significantly smaller fraction of RTT_typ.¶

1:  laqm(qdelay) {               % Returns Native L4S AQM probability
2:    if (qdelay >= maxTh)
3:      return 1
4:    else if (qdelay > minTh)
5:      return (qdelay - minTh)/range  % Divide could use a bit-shift
6:    else
7:      return 0
8:  }

1:  dualpi2_update(lq, cq) {                % Update p' every Tupdate
2:    curq = cq.time()  % use queuing time of first-in Classic packet
3:    p' = p' + alpha * (curq - target) + beta * (curq - prevq)
4:    p_CL = k * p'  % Coupled L4S prob = base prob * coupling factor
5:    p_C = p'^2                       % Classic prob = (base prob)^2
6:    prevq = curq
7:  }

(Note: Clamping p' within the range [0,1] omitted for clarity -- see below.)¶

The coupled marking probability p_CL depends on the base probability (p'), which is kept up to date by executing the core PI algorithm in Figure 6 every Tupdate.¶

Note that p' solely depends on the queuing time in the Classic queue. In line 2, the current queuing delay (curq) is evaluated from how long the head packet was in the Classic queue (cq). The function cq.time() (not shown) subtracts the time stamped at enqueue from the current time (see Note a below) and implicitly takes the current queuing delay as 0 if the queue is empty.¶

The algorithm centres on line 3, which is a classical PI controller that alters p' dependent on: a) the error between the current queuing delay (curq) and the target queuing delay (target) and b) the change in queuing delay since the last sample. The name 'PI' represents the fact that the second factor (how fast the queue is growing) is Proportional to load while the first is the Integral of the load (so it removes any standing queue in excess of the target).¶

The target parameter can be set based on local knowledge, but the aim is for the default to be a good compromise for anywhere in the intended deployment environment -- the public Internet. According to [PI2param], the target queuing delay on line 8 of Figure 2 is related to the typical base RTT worldwide, RTT_typ, by two factors: target = RTT_typ * g * f. Below, we summarize the rationale behind these factors and introduce a further adjustment. The two factors ensure that, in a large proportion of cases (say 90%), the sawtooth variations in RTT of a single flow will fit within the buffer without underutilizing the link. Frankly, these factors are educated guesses, but with the emphasis closer to 'educated' than to 'guess' (see [PI2param] for the full background):¶

• RTT_typ is taken as 25 ms. This is based on an average CDN latency measured in each country weighted by the number of Internet users in that country to produce an overall weighted average for the Internet [PI2param]. Countries were ranked by number of Internet users, and once 90% of Internet users were covered, smaller countries were excluded to avoid small sample sizes that would be less representative. Also, importantly, the data for the average CDN latency in China (with the largest number of Internet users) has been removed, because the CDN latency was a significant outlier and, on reflection, the experimental technique seemed inappropriate to the CDN market in China.¶
• g is taken as 0.38. The factor g is a geometry factor that characterizes the shape of the sawteeth of prevalent Classic congestion controllers. The geometry factor is the fraction of the amplitude of the sawtooth variability in queue delay that lies below the AQM's target. For instance, at low bitrates, the geometry factor of standard Reno is 0.5, but at higher rates, it tends towards just under 1. According to the census of congestion controllers conducted by Mishra et al. in Jul-Oct 2019 [CCcensus19], most Classic TCP traffic uses CUBIC. And, according to the analysis in [PI2param], if running over a PI2 AQM, a large proportion of this CUBIC traffic would be in its Reno-friendly mode, which has a geometry factor of ~0.39 (for all known implementations). The rest of the CUBIC traffic would be in true CUBIC mode, which has a geometry factor of ~0.36. Without modelling the sawtooth profiles from all the other less prevalent congestion controllers, we estimate a 7:3 weighted average of these two, resulting in an average geometry factor of 0.38.¶
• f is taken as 2. The factor f is a safety factor that increases the target queue to allow for the distribution of RTT_typ around its mean. Otherwise, the target queue would only avoid underutilization for those users below the mean. It also provides a safety margin for the proportion of paths in use that span beyond the distance between a user and their local CDN. Currently, no data is available on the variance of queue delay around the mean in each region, so there is plenty of room for this guess to become more educated.¶
• [PI2param] recommends target = RTT_typ * g * f = 25 ms * 0.38 * 2 = 19 ms. However, a further adjustment is warranted, because target is moving year-on-year. The paper is based on data collected in 2019, and it mentions evidence from the Speedtest Global Index that suggests RTT_typ reduced by 17% (fixed) or 12% (mobile) between 2020 and 2021. Therefore, we recommend a default of target = 15 ms at the time of writing (2021).¶

Operators can always use the data and discussion in [PI2param] to configure a more appropriate target for their environment. For instance, an operator might wish to question the assumptions called out in that paper, such as the goal of no underutilization for a large majority of single flow transfers (given many large transfers use multiple flows to avoid the scaling limitations of Classic flows).¶

The two 'gain factors' in line 3 of Figure 6, alpha and beta, respectively weight how strongly each of the two elements (Integral and Proportional) alters p'. They are in units of 'per second of delay' or Hz, because they transform differences in queuing delay into changes in probability (assuming probability has a value from 0 to 1).¶

Alpha and beta determine how much p' ought to change after each update interval (Tupdate). For a smaller Tupdate, p' should change by the same amount per second but in finer more frequent steps. So alpha depends on Tupdate (see line 13 of the initialization function in Figure 2). It is best to update p' as frequently as possible, but Tupdate will probably be constrained by hardware performance. As shown in line 12, the update interval should be frequent enough to update at least once in the time taken for the target queue to drain ('target') as long as it updates at least three times per maximum RTT. Tupdate defaults to 16 ms in the reference Linux implementation because it has to be rounded to a multiple of 4 ms. For link rates from 4 to 200 Mb/s and a maximum RTT of 100 ms, it has been verified through extensive testing that Tupdate = 16 ms (as also recommended in the PIE spec [RFC8033]) is sufficient.¶

The choice of alpha and beta also determines the AQM's stable operating range. The AQM ought to change p' as fast as possible in response to changes in load without overcompensating and therefore causing oscillations in the queue. Therefore, the values of alpha and beta also depend on the RTT of the expected worst-case flow (RTT_max).¶

The maximum RTT of a PI controller (RTT_max in line 9 of Figure 2) is not an absolute maximum, but more instability (more queue variability) sets in for long-running flows with an RTT above this value. The propagation delay halfway round the planet and back in glass fibre is 200 ms. However, hardly any traffic traverses such extreme paths and, since the significant consolidation of Internet traffic between 2007 and 2009 [Labovitz10], a high and growing proportion of all Internet traffic (roughly two-thirds at the time of writing) has been served from CDNs or 'cloud' services distributed close to end users. The Internet might change again, but for now, designing for a maximum RTT of 100 ms is a good compromise between faster queue control at low RTT and some instability on the occasions when a longer path is necessary.¶

Recommended derivations of the gain constants alpha and beta can be approximated for Reno over a PI2 AQM as: alpha = 0.1 * Tupdate / RTT_max^2; beta = 0.3 / RTT_max, as shown in lines 13 and 14 of Figure 2. These are derived from the stability analysis in [PI2]. For the default values of Tupdate = 16 ms and RTT_max = 100 ms, they result in alpha = 0.16; beta = 3.2 (discrepancies are due to rounding). These defaults have been verified with a wide range of link rates, target delays, and traffic models with mixed and similar RTTs, short and long flows, etc.¶

In corner cases, p' can overflow the range [0,1] so the resulting value of p' has to be bounded (omitted from the pseudocode). Then, as already explained, the coupled and Classic probabilities are derived from the new p' in lines 4 and 5 of Figure 6 as p_CL = k*p' and p_C = p'^2.¶

Because the coupled L4S marking probability (p_CL) is factored up by k, the dynamic gain parameters alpha and beta are also inherently factored up by k for the L4S queue. So, the effective gain factor for the L4S queue is k*alpha (with defaults alpha = 0.16 Hz and k = 2, effective L4S alpha = 0.32 Hz).¶

Unlike in PIE [RFC8033], alpha and beta do not need to be tuned every Tupdate dependent on p'. Instead, in PI2, alpha and beta are independent of p' because the squaring applied to Classic traffic tunes them inherently. This is explained in [PI2], which also explains why this more principled approach removes the need for most of the heuristics that had to be added to PIE.¶

Nonetheless, an implementer might wish to add selected details to either AQM. For instance, the Linux reference DualPI2 implementation includes the following (not shown in the pseudocode above):¶

• Classic and coupled marking or dropping (i.e., based on p_C and p_CL from the PI controller) is not applied to a packet if the aggregate queue length in bytes is < 2 MTU (prior to enqueuing the packet or dequeuing it, depending on whether the AQM is configured to be applied at enqueue or dequeue); and¶
• in the WRR scheduler, the 'credit' indicating which queue should transmit is only changed if there are packets in both queues (i.e., if there is actual resource contention). This means that a properly paced L flow might never be delayed by the WRR. The WRR credit is reset in favour of the L queue when the link is idle.¶

An implementer might also wish to add other heuristics, e.g., burst protection [RFC8033] or enhanced burst protection [RFC8034].¶

Notes:¶

1. The drain rate of the queue can vary if it is scheduled relative to other queues or if it accommodates fluctuations in a wireless medium. To auto-adjust to changes in drain rate, the queue needs to be measured in time, not bytes or packets [AQMmetrics] [CoDel]. Queuing delay could be measured directly as the sojourn time (a.k.a. service time) of the queue by storing a per-packet timestamp as each packet is enqueued and subtracting it from the system time when the packet is dequeued. If timestamping is not easy to introduce with certain hardware, queuing delay could be predicted indirectly by dividing the size of the queue by the predicted departure rate, which might be known precisely for some link technologies (see, for example, DOCSIS PIE [RFC8034]).¶

   However, sojourn time is slow to detect bursts. For instance, if a burst arrives at an empty queue, the sojourn time only fully measures the burst's delay when its last packet is dequeued, even though the queue has known the size of the burst since its last packet was enqueued -- so it could have signalled congestion earlier. To remedy this, each head packet can be marked when it is dequeued based on the expected delay of the tail packet behind it, as explained below, rather than based on the head packet's own delay due to the packets in front of it. "Underutilization with Bursty Traffic" in [Heist21] identifies a specific scenario where bursty traffic significantly hits utilization of the L queue. If this effect proves to be more widely applicable, using the delay behind the head could improve performance.¶

   The delay behind the head can be implemented by dividing the backlog at dequeue by the link rate or equivalently multiplying the backlog by the delay per unit of backlog. The implementation details will depend on whether the link rate is known; if it is not, a moving average of the delay per unit backlog can be maintained. This delay consists of serialization as well as media acquisition for shared media. So the details will depend strongly on the specific link technology. This approach should be less sensitive to timing errors and cost less in operations and memory than the otherwise equivalent 'scaled sojourn time' metric, which is the sojourn time of a packet scaled by the ratio of the queue sizes when the packet departed and arrived [SigQ-Dyn].¶

2. Line 2 of the dualpi2_enqueue() function (Figure 3) assumes an implementation where lq and cq share common buffer memory. An alternative implementation could use separate buffers for each queue, in which case the arriving packet would have to be classified first to determine which buffer to check for available space. The choice is a trade-off; a shared buffer can use less memory whereas separate buffers isolate the L4S queue from tail drop due to large bursts of Classic traffic (e.g., a Classic Reno TCP during slow-start over a long RTT).¶

3. There has been some concern that using the step function of DCTCP for the Native L4S AQM requires end systems to smooth the signal for an unnecessarily large number of round trips to ensure sufficient fidelity. A ramp is no worse than a step in initial experiments with existing DCTCP. Therefore, it is recommended that a ramp is configured in place of a step, which will allow congestion control algorithms to investigate faster smoothing algorithms.¶

   A ramp is more general than a step, because an operator can effectively turn the ramp into a step function, as used by DCTCP, by setting the range to zero. There will not be a divide by zero problem at line 5 of Figure 5 because, if minTh is equal to maxTh, the condition for this ramp calculation cannot arise.¶

A.2. Pass #2: Edge-Case Details

This section takes a second pass through the pseudocode to add details of two edge-cases: low link rate and overload. Figure 7 repeats the dequeue function of Figure 4, but with details of both edge-cases added. Similarly, Figure 8 repeats the core PI algorithm of Figure 6, but with overload details added. The initialization, enqueue, L4S AQM, and recur functions are unchanged.¶

The link rate can be so low that it takes a single packet queue longer to serialize than the threshold delay at which ECN marking starts to be applied in the L queue. Therefore, a minimum marking threshold parameter in units of packets rather than time is necessary (Th_len, default 1 packet in line 19 of Figure 2) to ensure that the ramp does not trigger excessive marking on slow links. Where an implementation knows the link rate, it can set up this minimum at the time it is configured. For instance, it would divide 1 MTU by the link rate to convert it into a serialization time, then if the lower threshold of the Native L AQM ramp was lower than this serialization time, it could increase the thresholds to shift the bottom of the ramp to 2 MTU. This is the approach used in DOCSIS [DOCSIS3.1], because the configured link rate is dedicated to the DualQ.¶

The pseudocode given here applies where the link rate is unknown, which is more common for software implementations that might be deployed in scenarios where the link is shared with other queues. In lines 5a to 5d in Figure 7, the native L4S marking probability, p'_L, is zeroed if the queue is only 1 packet (in the default configuration).¶

Persistent overload is deemed to have occurred when Classic drop/marking probability reaches p_Cmax. Above this point, the Classic drop probability is applied to both the L and C queues, irrespective of whether any packet is ECN-capable. ECT packets that are not dropped can still be ECN-marked.¶

In line 11 of the initialization function (Figure 2), the maximum Classic drop probability p_Cmax = min(1/k^2, 1) or 1/4 for the default coupling factor k = 2. In practice, 25% has been found to be a good threshold to preserve fairness between ECN-capable and non-ECN-capable traffic. This protects the queues against both temporary overload from responsive flows and more persistent overload from any unresponsive traffic that falsely claims to be responsive to ECN.¶

When the Classic ECN-marking probability reaches the p_Cmax threshold (1/k^2), the marking probability that is coupled to the L4S queue, p_CL, will always be 100% for any k (by equation (1) in Section 2.1). So, for readability, the constant p_Lmax is defined as 1 in line 21 of the initialization function (Figure 2). This is intended to ensure that the L4S queue starts to introduce dropping once ECN marking saturates at 100% and can rise no further. The 'Prague L4S requirements' [RFC9331] state that when an L4S congestion control detects a drop, it falls back to a response that coexists with 'Classic' Reno congestion control. So, it is correct that when the L4S queue drops packets, it drops them proportional to p'^2, as if they are Classic packets.¶

The two queues each test for overload in lines 4b and 12b of the dequeue function (Figure 7). Lines 8c to 8g drop L4S packets with probability p'^2. Lines 8h to 8i mark the remaining packets with probability p_CL. Given p_Lmax = 1, all remaining packets will be marked because, to have reached the else block at line 8b, p_CL >= 1.¶

Line 2a in the core PI algorithm (Figure 8) deals with overload of the L4S queue when there is little or no Classic traffic. This is necessary, because the core PI algorithm maintains the appropriate drop probability to regulate overload, but it depends on the length of the Classic queue. If there is little or no Classic queue, the naive PI-update function (Figure 6) would drop nothing, even if the L4S queue were overloaded -- so tail drop would have to take over (lines 2 and 3 of Figure 3).¶

Instead, line 2a of the full PI-update function (Figure 8) ensures that the Base PI AQM in line 3 is driven by whichever of the two queue delays is greater, but line 3 still always uses the same Classic target (default 15 ms). If L queue delay is greater just because there is little or no Classic traffic, normally it will still be well below the Base AQM target. This is because L4S traffic is also governed by the shallow threshold of its own Native AQM (lines 5a to 6 of the dequeue algorithm in Figure 7). So the Base AQM will be driven to zero and not contribute. However, if the L queue is overloaded by traffic that is unresponsive to its marking, the max() in line 2a of Figure 8 enables the L queue to smoothly take over driving the Base AQM into overload mode even if there is little or no Classic traffic. Then the Base AQM will keep the L queue to the Classic target (default 15 ms) by shedding L packets.¶

1:  dualpi2_dequeue(lq, cq, pkt) {     % Couples L4S & Classic queues
2:    while ( lq.byt() + cq.byt() > 0 ) {
3:      if ( scheduler() == lq ) {
4a:       lq.dequeue(pkt)                             % L4S scheduled
4b:       if ( p_CL < p_Lmax ) {      % Check for overload saturation
5a:         if (lq.len()>Th_len)                   % >1 packet queued
5b:           p'_L = laqm(lq.time())                    % Native LAQM
5c:         else
5d:           p'_L = 0                 % Suppress marking 1 pkt queue
6:          p_L = max(p'_L, p_CL)                % Combining function
7:          if ( recur(lq, p_L)                       %Linear marking
8a:           mark(pkt)
8b:       } else {                              % overload saturation
8c:         if ( recur(lq, p_C) ) {          % probability p_C = p'^2
8e:           drop(pkt)      % revert to Classic drop due to overload
8f:           continue        % continue to the top of the while loop
8g:         }
8h:         if ( recur(lq, p_CL) )        % probability p_CL = k * p'
8i:           mark(pkt)         % linear marking of remaining packets
8j:       }
9:      } else {
10:       cq.dequeue(pkt)                         % Classic scheduled
11:       if ( recur(cq, p_C) ) {            % probability p_C = p'^2
12a:        if ( (ecn(pkt) == 0)                % ECN field = not-ECT
12b:             OR (p_C >= p_Cmax) ) {       % Overload disables ECN
13:           drop(pkt)                     % squared drop, redo loop
14:           continue        % continue to the top of the while loop
15:         }
16:         mark(pkt)                                  % squared mark
17:       }
18:     }
19:     return(pkt)                      % return the packet and stop
20:   }
21:   return(NULL)                             % no packet to dequeue
22: }

1:  dualpi2_update(lq, cq) {                % Update p' every Tupdate
2a:   curq = max(cq.time(), lq.time())    % use greatest queuing time
3:    p' = p' + alpha * (curq - target) + beta * (curq - prevq)
4:    p_CL = p' * k  % Coupled L4S prob = base prob * coupling factor
5:    p_C = p'^2                       % Classic prob = (base prob)^2
6:    prevq = curq
7:  }

The choice of scheduler technology is critical to overload protection (see Section 4.2.2).¶

• A well-understood weighted scheduler such as WRR is recommended. As long as the scheduler weight for Classic is small (e.g., 1/16), its exact value is unimportant, because it does not normally determine capacity shares. The weight is only important to prevent unresponsive L4S traffic starving Classic traffic in the short term (see Section 4.2.2). This is because capacity sharing between the queues is normally determined by the coupled congestion signal, which overrides the scheduler, by making L4S sources leave roughly equal per-flow capacity available for Classic flows.¶

• Alternatively, a time-shifted FIFO (TS-FIFO) could be used. It works by selecting the head packet that has waited the longest, biased against the Classic traffic by a time-shift of tshift. To implement TS-FIFO, the scheduler() function in line 3 of the dequeue code would simply be implemented as the scheduler() function at the bottom of Figure 10 in Appendix B. For the public Internet, a good value for tshift is 50 ms. For private networks with smaller diameter, about 4*target would be reasonable. TS-FIFO is a very simple scheduler, but complexity might need to be added to address some deficiencies (which is why it is not recommended over WRR):¶

  • TS-FIFO does not fully isolate latency in the L4S queue from uncontrolled bursts in the Classic queue;¶
  • using sojourn time for TS-FIFO is only appropriate if timestamping of packets is feasible; and¶
  • even if timestamping is supported, the sojourn time of the head packet is always stale, so a more instantaneous measure of queue delay could be used (see Note a in Appendix A.1).¶
• A strict priority scheduler would be inappropriate as discussed in Section 4.2.2.¶

B.1. Curvy RED in Pseudocode

The pseudocode manipulates three main structures of variables: the packet (pkt), the L4S queue (lq), and the Classic queue (cq). It is defined and described below in the following three functions:¶

• the initialization function cred_params_init(...) (Figure 2) that sets parameter defaults (the API for setting non-default values is omitted for brevity);¶
• the dequeue function cred_dequeue(lq, cq, pkt) (Figure 4); and¶
• the scheduling function scheduler(), which selects between the head packets of the two queues.¶

It also uses the following functions that are either shown elsewhere or not shown in full here:¶

• the enqueue function, which is identical to that used for DualPI2, dualpi2_enqueue(lq, cq, pkt) in Figure 3;¶
• mark(pkt) and drop(pkt) for ECN marking and dropping a packet;¶
• cq.byt() or lq.byt() returns the current length (a.k.a. backlog) of the relevant queue in bytes; and¶
• cq.time() or lq.time() returns the current queuing delay of the relevant queue in units of time (see Note a in Appendix A.1).¶

Because Curvy RED was evaluated before DualPI2, certain improvements introduced for DualPI2 were not evaluated for Curvy RED. In the pseudocode below, the straightforward improvements have been added on the assumption they will provide similar benefits, but that has not been proven experimentally. They are: i) a conditional priority scheduler instead of strict priority; ii) a time-based threshold for the Native L4S AQM; and iii) ECN support for the Classic AQM. A recent evaluation has proved that a minimum ECN-marking threshold (minTh) greatly improves performance, so this is also included in the pseudocode.¶

Overload protection has not been added to the Curvy RED pseudocode below so as not to detract from the main features. It would be added in exactly the same way as in Appendix A.2 for the DualPI2 pseudocode. The Native L4S AQM uses a step threshold, but a ramp like that described for DualPI2 could be used instead. The scheduler uses the simple TS-FIFO algorithm, but it could be replaced with WRR.¶

The Curvy RED algorithm has not been maintained or evaluated to the same degree as the DualPI2 algorithm. In initial experiments on broadband access links ranging from 4 Mb/s to 200 Mb/s with base RTTs from 5 ms to 100 ms, Curvy RED achieved good results with the default parameters in Figure 9.¶

The parameters are categorized by whether they relate to the Classic AQM, the L4S AQM, or the framework coupling them together. Constants and variables derived from these parameters are also included at the end of each category. These are the raw input parameters for the algorithm. A configuration front-end could accept more meaningful parameters (e.g., RTT_max and RTT_typ) and convert them into these raw parameters, as has been done for DualPI2 in Appendix A. Where necessary, parameters are explained further in the walk-through of the pseudocode below.¶

1:  cred_params_init(...) {            % Set input parameter defaults
2:    % DualQ Coupled framework parameters
3:    limit = MAX_LINK_RATE * 250 ms               % Dual buffer size
4:    k' = 1                        % Coupling factor as a power of 2
5:    tshift = 50 ms                % Time-shift of TS-FIFO scheduler
6:    % Constants derived from Classic AQM parameters
7:    k = 2^k'                    % Coupling factor from equation (1)
6:
7:    % Classic AQM parameters
8:    g_C = 5            % EWMA smoothing parameter as a power of 1/2
9:    S_C = -1          % Classic ramp scaling factor as a power of 2
10:   minTh = 500 ms    % No Classic drop/mark below this queue delay
11:   % Constants derived from Classic AQM parameters
12:   gamma = 2^(-g_C)                     % EWMA smoothing parameter
13:   range_C = 2^S_C                         % Range of Classic ramp
14:
15:   % L4S AQM parameters
16:   T = 1 ms             % Queue delay threshold for Native L4S AQM
17:   % Constants derived from above parameters
18:   S_L = S_C - k'        % L4S ramp scaling factor as a power of 2
19:   range_L = 2^S_L                             % Range of L4S ramp
20: }

1:  cred_dequeue(lq, cq, pkt) {       % Couples L4S & Classic queues
2:    while ( lq.byt() + cq.byt() > 0 ) {
3:      if ( scheduler() == lq ) {
4:        lq.dequeue(pkt)                            % L4S scheduled
5a:       p_CL = (Q_C - minTh) / range_L
5b:       if (  ( lq.time() > T )
5c:          OR ( p_CL > maxrand(U) ) )
6:          mark(pkt)
7:      } else {
8:        cq.dequeue(pkt)                        % Classic scheduled
9a:       Q_C = gamma * cq.time() + (1-gamma) * Q_C % Classic Q EWMA
10a:      sqrt_p_C = (Q_C - minTh) / range_C
10b:      if ( sqrt_p_C > maxrand(2*U) ) {
11:         if ( (ecn(pkt) == 0)  {            % ECN field = not-ECT
12:           drop(pkt)                    % Squared drop, redo loop
13:           continue       % continue to the top of the while loop
14:         }
15:         mark(pkt)
16:       }
17:     }
18:     return(pkt)                % return the packet and stop here
19:   }
20:   return(NULL)                            % no packet to dequeue
21: }

22: maxrand(u) {                % return the max of u random numbers
23:   maxr=0
24:   while (u-- > 0)
25:     maxr = max(maxr, rand())                   % 0 <= rand() < 1
26:   return(maxr)
27: }

28: scheduler() {
29:   if ( lq.time() + tshift >= cq.time() )
30:     return lq;
31:   else
32:     return cq;
33: }

The dequeue pseudocode (Figure 10) is repeatedly called whenever the lower layer is ready to forward a packet. It schedules one packet for dequeuing (or zero if the queue is empty) then returns control to the caller so that it does not block while that packet is being forwarded. While making this dequeue decision, it also makes the necessary AQM decisions on dropping or marking. The alternative of applying the AQMs at enqueue would shift some processing from the critical time when each packet is dequeued. However, it would also add a whole queue of delay to the control signals, making the control loop very sloppy.¶

The code is written assuming the AQMs are applied on dequeue (Note 1). All the dequeue code is contained within a large while loop so that if it decides to drop a packet, it will continue until it selects a packet to schedule. If both queues are empty, the routine returns NULL at line 20. Line 3 of the dequeue pseudocode is where the conditional priority scheduler chooses between the L4S queue (lq) and the Classic queue (cq). The TS-FIFO scheduler is shown at lines 28-33, which would be suitable if simplicity is paramount (see Note 2).¶

Within each queue, the decision whether to forward, drop, or mark is taken as follows (to simplify the explanation, it is assumed that U = 1):¶

L4S:

If the test at line 3 determines there is an L4S packet to dequeue, the tests at lines 5b and 5c determine whether to mark it. The first is a simple test of whether the L4S queue delay (lq.time()) is greater than a step threshold T (Note 3). The second test is similar to the random ECN marking in RED but with the following differences: i) marking depends on queuing time, not bytes, in order to scale for any link rate without being reconfigured; ii) marking of the L4S queue depends on a logical OR of two tests: one against its own queuing time and one against the queuing time of the other (Classic) queue; iii) the tests are against the instantaneous queuing time of the L4S queue but against a smoothed average of the other (Classic) queue; and iv) the queue is compared with the maximum of U random numbers (but if U = 1, this is the same as the single random number used in RED).¶

Specifically, in line 5a, the coupled marking probability p_CL is set to the amount by which the averaged Classic queuing delay Q_C exceeds the minimum queuing delay threshold (minTh), all divided by the L4S scaling parameter range_L. range_L represents the queuing delay (in seconds) added to minTh at which marking probability would hit 100%. Then, in line 5c (if U = 1), the result is compared with a uniformly distributed random number between 0 and 1, which ensures that, over range_L, marking probability will linearly increase with queuing time.¶

Classic:

If the scheduler at line 3 chooses to dequeue a Classic packet and jumps to line 7, the test at line 10b determines whether to drop or mark it. But before that, line 9a updates Q_C, which is an exponentially weighted moving average (Note 4) of the queuing time of the Classic queue, where cq.time() is the current instantaneous queuing time of the packet at the head of the Classic queue (zero if empty), and gamma is the exponentially weighted moving average (EWMA) constant (default 1/32; see line 12 of the initialization function).¶

Lines 10a and 10b implement the Classic AQM. In line 10a, the averaged queuing time Q_C is divided by the Classic scaling parameter range_C, in the same way that queuing time was scaled for L4S marking. This scaled queuing time will be squared to compute Classic drop probability. So, before it is squared, it is effectively the square root of the drop probability; hence, it is given the variable name sqrt_p_C. The squaring is done by comparing it with the maximum out of two random numbers (assuming U = 1). Comparing it with the maximum out of two is the same as the logical 'AND' of two tests, which ensures drop probability rises with the square of queuing time.¶

The AQM functions in each queue (lines 5c and 10b) are two cases of a new generalization of RED called 'Curvy RED', motivated as follows. When the performance of this AQM was compared with FQ-CoDel and PIE, their goal of holding queuing delay to a fixed target seemed misguided [CRED_Insights]. As the number of flows increases, if the AQM does not allow host congestion controllers to increase queuing delay, it has to introduce abnormally high levels of loss. Then loss rather than queuing becomes the dominant cause of delay for short flows, due to timeouts and tail losses.¶

Curvy RED constrains delay with a softened target that allows some increase in delay as load increases. This is achieved by increasing drop probability on a convex curve relative to queue growth (the square curve in the Classic queue, if U = 1). Like RED, the curve hugs the zero axis while the queue is shallow. Then, as load increases, it introduces a growing barrier to higher delay. But, unlike RED, it requires only two parameters, not three. The disadvantage of Curvy RED (compared to a PI controller, for example) is that it is not adapted to a wide range of RTTs. Curvy RED can be used as is when the RTT range to be supported is limited; otherwise, an adaptation mechanism is needed.¶

From our limited experiments with Curvy RED so far, recommended values of these parameters are: S_C = -1; g_C = 5; T = 5 * MTU at the link rate (about 1 ms at 60 Mb/s) for the range of base RTTs typical on the public Internet. [CRED_Insights] explains why these parameters are applicable whatever rate link this AQM implementation is deployed on and how the parameters would need to be adjusted for a scenario with a different range of RTTs (e.g., a data centre). The setting of k depends on policy (see Section 2.5 and Appendix C.2, respectively, for its recommended setting and guidance on alternatives).¶

There is also a cUrviness parameter, U, which is a small positive integer. It is likely to take the same hard-coded value for all implementations, once experiments have determined a good value. Only U = 1 has been used in experiments so far, but results might be even better with U = 2 or higher.¶

Notes:¶

1. The alternative of applying the AQMs at enqueue would shift some processing from the critical time when each packet is dequeued. However, it would also add a whole queue of delay to the control signals, making the control loop sloppier (for a typical RTT, it would double the Classic queue's feedback delay). On a platform where packet timestamping is feasible, e.g., Linux, it is also easiest to apply the AQMs at dequeue, because that is where queuing time is also measured.¶
2. WRR better isolates the L4S queue from large delay bursts in the Classic queue, but it is slightly less simple than TS-FIFO. If WRR were used, a low default Classic weight (e.g., 1/16) would need to be configured in place of the time-shift in line 5 of the initialization function (Figure 9).¶
3. A step function is shown for simplicity. A ramp function (see Figure 5 and the discussion around it in Appendix A.1) is recommended, because it is more general than a step and has the potential to enable L4S congestion controls to converge more rapidly.¶
4. An EWMA is only one possible way to filter bursts; other more adaptive smoothing methods could be valid, and it might be appropriate to decrease the EWMA faster than it increases, e.g., by using the minimum of the smoothed and instantaneous queue delays, min(Q_C, qc.time()).¶