<?xml version='1.0' encoding='utf-8'?>
<!DOCTYPE rfc [
  <!ENTITY nbsp    "&#160;">
  <!ENTITY zwsp   "&#8203;">
  <!ENTITY nbhy   "&#8209;">
  <!ENTITY wj     "&#8288;">
]>
<?xml-stylesheet type="text/xsl" href="rfc2629.xslt" ?>
<!-- generated by https://github.com/cabo/kramdown-rfc version 1.7.43 (Ruby 3.2.3) -->
<rfc xmlns:xi="http://www.w3.org/2001/XInclude" ipr="trust200902" docName="draft-irtf-iccrg-pacing-05" category="info" consensus="true" submissionType="IRTF" tocInclude="true" sortRefs="true" symRefs="true" version="3">
  <!-- xml2rfc v2v3 conversion 3.35.0 -->
  <front>
    <title abbrev="Pacing in Transport Protocols">Pacing in Transport Protocols</title>
    <seriesInfo name="Internet-Draft" value="draft-irtf-iccrg-pacing-05"/>
    <author initials="M." surname="Welzl" fullname="Michael Welzl">
      <organization>University of Oslo</organization>
      <address>
        <postal>
          <street>PO Box 1080 Blindern</street>
          <city>0316  Oslo</city>
          <country>Norway</country>
        </postal>
        <email>michawe@ifi.uio.no</email>
        <uri>http://welzl.at/</uri>
      </address>
    </author>
    <author initials="W." surname="Eddy" fullname="Wesley Eddy">
      <organization>Aalyria Technologies</organization>
      <address>
        <postal>
          <city>Avon, OH 44011</city>
          <country>United States of America</country>
        </postal>
        <email>wes@aalyria.com</email>
      </address>
    </author>
    <author initials="V." surname="Goel" fullname="Vidhi Goel">
      <organization>Apple Inc.</organization>
      <address>
        <postal>
          <street>One Apple Park Way</street>
          <city>Cupertino, California 95014</city>
          <country>United States of America</country>
        </postal>
        <email>vidhi_goel@apple.com</email>
      </address>
    </author>
    <author initials="M." surname="Tüxen" fullname="Michael Tüxen">
      <organization>Münster University of Applied Sciences</organization>
      <address>
        <postal>
          <street>Stegerwaldstrasse 39</street>
          <city>48565 Steinfurt</city>
          <country>Germany</country>
        </postal>
        <email>tuexen@fh-muenster.de</email>
      </address>
    </author>
    <date year="2026" month="November" day="16"/>
    <area>IRTF</area>
    <workgroup>Internet Congestion Control</workgroup>
    <keyword>pacing</keyword>
    <keyword>congestion control</keyword>
    <abstract>
      <?line 187?>

<t>Applications or congestion control mechanisms can produce bursty traffic, which can cause unnecessary queuing and packet loss. To reduce the burstiness of traffic, the concept of evenly spacing out the traffic from a data sender over a round-trip time known as "pacing" has been used in many transport protocol implementations. This document gives an overview of pacing and how some known pacing implementations work. It is a product of the Internet Congestion Control Research Group (ICCRG) of the Internet Research Task Force (IRTF).</t>
    </abstract>
    <note removeInRFC="true">
      <name>About This Document</name>
      <t>
        The latest revision of this draft can be found at <eref target="https://mwelzl.github.io/draft-iccrg-pacing/draft-irtf-iccrg-pacing.html"/>.
        Status information for this document may be found at <eref target="https://datatracker.ietf.org/doc/draft-irtf-iccrg-pacing/"/>.
      </t>
      <t>
        Discussion of this document takes place on the
        Internet Congestion Control Research Group mailing list (<eref target="mailto:iccrg@irtf.org"/>),
        which is archived at <eref target="https://mailarchive.ietf.org/arch/browse/iccrg"/>.
        Subscribe at <eref target="https://www.ietf.org/mailman/listinfo/iccrg/"/>.
      </t>
      <t>Source for this draft and an issue tracker can be found at
        <eref target="https://github.com/mwelzl/draft-iccrg-pacing"/>.</t>
    </note>
  </front>
  <middle>
    <?line 192?>

<section anchor="introduction">
      <name>Introduction</name>
      <t>Applications commonly generate either bulk data (e.g., files) or bursts of data (e.g., segments of media) that transport protocols deliver into the network based on congestion control algorithms.</t>
      <t>RFCs describing congestion control generally refer to a congestion window (cwnd) state variable as an upper limit for either the number of unacknowledged packets or bytes that a sender is allowed to emit. This limits the sender's transmission rate at the granularity of a round-trip time (RTT). If the sender transmits the entire cwnd sized data in an instant, this can result in unnecessarily high queuing and eventually packet losses at the bottleneck. Such consequences are detrimental to users' applications in terms of both responsiveness and goodput. To solve this problem, the concept of pacing was introduced. Pacing allows a sender to send a cwnd of data, but to spread transmission more evenly across a round-trip time.</t>
      <t>Congestion control specifications always allow to send less than the cwnd, or temporarily emit packets at a lower rate. Accordingly, it is in line with these specifications to pace packets. Pacing is known to have advantages -- if some packets arrive at a bottleneck as a burst (all packets sent back-to-back), loss can be more likely than in a case where there are time gaps between packets (e.g., when they are spread over the RTT). It also means that pacing is less likely to cause any sudden, ephemeral increases in queuing delay. Since keeping the queues short reduces packet losses, pacing can also yield higher goodput by reducing the time lost in loss recovery.</t>
      <t>Because of its known advantages, pacing has become common in implementations of congestion controlled transports. It is also an integral element of the "BBR" congestion control mechanism <xref target="I-D.ietf-ccwg-bbr"/>.</t>
      <t>This document provides an overview of pacing and describes how it is implemented in several well-known systems. It is a product of the IRTF Internet Congestion Control Research Group (ICCRG), and represents the consensus of this group, based on reviews by several members of the group, discussions in the mailing list and in the group's meetings. This document is not an IETF product and is not an Internet Standards Track specification.</t>
    </section>
    <section anchor="conventions-and-definitions">
      <name>Conventions and Definitions</name>
      <t>The key words "<bcp14>MUST</bcp14>", "<bcp14>MUST NOT</bcp14>", "<bcp14>REQUIRED</bcp14>", "<bcp14>SHALL</bcp14>", "<bcp14>SHALL
NOT</bcp14>", "<bcp14>SHOULD</bcp14>", "<bcp14>SHOULD NOT</bcp14>", "<bcp14>RECOMMENDED</bcp14>", "<bcp14>NOT RECOMMENDED</bcp14>",
"<bcp14>MAY</bcp14>", and "<bcp14>OPTIONAL</bcp14>" in this document are to be interpreted as
described in BCP 14 <xref target="RFC2119"/> <xref target="RFC8174"/> when, and only when, they
appear in all capitals, as shown here.</t>
      <?line -18?>

</section>
    <section anchor="motivations-for-pacing">
      <name>Motivations for Pacing</name>
      <t>Pacing is an old idea that did not see much deployment for decades. This may be
due to the need for efficient fine-grain timers, which were not previously available
in software. Also, at least one early analysis has documented disadvantages of
pacing, primarily in terms of throughput <xref target="UnderstandingPacing"/>. At the time of
writing, this article has become 25 years old; it is limited to Reno congestion control,
and defines a pacing method that is not in line with any of the implementations that
we document in <xref target="implementation"/>. A part of the critical analysis in the article
includes an example of a double back-off after slow start (fig. 8); the authors
attribute the performance drop to synchronization, but it may instead be
caused by the back-off factor beta being too large, as described in <xref target="backoff"/>.
In the latter case, it is not a problem with pacing per se.</t>
      <t>There are several potential benefits to pacing, both for the end-host protocol stacks
and within the network.  This section provides a short summary of the
motivations for performing pacing, with specific examples worked through in the
following (<xref target="considerations"/>).</t>
      <section anchor="network-benefits">
        <name>Network Benefits</name>
        <t>Senders generating large bursts create challenges for network queue management
to maintain low latencies, low loss rates, and low correlated bursts of loss.
This is described in more detail in <xref target="losstypes"/>, with examples.</t>
        <t>A number of causes within the network may lead to "ACK compression", where the
spacing of incoming packets (with new ACKs) becomes bunched.  This could happen
due to many factors, such as congestion at a bottleneck, packet send
aggregation in the MAC layer or device drivers, etc.  ACKs can also be aggregated such that instead
of acknowledging one or two packets of data, a received ACK may cover many packets,
and cause a large change in the congestion window, allowing many packets to be
released in a burst (if pacing is not used).  This can happen due to coalescing of
ACKs or ACK "thinning" within the network, or as a means to deal with highly
asymmetric connectivity.  In any case, a sender that performs pacing is not
susceptible to ACK compression, aggregation, or thinning causing its own
sending patterns to become bursty in turn, which allows the network more latitude
in how ACKs are handled.</t>
      </section>
      <section anchor="end-host-benefits">
        <name>End Host Benefits</name>
        <t>Pacing enables good operation with shorter queues, and can create an incentive
to reduce the size of queues being configured within the network, leading to
lower maximum latency for end host applications. When applications use pacing
to limit the rate, this can also reduce latency irrespective of the
size of the queues available in the network (see <xref target="appcontrol"/>).</t>
        <t>Improved RTT measurements can result from pacing, since samples are spread out
further across time rather than clumped in time, as explained in <xref target="rtt"/>.</t>
        <t>At a receiver, processing of the received packets is impacted by pacing, since
it will result in a more steady workload, rather than large incoming bursts of
data.  For some applications, pacing can therefore be beneficial in avoiding long periods
where the system is busy devoted to processing incoming data, and unable to
perform other work.  However, some systems are designed such that incoming
bursts are more efficient to process than a steadily paced stream, so benefits
may differ.</t>
      </section>
      <section anchor="other-motivations">
        <name>Other Motivations</name>
        <t>In some special situations, outside general Internet usage, the path properties may
be well-known in advance (e.g., due to scheduling of capacity, etc.).  In this
case, senders should pace packets at the scheduled rates in order to
efficiently utilize that capacity.  In some of these cases, the rates may be
very high, and any sender burstiness might require large expensive buffers
within the network to accommodate bursts without losses.  Situations where this
applies may include supercomputing grids, private datacenter interconnection,
and space mission communications <xref target="I-D.draft-many-tiptop-usecase"/>.</t>
      </section>
    </section>
    <section anchor="considerations">
      <name>Pacing: general considerations and consequences</name>
      <t>This section explores pacing scenarios in more detail, explains considerations
important for using and tuning pacing, and the resulting consequences.</t>
      <section anchor="losstypes">
        <name>More likely to saturate a bottleneck</name>
        <t>Two reasons for packet losses that are due to congestion at a bottleneck with a DropTail (FIFO) queue can be distinguished:</t>
        <ol spacing="normal" type="1"><li>
            <t>A flight of N packets arrives. The amount of data in this flight exceeds the amount of data that can be transmitted by the bottleneck during the flight's arrival plus the queue length, i.e. some data do not fit into the queue.</t>
          </li>
          <li>
            <t>The bottleneck is fully saturated. The queue is full, and packets drain from it more slowly than new packets arrive.</t>
          </li>
        </ol>
        <t>The second type of loss matches the typical expectation of a congestion control algorithm: the cwnd value when loss happens is indicative of the bottleneck being fully saturated. When the first type of loss happens, however, a sender's cwnd can be much smaller than the Bandwidth-Delay Product (BDP) of the path (the amount of data that can be in flight, ignoring the queue). In the absence of other traffic, the probability for the first type of loss depends on the queue length and the ratio between the departure and the arrival rate during the flight's arrival. By introducing time gaps between the packets of a burst, this ratio is increased, i.e. the difference between the departure and the arrival rate becomes smaller, and the second type of loss is more likely.</t>
        <t>For example, consider a network path with a bottleneck capacity of 50 Mbit/s, a queue length of 15000 bytes (or 10 packets of size 1500 bytes) and an RTT of 30 ms. Assume that all packets emitted by the sender have a size of 1500 bytes. Then, the BDP equals 125 packets. The bottleneck of this network path is fully saturated when a (BDP + queue length) amount of bytes are in flight: 135 packets.</t>
        <t>In this network, the first type of loss can happen as follows: say, N=40 packets arrive at this bottleneck at a rate of 100 Mbit/s. In an otherwise empty network and assuming an initial window of 10 packets and no delayed ACKs, this occurs in the third round of slow start without pacing, provided that the capacities of all links before the bottleneck are at least 100 Mbit/s.
In this case, an overshoot will occur: packets are forwarded with half their arrival rate, i.e. less than 20 packets can be forwarded during the burst's arrival. The remaining 20 (or more) packets cannot fit into the 10-packet queue. A cwnd of 40 packets is much smaller than the (BDP + queue) limit of 135 packets, and the bottleneck is not fully saturated.</t>
        <t>Let us now assume that the flight of 40 packets is instead paced, such that the arrival rate only mildly exceeds the departure rate -- e.g., they arrive at a rate of 60 Mbit/s. When the last packet of this flight arrives at the bottleneck, the bottleneck should already have forwarded 5/6 * 39 = 32.5 packets. Since only complete packets can be sent, the bottleneck has really forwarded 32 packets, and the remaining 40-32 = 8 packets fit in the queue. No loss occurs.</t>
        <t>This example explains how pacing can enable a rate increase to last longer than without pacing. This makes it more likely that a bottleneck is saturated, such that cwnd reflects the BDP plus the queue length (loss type 2).</t>
        <section anchor="backoff">
          <name>Backing off after the increase</name>
          <t>The two loss types explained in <xref target="losstypes"/> require a different back-off factor to allow the queue to drain and congestion to dissipate. Specifically, in the single-sender single-bottleneck example above, when a slow start overshoot occurs as loss type 2, a back-off function such as: ssthresh = cwnd * beta with beta &gt;= 0.5 is guaranteed to cause a second loss after the end of loss recovery. This is because, when cwnd exceeds a fully saturated bottleneck (i.e., cwnd &gt; BDP + queue length), cwnd will have grown further by another (BDP + queue length) by the time the sender learns about the loss. In this case, beta = 0.5 will cause ssthresh to exceed (BDP + queue length) again.</t>
          <t>Since pacing makes loss type 2 more likely, beta &lt; 0.5 may be a better choice after slow start overshoot when pacing is used.</t>
          <t>The following example illustrates this: consider a TCP sender that transmits data across a single bottleneck router that uses a FIFO queue towards a TCP receiver; assume that the sender is paced well, and only loss type 2 happens. For simplicity, we consider the congestion window in units of segments rather than bytes. The initial window (IW) is 10 segments, and the path's capacity limit (the BDP plus the bottleneck queue length) equals 30 segments. In slow start, after receiving ACKs for the first 10 segments, the sender will have transmitted 20 more segments and the value of cwnd will be 20. After receiving ACKs for these 20 segments, the sender will have transmitted 40 more segments and the value of cwnd will be 40. The first 30 of the 40 newly transmitted segments will pass through the bottleneck, but the remaining 10 segments will be dropped. The ACKs that are caused by the first 30 segments then cause the sender to transmit another 60 segments, and cwnd will be increased to 70. When the first of these 60 segments arrive at the receiver, they cause DupACKs; when these DupACKs arrive the sender, backing off with beta=0.5 yields ssthresh = 35, which is more than the path's capacity limit, and another loss will occur.</t>
        </section>
        <section anchor="able-to-work-with-smaller-queues">
          <name>Able to work with smaller queues</name>
          <t>The probability of loss type 1 in <xref target="losstypes"/> is indirectly proportional to the queue length. Pacing therefore enables a rate increase to continue with a smaller queue at the bottleneck than in the case without pacing.</t>
        </section>
      </section>
      <section anchor="queue-dynamics">
        <name>Queue dynamics</name>
        <t>When it enters the queue at a network bottleneck, unpaced traffic causes more sudden, drastic delay growth than paced traffic, and has a higher risk of packet loss, as discussed in <xref target="losstypes"/>. Paced traffic, on the other hand, can cause a bottleneck queue to grow more slowly and steadily, incurring delay growth over a longer time interval. Aside from the direct problems that delay can cause, such sustained queue and delay growth is also more likely to provoke an Active Queue Management (AQM) algorithm <xref target="RFC7567"/> to drop packets or mark them using Explicit Congestion Notification (ECN) <xref target="RFC8087"/>. This is because AQM algorithms are commonly designed to allow short, transient traffic bursts to pass unharmed, but react upon longer-term average queue growth.</t>
      </section>
      <section anchor="rtt">
        <name>Getting good RTT estimates</name>
        <t>Since pacing algorithms generally attempt to spread out packets evenly across an RTT, it is important to have a good RTT estimate. Especially in the beginning of a transfer, when sending the initial window, the only RTT estimate available may be from the connection establishment handshake. Being based on only one sample, this is a very unreliable estimate. Moreover, a new transport connection may be preceded by a longer period of quiescence on the path between two endpoints than there  might normally occur when a connection is active. Such a silence period can provoke behavior of lower layers that increases the RTT. For example, idle periods commonly cause a handshake procedure on 5G links before communication can continue, inflating the RTT.</t>
        <t>Thus, using this sample to pace the initial window can cause the pacing rate to become unnecessarily low. Accordingly, the Linux TCP implementation does not pace the first 10 packets (see <xref target="linux"/>). As a possible improvement, the initial RTT estimate could also be based on a previous connection (temporal sharing) or on another ongoing connection (ensemble sharing) <xref target="RFC9040"/>.</t>
        <t>Since having an accurate RTT estimate is important for pacing also after the initial round trip, delayed ACKs can be detrimental to pacing in two ways: first, by potentially reducing the number of ACKs, they can reduce the sampling frequency. Second, by adding delay to the transmission of ACKs, they can worsen the quality of the signal. Here, QUIC has a benefit over TCP: since a host reports the local delay between data reception and generation of the corresponding ACK, QUIC's RTT calculation can be more precise <xref target="RFC9000"/>.</t>
      </section>
      <section anchor="mini-bursts-and-their-trade-offs">
        <name>Mini-bursts and their trade-offs</name>
        <t>Generally, hardware can perform better on large blocks of data than on multiple
small data blocks (fewer copy operations). Hardware offload capabilities such
as TCP Segment Offload (TSO) and Generic Segmentation Offload (GSO) are
popularly used in cases with high data rates or volumes (e.g., datacenters,
hyperscaler servers, etc.) and important to efficiency in compute and power
budgets.  When using TSO and GSO efficiently, there will be large writes
between software and hardware.  Since the hardware itself does not typically
perform pacing, this results in burstiness, or if the sending software is
trying to perform pacing, it could defeat the goal of efficiently using the
offload hardware.  A strategy to work with this is to avoid pacing every single
packet, but instead pace such that a pause is introduced between batches of
some packets that are sent as a mini-burst. Such a strategy is implemented in
Linux, for example.</t>
        <t>At the receiving side, offload techniques like Large Receive Offload (LRO),
Generic Receive Offload (GRO), interrupt coalescing, and other features may
also be impacted by pacing.  Paced packets reduce the ability to group
together incoming packets for upper layer processing, but
end systems may be tuned to handle incoming mini-bursts and maintain some efficiency.</t>
        <t>Clearly, the size of mini-bursts embeds some trade-offs. Even mini-bursts that are very short in terms of time when they leave the sender may cause significant delay further away on an Internet path, where the link capacity is smaller. For example, consider a server that is connected to a 100 Gbps link, serving a client that is behind a 15 Mbps bottleneck link. If that server emits bursts that are 50 kbyte long, the duration of these bursts at the server-side link is negligible (4.1 microseconds). When they reach the bottleneck, however, their duration becomes as large as 27.3 milliseconds. This is an argument for minimizing the size of mini-bursts. On the other hand, wireless link layers such as WiFi or 5G can benefit from having more than one packet available at the local send buffer, to make use of frame aggregation methods. This can significantly reduce overhead, and allow a wireless sender to make better use of its transmission opportunity; eliminating these benefits with pacing may in some cases be counter-productive. This is an argument for making the size of mini-bursts larger.</t>
        <t>Without pacing, increasing cwnd by the number of acknowledged bytes in slow start can cause large mini-bursts when a single ACK acknowledges multiple segments worth of data. This increase is therefore recommended to be upper-bound with one SMSS in <xref target="RFC5681"/>. With pacing, the earlier discussed trade-offs for mini-bursts apply, and such a slow start limitation may not be necessary. Accordingly, HyStart++ <xref target="RFC9406"/>, which CUBIC implementations should use according to <xref target="RFC9406"/>, specifies that no such limit should be applied ("L=infinity").</t>
      </section>
      <section anchor="appcontrol">
        <name>Application control</name>
        <t>When an application produces data at a certain (known) bitrate, it can be beneficial to make use of pacing to
limit the transport bitrate on this basis such that it is not exceedingly large. The Linux, FreeBSD, and Apple OS
applications allow the application to set an upper limit.</t>
        <t>For example, frame based video transmission typically generates data chunks at a varying size at regular intervals.
Such an application could request a data chunk to be transmitted at a lower rate, taking up a larger fraction of
such an interval. This would
allow a more sustained data transmission at a lower rate than a transport protocol's congestion control
might choose, rather than using a shorter time period within the interval with a high rate. This has
the benefit that queue growth is less likely, i.e. this form of pacing can reduce latency.</t>
        <t>Spreading over the interval from one video chunk to the next needs to be done with some caution;
"ideally" spreading data across the
entire interval risks that some of data will not arrive in time, e.g., when delays are introduced
by other traffic. SCReAM and SAMMY pace packets at a somewhat higher rate (50% in case of SCReAM)
to reduce this risk <xref target="I-D.draft-johansson-ccwg-rfc8298bis-screamv2-03"/>, <xref target="Sammy"/>.</t>
      </section>
    </section>
    <section anchor="implementation">
      <name>Implementation examples</name>
      <section anchor="linux">
        <name>Linux TCP</name>
        <t>Pacing was first implemented in Linux kernel version 3.12 in 2013. The following description is based on Linux kernel version 6.12.</t>
        <t>There are two ways to enable pacing in Linux: 1) via a socket option, 2) by configuring the FQ queue discipline. We describe case 1.</t>
        <t>Independent of the value of the Initial Window (IW), the first 10 (hardcoded) packets are not paced. Later, 10 packets will generally be sent without pacing every 2^32 packets.</t>
        <t>Every time an ACK arrives, a pacing rate is calculated, as: factor * MSS * cwnd / SRTT, where "factor" is a configurable value that, by default, is 2 in slow start and 1.2 in congestion avoidance. MSS is the sender maximum segment size <xref target="RFC5681"/>, and SRTT is the smoothed round-trip time <xref target="RFC6298"/>.
The sender transmits data in line with the calculated pacing rate; this is approximated by calculating the rate per millisecond, and generally sending the resulting amount of data per millisecond as a small burst, every millisecond. As an exception, the per-millisecond amount of data can be a little larger when the peer is very close, depending on a configurable value (per default, when the minimum RTT is less than 3 milliseconds).</t>
        <t>If the pacing rate is smaller than 2 packets per millisecond, these bursts will become 2 packets in size, and they will not be sent every millisecond but with a varying time delay (depending on the pacing rate).
If the pacing rate is larger than 64 Kbyte per millisecond, these bursts will be 64 Kbyte in size, and they will not be sent every millisecond but with a varying time delay (depending on the pacing rate).
Bursts can always be smaller than described above, or be "nothing", if a limiting factor such as the receiver window (rwnd) <xref target="RFC5681"/> or the current cwnd disallows transmission.
If the previous packet was not sent when expected by the pacing logic, but more than half of a pacing gap ago (e.g., due to a cwnd limitation), the pacing gap is halved.</t>
        <t>This description is based on the longer Linux pacing analysis text in <xref target="LinuxPacing"/>.</t>
      </section>
      <section anchor="apple-oses">
        <name>Apple OSes</name>
        <t>Starting with iOS 27 and macOS 27, an application can enable pacing on a TCP connection, and set the maximum pacing rate, through two equivalent APIs.</t>
        <t>The Network.framework public API has the following C function:</t>
        <artwork><![CDATA[
int nw_tcp_set_max_pacing_rate(nw_protocol_metadata_t metadata,
                               uint64_t max_pacing_rate);
]]></artwork>
        <t>and a corresponding Swift overlay on <tt>NWProtocolTCP.Metadata</tt>:</t>
        <artwork><![CDATA[
func setMaximumPacingRateBytesPerSecond(_ maximumPacingRateBytesPerSecond: UInt64?)
]]></artwork>
        <t>For applications using the BSD sockets API, the same cap is exposed as the socket-level <tt>SO_MAX_PACING_RATE</tt> option, taking a <tt>uint64_t</tt> rate in bytes per second and supported on <tt>AF_INET</tt> and <tt>AF_INET6</tt> sockets:</t>
        <artwork><![CDATA[
uint64_t rate = 2 * 1024 * 1024;  /* 2 MB/s */
setsockopt(fd, SOL_SOCKET, SO_MAX_PACING_RATE, &rate, sizeof(rate));
]]></artwork>
        <t>The same option may be read back with <tt>getsockopt</tt>.</t>
        <t>In all cases, the value supplied is an upper bound, in bytes per second, on the on-wire rate of a single connection. The actual pacing rate used by the TCP stack is the minimum of (a) this cap and (b) the rate calculated by the transport protocol. The cap can therefore only reduce the on-wire rate, and never raise it above what congestion control allows -- it is exactly the "upper limit" form of application control discussed in <xref target="appcontrol"/>.</t>
        <t>Passing 0 or UINT64_MAX in C, or <tt>nil</tt> in Swift, disables pacing on that connection; the stack then sends as congestion and flow control allow. The cap may be changed at any time during the lifetime of an established connection, and each call replaces the previous value.</t>
        <t>Caps strictly between 0 and 12500 bytes/second (i.e. below 100 Kbps) are silently clamped up to 12500 bytes/second. Applications that need a genuinely lower cap have to shape at the application layer.</t>
        <section anchor="rate-computation-and-packet-scheduling">
          <name>Rate computation and packet scheduling</name>
          <t>On every cwnd update the kernel recomputes a target rate as <tt>cwnd / SRTT</tt> in bytes per second, doubled while the sender is in slow start so that pacing does not throttle the exponential cwnd growth. The application cap, if set, is then applied: the effective pacing rate is the minimum of the computed rate and the cap. The stack then derives a burst budget of roughly 244 µs of data, with a minimum of one MSS.</t>
          <t>Packets are assigned transmit timestamps using a leaky-bucket scheme: consecutive packets share a timestamp while their cumulative size stays within the current burst budget, and once the budget is exhausted the next timestamp is advanced by <tt>budget / rate</tt>. The timestamp is then carried with the packet(s) into the per-interface AQM, which holds the packet in its per-flow queue until the wall clock reaches that timestamp before transmitting it.</t>
        </section>
      </section>
      <section anchor="freebsd">
        <name>FreeBSD</name>
        <t>FreeBSD has the infrastructure to support multiple TCP stacks.
Each TCP stack has a <tt>tcp_output()</tt> function, which handles most of the sending of TCP segments.
The default stack does not support pacing and its <tt>tcp_output()</tt> may be called whenever</t>
        <ol spacing="normal" type="1"><li>
            <t>a TCP segment is received or</t>
          </li>
          <li>
            <t>a TCP timer (like the retransmission or delayed ACK timer) runs off or</t>
          </li>
          <li>
            <t>the application provides new data to be sent</t>
          </li>
        </ol>
        <t>and sends as many TCP segments as is allowed by the congestion and flow control resulting in burst of TCP segments.
However, this also allows to make use of TCP Segment Offload (TSO), which reduces the CPU load.</t>
        <t>The RACK <xref target="RACK"/> and BBR stacks both support pacing by leveraging the TCP High Precision Timer System (HPTS) <xref target="HPTS"/><xref target="HPTSCode"/>, which is a kernel loadable module available in FreeBSD 14 and higher.
The <tt>tcp_output()</tt> function of a TCP stack which supports pacing will not send as much as is allowed by congestion and flow control, but may only send a micro burst and schedule itself for being called after the inter-burst send time using the HPTS.
The RACK stack supports an application setting a pacing rate and a maximum burst size using TCP socket options.
The RACK stack then uses these values to compute the actual micro burst size and the inter-burst send time.</t>
        <t>The following <tt>IPPROTO_TCP</tt>-level socket options are used to control static pacing:</t>
        <table anchor="socket-options-table">
          <name>Socket Options</name>
          <thead>
            <tr>
              <th align="left">Option Name</th>
              <th align="left">Data Type</th>
              <th align="left">Semantic</th>
            </tr>
          </thead>
          <tbody>
            <tr>
              <td align="left">
                <tt>TCP_RACK_PACE_RATE_SS</tt></td>
              <td align="left">
                <tt>uint64_t</tt></td>
              <td align="left">Pace rate in B/s during slow start</td>
            </tr>
            <tr>
              <td align="left">
                <tt>TCP_RACK_PACE_RATE_CA</tt></td>
              <td align="left">
                <tt>uint64_t</tt></td>
              <td align="left">Pace rate in B/s during congestion avoidance</td>
            </tr>
            <tr>
              <td align="left">
                <tt>TCP_RACK_PACE_RATE_REC</tt></td>
              <td align="left">
                <tt>uint64_t</tt></td>
              <td align="left">Pace rate in B/s during recovery</td>
            </tr>
            <tr>
              <td align="left">
                <tt>TCP_RACK_PACE_ALWAYS</tt></td>
              <td align="left">
                <tt>int</tt></td>
              <td align="left">Enable/Disable pacing</td>
            </tr>
            <tr>
              <td align="left">
                <tt>TCP_RACK_PACE_MAX_SEG</tt></td>
              <td align="left">
                <tt>int</tt></td>
              <td align="left">Micro burst size in number of full sized segments</td>
            </tr>
          </tbody>
        </table>
        <t>The first three options can be used to control the pace rate in B/s.
It is possible to specify individual pace rates for slow start, congestion avoidance,
and recovery.
When initializing one of the three pace rates, the other two pace rates are also
initialized to the same rate.
These pacing rates limit the maximum sending rate.
The congestion control and the flow control are always honored.
With the fourth socket option static pacing can be enabled and disabled.
The last socket option allows to control the size of the micro burst in
full sized segments. Its default value is 40.</t>
        <t>The following <tt>packetdrill</tt>-script illustrates the behaviour of a sender
using a pace rate of 1200000 b/s and a micro burst size of 4 full sized segments.
<tt>packetdrill</tt> is available in the FreeBSD ports collection.
Please note that 1200000 b/s correspond to 150000 B/s.
Since FreeBSD takes the size of the IP packet into account this corresponds
to 100 full sized segments per second on a path with an MTU of 1500 bytes.
This means sending a full sized segment every 10 ms or sending a micro burst
of 4 full sized segments every 40 ms.
The script uses a round trip time of 100 ms.
Please note that this behaviour is specific for the RACK stack;
the FreeBSD default stack does not support pacing.</t>
        <artwork><![CDATA[
--ip_version=ipv4

 0.000 `kldload -n tcp_rack`
+0.000 `kldload -n cc_newreno`
+0.000 `sysctl kern.timecounter.alloweddeviation=0`

+0.000 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3
+0.000 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0
+0.000 setsockopt(3, IPPROTO_TCP, TCP_FUNCTION_BLK, {function_set_name="rack",
                                                     pcbcnt=0}, 36) = 0
+0.000 setsockopt(3, IPPROTO_TCP, TCP_CONGESTION, "newreno", 8) = 0
+0.000 bind(3, ..., ...) = 0
+0.000 listen(3, 1) = 0
+0.000 < S      0:0(0)                  win 65535 <mss 1460,sackOK,eol,eol>
+0.000 > S.     0:0(0)        ack     1 win 65535 <mss 1460,sackOK,eol,eol>
+0.100 <  .     1:1(0)        ack     1 win 65535
+0.000 accept(3, ..., ...) = 4
+0.000 close(3) = 0
+0.000 setsockopt(4, IPPROTO_TCP, TCP_LOG, [TCP_LOG_STATE_CONTINUAL], 4) = 0
+0.000 setsockopt(4, IPPROTO_TCP, TCP_RACK_PACE_RATE_SS, [150000], 8) = 0
+0.000 setsockopt(4, IPPROTO_TCP, TCP_RACK_PACE_MAX_SEG, [4], 4) = 0
+0.000 setsockopt(4, IPPROTO_TCP, TCP_RACK_PACE_ALWAYS, [1], 4) = 0
// Provide user data for 10 full sized segments to the TCP stack.
+1.000 send(4, ..., 14600, 0) = 14600
// Send the first micro burst of 4 full sized segments.
+0.000 >  .     1:1461(1460)  ack     1 win 65535
+0.000 >  .  1461:2921(1460)  ack     1 win 65535
+0.000 >  .  2921:4381(1460)  ack     1 win 65535
+0.000 >  .  4381:5841(1460)  ack     1 win 65535
// Send the second micro burst of 4 full sized segments.
+0.040 >  .  5841:7301(1460)  ack     1 win 65535
+0.000 >  .  7301:8761(1460)  ack     1 win 65535
+0.000 >  .  8761:10221(1460) ack     1 win 65535
+0.000 >  . 10221:11681(1460) ack     1 win 65535
// Send the third micro burst of the remaining 2 full sized segments.
+0.040 >  . 11681:13141(1460) ack     1 win 65535
+0.000 > P. 13141:14601(1460) ack     1 win 65535
+0.020 <  .     1:1(0)        ack  2921 win 65535
+0.000 <  .     1:1(0)        ack  5841 win 65535
+0.040 <  .     1:1(0)        ack  8761 win 65535
+0.000 <  .     1:1(0)        ack 11681 win 65535
+0.040 <  .     1:1(0)        ack 14601 win 65535
+0.000 close(4) = 0
+0.000 > F. 14601:14601(0)    ack     1 win 65535
+0.100 < F.     1:1(0)        ack 14602 win 65535
+0.000 >  . 14602:14602(0)    ack     2
]]></artwork>
        <t>The HPTS is optimized for handling a large number of TCP connections and the <tt>tcp_output()</tt> function of the RACK stack is also optimized for being called more often than the <tt>tcp_output()</tt> function of the default stack.
This allows to use TSO in combination with TCP pacing.</t>
        <t>This subsystem underpins recently published research by Netflix and Stanford into application-informed pacing at scale <xref target="Sammy"/>.</t>
      </section>
      <section anchor="quic-bbr-implementations">
        <name>QUIC BBR implementations</name>
        <t>Pacing capability is expected in QUIC senders.  While standard QUIC congestion control <xref target="RFC9002"/> is based on TCP Reno, which is not defined to include pacing (but also does not prohibit it), QUIC congestion control requires either pacing or some other burst limitation (<xref section="7.7" sectionFormat="of" target="RFC9002"/>).  RFC 9002 does not require a specific approach or set of parameters for pacing, leaving endpoints to implement pacing as they choose.  BBR congestion control implementations are common in QUIC stacks, and pacing is integral to BBR, so this document focuses on it, though the same methods are applicable for congestion control based on RFC 9002.</t>
        <t>Pacing in QUIC stacks can rely on one or more of:</t>
        <ol spacing="normal" type="1"><li>
            <t>Access to lower-level (e.g., OS and hardware) capabilities needed for effective pacing.</t>
          </li>
          <li>
            <t>Managing additional timers related to pacing, along with those already needed for retransmission, and other events.</t>
          </li>
          <li>
            <t>Details of the actual pacing algorithm (e.g., granularity of bursts allowed, etc.).</t>
          </li>
        </ol>
        <t>Examples of different approaches to dealing with these challenges in ways that work on multiple operating systems and hardware platforms can be found in open source QUIC stacks, such as Google's QUIC implementation and Meta's "mvfst". These provide examples for some of the concepts discussed below.</t>
        <t>Unlike TCP implementations that typically run within the operating system kernel, QUIC implementations more typically run in user space and are thus faced with more challenges regarding timing and coupling with the underlying protocol stack and hardware needed to achieve pacing.  For instance, if an application trying to do pacing is running on a highly loaded system, it may often "wake up late" and miss the times that it intends to pace packets.  When available, QUIC implementations may choose to rely on operating system features that support pacing for outgoing UDP packets.</t>
        <t>Some operating systems, like Linux <xref target="linux"/>, can support pacing simply through use of an FQ queue discipline applied for outgoing QUIC packets.  However, since multiple QUIC connections can share a socket, this should be carefully applied.  Linux also includes an "SO_TXTIME" option that allows a per-packet transmission time to be specified, however, this may not be compatible with simultaneous use of GSO.  Effectiveness of different OS capabilities for pacing has been evaluated with multiple QUIC stacks <xref target="QuicSteps"/>.</t>
        <t>To support cases where operating systems mechanisms are not available or sufficient alone, QUIC stacks may include their own pacing logic.  When a large amount of data needs to be sent, pacing naively could result in an excessive number of timers to be managed and adjusted along with all of the other timers that the QUIC stack and rest of the application require.  The Hashed Hierarchical Timing Wheel <xref target="VL87"/> provides one approach for such cases, but implementations may also simply schedule the next send event based on the current pacing rate, and then schedule subsequent events as needed, rather than adjusting timers for them.  In any case, typically a pacing algorithm should allow for some amount of burstiness, in order to efficiently use the hardware as well as to be responsive for bursty (but low overall rate) applications, and to avoid excessive timer management.</t>
        <t>Pacing can be done based on different approaches such as a token-based or tokenless algorithm.  For instance, a tokenless algorithm (e.g., as used in mvfst) might compute a regular interval time and batch size (number of packets) to be released every interval and achieve the pacing rate.  This allows specific future transmissions to be scheduled.  In contrast, a token-based algorithm accumulates tokens to permit transmission based on the pacing rate, using a "leaky bucket" to control bursts.  In this case the size of bursts may be more granular, depending on how much time has elapsed between evaluations.</t>
        <t>The additional notion of "burst tokens" (or other burst allowance) may be present in order to rapidly transmit data if coming out of a quiescent period (e.g., when a flow has been application-limited without data to send, e.g., as used in Google's implementation).  A number of burst tokens, representing packets that can be sent unpaced, is initialized to some value (e.g., 10) when a flow starts or becomes quiescent.  If burst tokens are available, outgoing packets are sent immediately, without pacing, up to the limit permitted by the congestion window, and the burst tokens are depleted by each packet sent.  The number of burst tokens is reduced to zero on congestion events.  When coming out of quiescence, it is set to the minimum of the initial burst size, or the amount of packets that the congestion window (in bytes) represents.</t>
        <t>There may be additional "lumpy tokens" that further allow unpaced packets after the burst tokens have been consumed, and the congestion window does not limit sending.  The amount of lumpy tokens that might be present is determined using heuristics, generally limiting to a small number of packets (e.g., 1 or 2).</t>
      </section>
    </section>
    <section anchor="security-considerations">
      <name>Security Considerations</name>
      <t>While congestion control designs, including aspects such as pacing, could result in unwanted competing traffic, they do not directly result in new security considerations.</t>
      <t>Transport protocols that provide authentication (including those using encryption), or are carried over protocols that provide authentication, can protect their congestion control algorithm from network attack. This is orthogonal to the congestion control rules.</t>
    </section>
    <section anchor="iana-considerations">
      <name>IANA Considerations</name>
      <t>This document has no IANA actions.</t>
    </section>
  </middle>
  <back>
    <references anchor="sec-combined-references">
      <name>References</name>
      <references anchor="sec-normative-references">
        <name>Normative References</name>
        <reference anchor="RFC2119">
          <front>
            <title>Key words for use in RFCs to Indicate Requirement Levels</title>
            <author fullname="S. Bradner" initials="S." surname="Bradner"/>
            <date month="March" year="1997"/>
            <abstract>
              <t>In many standards track documents several words are used to signify the requirements in the specification. These words are often capitalized. This document defines these words as they should be interpreted in IETF documents. This document specifies an Internet Best Current Practices for the Internet Community, and requests discussion and suggestions for improvements.</t>
            </abstract>
          </front>
          <seriesInfo name="BCP" value="14"/>
          <seriesInfo name="RFC" value="2119"/>
          <seriesInfo name="DOI" value="10.17487/RFC2119"/>
        </reference>
        <reference anchor="RFC8174">
          <front>
            <title>Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words</title>
            <author fullname="B. Leiba" initials="B." surname="Leiba"/>
            <date month="May" year="2017"/>
            <abstract>
              <t>RFC 2119 specifies common key words that may be used in protocol specifications. This document aims to reduce the ambiguity by clarifying that only UPPERCASE usage of the key words have the defined special meanings.</t>
            </abstract>
          </front>
          <seriesInfo name="BCP" value="14"/>
          <seriesInfo name="RFC" value="8174"/>
          <seriesInfo name="DOI" value="10.17487/RFC8174"/>
        </reference>
      </references>
      <references anchor="sec-informative-references">
        <name>Informative References</name>
        <reference anchor="RFC9002">
          <front>
            <title>QUIC Loss Detection and Congestion Control</title>
            <author fullname="J. Iyengar" initials="J." role="editor" surname="Iyengar"/>
            <author fullname="I. Swett" initials="I." role="editor" surname="Swett"/>
            <date month="May" year="2021"/>
            <abstract>
              <t>This document describes loss detection and congestion control mechanisms for QUIC.</t>
            </abstract>
          </front>
          <seriesInfo name="RFC" value="9002"/>
          <seriesInfo name="DOI" value="10.17487/RFC9002"/>
        </reference>
        <reference anchor="I-D.draft-many-tiptop-usecase">
          <front>
            <title>IP in Deep Space: Key Characteristics, Use Cases and Requirements</title>
            <author fullname="Marc Blanchet" initials="M." surname="Blanchet">
              <organization>Viagenie</organization>
            </author>
            <author fullname="Wesley Eddy" initials="W." surname="Eddy">
              <organization>MTI Systems</organization>
            </author>
            <author fullname="Marshall Eubanks" initials="M." surname="Eubanks">
              <organization>Space Initiatives Inc</organization>
            </author>
            <date day="18" month="June" year="2025"/>
            <abstract>
              <t>   Deep space communications involve long delays (e.g., Earth to Mars
   has one-way delays 4-24 minutes) and intermittent communications,
   mainly because of orbital dynamics.  The IP protocol stack used on
   the Internet is based on the assumptions of shorter delays and mostly
   uninterrupted communications.  This document describes the key
   characteristics, use cases, and requirements for deep space
   networking, intended to help when profiling IP protocols in such
   environment.

              </t>
            </abstract>
          </front>
          <seriesInfo name="Internet-Draft" value="draft-many-tiptop-usecase-03"/>
        </reference>
        <reference anchor="I-D.draft-johansson-ccwg-rfc8298bis-screamv2-03">
          <front>
            <title>Self-Clocked Rate Adaptation for Multimedia</title>
            <author fullname="Ingemar Johansson" initials="I." surname="Johansson">
              <organization>Ericsson</organization>
            </author>
            <author fullname="Magnus Westerlund" initials="M." surname="Westerlund">
              <organization>Ericsson</organization>
            </author>
            <author fullname="Mirja Kühlewind" initials="M." surname="Kühlewind">
              <organization>Ericsson</organization>
            </author>
            <date day="3" month="March" year="2025"/>
            <abstract>
              <t>   This memo describes Self-Clocked Rate Adaptation for Multimedia
   version 2 (SCReAMv2), an update to SCReAM congestion control for
   media streams such as RTP [RFC3550].  SCReAMv2 includes several
   algorithm simplifications and adds support for L4S.  The algorithm
   supports handling of multiple media streams, typical use cases are
   streaming for remote control, AR and 3D VR googles.  This
   specification obsoletes RFC 8298.

              </t>
            </abstract>
          </front>
          <seriesInfo name="Internet-Draft" value="draft-johansson-ccwg-rfc8298bis-screamv2-03"/>
        </reference>
        <reference anchor="VL87">
          <front>
            <title>Hashed and hierarchical timing wheels: data structures for the efficient implementation of a timer facility</title>
            <author initials="G." surname="Varghese">
              <organization/>
            </author>
            <author initials="T." surname="Tauck">
              <organization/>
            </author>
            <date year="1987" month="November" day="01"/>
          </front>
          <seriesInfo name="DOI" value="10.1145/37499.37504"/>
        </reference>
        <reference anchor="Sammy" target="https://doi.org/10.1145/3603269.3604839">
          <front>
            <title>Sammy: Smoothing Video Traffic to be a Friendly Internet Neighbor</title>
            <author initials="B." surname="Spang" fullname="Bruce Spang">
              <organization/>
            </author>
            <author initials="S." surname="Kunamalla" fullname="Shravya Kunamalla">
              <organization/>
            </author>
            <author initials="R." surname="Teixeira" fullname="Renata Teixeira">
              <organization/>
            </author>
            <author initials="T." surname="Huang" fullname="Te-Yuan Huang">
              <organization/>
            </author>
            <author initials="G." surname="Armitage" fullname="Grenville Armitage">
              <organization/>
            </author>
            <author initials="R." surname="Johari" fullname="Ramesh Johari">
              <organization/>
            </author>
            <author initials="N." surname="McKeown" fullname="Nick McKeown">
              <organization/>
            </author>
            <date year="2023" month="September" day="01"/>
          </front>
          <seriesInfo name="ACM SIGCOMM '23: Proceedings of the ACM SIGCOMM 2023 Conference" value=""/>
        </reference>
        <reference anchor="RACK" target="https://freebsdfoundation.org/our-work/journal/browser-based-edition/networking-10th-anniversary/rack-and-alternate-tcp-stacks-for-freebsd/">
          <front>
            <title>RACK and Alternate TCP Stacks for FreeBSD</title>
            <author initials="R." surname="Stewart" fullname="Randall Stewart">
              <organization/>
            </author>
            <author initials="M." surname="Tüxen" fullname="Michael Tüxen">
              <organization/>
            </author>
            <date year="2024" month="February"/>
          </front>
          <seriesInfo name="FreeBSD Journal (January/February 2024)" value=""/>
        </reference>
        <reference anchor="HPTS" target="https://freebsdfoundation.org/wp-content/uploads/2024/11/stewart-adventures.pdf">
          <front>
            <title>Pacing in the FreeBSD TCP Stack</title>
            <author initials="R." surname="Stewart" fullname="Randall Stewart">
              <organization/>
            </author>
            <author initials="M." surname="Tüxen" fullname="Michael Tüxen">
              <organization/>
            </author>
            <date year="2024" month="October"/>
          </front>
          <seriesInfo name="FreeBSD Journal" value=""/>
        </reference>
        <reference anchor="HPTSCode" target="https://github.com/freebsd/freebsd-src/blob/main/sys/netinet/tcp_hpts.c#L31-L99">
          <front>
            <title>tcp_hpts.c</title>
            <author>
              <organization/>
            </author>
            <date year="2024" month="October"/>
          </front>
        </reference>
        <reference anchor="LinuxPacing" target="https://folk.universitetetioslo.no/michawe/research/publications/icnc2025-pacing.pdf">
          <front>
            <title>TCP Pacing in the Linux Kernel</title>
            <author initials="M." surname="Welzl" fullname="Michael Welzl">
              <organization/>
            </author>
            <date year="2025" month="February" day="17"/>
          </front>
          <seriesInfo name="IEEE ICNC 2025" value=""/>
        </reference>
        <reference anchor="UnderstandingPacing" target="https://doi.org/10.1109/INFCOM.2000.832483">
          <front>
            <title>Understanding the performance of TCP pacing</title>
            <author initials="A." surname="Aggarwal" fullname="Amit Aggarwal">
              <organization/>
            </author>
            <author initials="S." surname="Savage" fullname="Stefan Savage">
              <organization/>
            </author>
            <author initials="T." surname="Anderson" fullname="Thomas Anderson">
              <organization/>
            </author>
            <date year="2000" month="March"/>
          </front>
          <seriesInfo name="IEEE Infocom 2000" value=""/>
        </reference>
        <reference anchor="QuicSteps" target="https://doi.org/10.1145/3730985">
          <front>
            <title>QUIC Steps: Evaluating Pacing Strategies in QUIC Implementations</title>
            <author initials="M." surname="Kempf" fullname="Marcel Kempf">
              <organization/>
            </author>
            <author initials="S." surname="Tietz" fullname="Simon Tietz">
              <organization/>
            </author>
            <author initials="B." surname="Jaeger" fullname="Benedikt Jaeger">
              <organization/>
            </author>
            <author initials="J." surname="Spath" fullname="Johannes Spath">
              <organization/>
            </author>
            <author initials="G." surname="Carle" fullname="Georg Carle">
              <organization/>
            </author>
            <author initials="J." surname="Zirngibl" fullname="Johnnes Zirngibl">
              <organization/>
            </author>
            <date year="2025" month="June"/>
          </front>
          <seriesInfo name="Proceedings of the ACM on Networking" value=""/>
        </reference>
        <reference anchor="I-D.ietf-ccwg-bbr">
          <front>
            <title>BBR Congestion Control</title>
            <author fullname="Neal Cardwell" initials="N." surname="Cardwell">
              <organization>Google</organization>
            </author>
            <author fullname="Ian Swett" initials="I." surname="Swett">
              <organization>Google</organization>
            </author>
            <author fullname="Joseph Beshay" initials="J." surname="Beshay">
              <organization>Meta</organization>
            </author>
            <date day="2" month="November" year="2026"/>
            <abstract>
              <t>   This document specifies the BBR congestion control algorithm.  BBR
   ("Bottleneck Bandwidth and Round-trip propagation time") uses recent
   measurements of a transport connection's delivery rate, round-trip
   time, and packet loss rate to build an explicit model of the network
   path.  BBR then uses this model to control both how fast it sends
   data and the maximum volume of data it allows in flight in the
   network at any time.  Relative to loss-based congestion control
   algorithms such as Reno [RFC5681] or CUBIC [RFC9438], BBR offers
   substantially higher throughput for bottlenecks with shallow buffers
   or random losses, and substantially lower queueing delays for
   bottlenecks with deep buffers (avoiding "bufferbloat").  BBR can be
   implemented in any transport protocol that supports packet-delivery
   acknowledgment.  Thus far, open source implementations are available
   for TCP [RFC9293] and QUIC [RFC9000].  This document specifies
   version 3 of the BBR algorithm, BBRv3.

              </t>
            </abstract>
          </front>
          <seriesInfo name="Internet-Draft" value="draft-ietf-ccwg-bbr-07"/>
        </reference>
        <reference anchor="RFC7567">
          <front>
            <title>IETF Recommendations Regarding Active Queue Management</title>
            <author fullname="F. Baker" initials="F." role="editor" surname="Baker"/>
            <author fullname="G. Fairhurst" initials="G." role="editor" surname="Fairhurst"/>
            <date month="July" year="2015"/>
            <abstract>
              <t>This memo presents recommendations to the Internet community concerning measures to improve and preserve Internet performance. It presents a strong recommendation for testing, standardization, and widespread deployment of active queue management (AQM) in network devices to improve the performance of today's Internet. It also urges a concerted effort of research, measurement, and ultimate deployment of AQM mechanisms to protect the Internet from flows that are not sufficiently responsive to congestion notification.</t>
              <t>Based on 15 years of experience and new research, this document replaces the recommendations of RFC 2309.</t>
            </abstract>
          </front>
          <seriesInfo name="BCP" value="197"/>
          <seriesInfo name="RFC" value="7567"/>
          <seriesInfo name="DOI" value="10.17487/RFC7567"/>
        </reference>
        <reference anchor="RFC8087">
          <front>
            <title>The Benefits of Using Explicit Congestion Notification (ECN)</title>
            <author fullname="G. Fairhurst" initials="G." surname="Fairhurst"/>
            <author fullname="M. Welzl" initials="M." surname="Welzl"/>
            <date month="March" year="2017"/>
            <abstract>
              <t>The goal of this document is to describe the potential benefits of applications using a transport that enables Explicit Congestion Notification (ECN). The document outlines the principal gains in terms of increased throughput, reduced delay, and other benefits when ECN is used over a network path that includes equipment that supports Congestion Experienced (CE) marking. It also discusses challenges for successful deployment of ECN. It does not propose new algorithms to use ECN nor does it describe the details of implementation of ECN in endpoint devices (Internet hosts), routers, or other network devices.</t>
            </abstract>
          </front>
          <seriesInfo name="RFC" value="8087"/>
          <seriesInfo name="DOI" value="10.17487/RFC8087"/>
        </reference>
        <reference anchor="RFC9040">
          <front>
            <title>TCP Control Block Interdependence</title>
            <author fullname="J. Touch" initials="J." surname="Touch"/>
            <author fullname="M. Welzl" initials="M." surname="Welzl"/>
            <author fullname="S. Islam" initials="S." surname="Islam"/>
            <date month="July" year="2021"/>
            <abstract>
              <t>This memo provides guidance to TCP implementers that is intended to help improve connection convergence to steady-state operation without affecting interoperability. It updates and replaces RFC 2140's description of sharing TCP state, as typically represented in TCP Control Blocks, among similar concurrent or consecutive connections.</t>
            </abstract>
          </front>
          <seriesInfo name="RFC" value="9040"/>
          <seriesInfo name="DOI" value="10.17487/RFC9040"/>
        </reference>
        <reference anchor="RFC9000">
          <front>
            <title>QUIC: A UDP-Based Multiplexed and Secure Transport</title>
            <author fullname="J. Iyengar" initials="J." role="editor" surname="Iyengar"/>
            <author fullname="M. Thomson" initials="M." role="editor" surname="Thomson"/>
            <date month="May" year="2021"/>
            <abstract>
              <t>This document defines the core of the QUIC transport protocol. QUIC provides applications with flow-controlled streams for structured communication, low-latency connection establishment, and network path migration. QUIC includes security measures that ensure confidentiality, integrity, and availability in a range of deployment circumstances. Accompanying documents describe the integration of TLS for key negotiation, loss detection, and an exemplary congestion control algorithm.</t>
            </abstract>
          </front>
          <seriesInfo name="RFC" value="9000"/>
          <seriesInfo name="DOI" value="10.17487/RFC9000"/>
        </reference>
        <reference anchor="RFC5681">
          <front>
            <title>TCP Congestion Control</title>
            <author fullname="M. Allman" initials="M." surname="Allman"/>
            <author fullname="V. Paxson" initials="V." surname="Paxson"/>
            <author fullname="E. Blanton" initials="E." surname="Blanton"/>
            <date month="September" year="2009"/>
            <abstract>
              <t>This document defines TCP's four intertwined congestion control algorithms: slow start, congestion avoidance, fast retransmit, and fast recovery. In addition, the document specifies how TCP should begin transmission after a relatively long idle period, as well as discussing various acknowledgment generation methods. This document obsoletes RFC 2581. [STANDARDS-TRACK]</t>
            </abstract>
          </front>
          <seriesInfo name="RFC" value="5681"/>
          <seriesInfo name="DOI" value="10.17487/RFC5681"/>
        </reference>
        <reference anchor="RFC9406">
          <front>
            <title>HyStart++: Modified Slow Start for TCP</title>
            <author fullname="P. Balasubramanian" initials="P." surname="Balasubramanian"/>
            <author fullname="Y. Huang" initials="Y." surname="Huang"/>
            <author fullname="M. Olson" initials="M." surname="Olson"/>
            <date month="May" year="2023"/>
            <abstract>
              <t>This document describes HyStart++, a simple modification to the slow start phase of congestion control algorithms. Slow start can overshoot the ideal send rate in many cases, causing high packet loss and poor performance. HyStart++ uses increase in round-trip delay as a heuristic to find an exit point before possible overshoot. It also adds a mitigation to prevent jitter from causing premature slow start exit.</t>
            </abstract>
          </front>
          <seriesInfo name="RFC" value="9406"/>
          <seriesInfo name="DOI" value="10.17487/RFC9406"/>
        </reference>
        <reference anchor="RFC6298">
          <front>
            <title>Computing TCP's Retransmission Timer</title>
            <author fullname="V. Paxson" initials="V." surname="Paxson"/>
            <author fullname="M. Allman" initials="M." surname="Allman"/>
            <author fullname="J. Chu" initials="J." surname="Chu"/>
            <author fullname="M. Sargent" initials="M." surname="Sargent"/>
            <date month="June" year="2011"/>
            <abstract>
              <t>This document defines the standard algorithm that Transmission Control Protocol (TCP) senders are required to use to compute and manage their retransmission timer. It expands on the discussion in Section 4.2.3.1 of RFC 1122 and upgrades the requirement of supporting the algorithm from a SHOULD to a MUST. This document obsoletes RFC 2988. [STANDARDS-TRACK]</t>
            </abstract>
          </front>
          <seriesInfo name="RFC" value="6298"/>
          <seriesInfo name="DOI" value="10.17487/RFC6298"/>
        </reference>
      </references>
    </references>
    <?line 597?>

<!-- # Acknowledgments
{:numbered="false"}

TODO acknowledge. Note, numbered sections shouldn't appear
after an unnumbered one - so either move this last, or take
the numbering rule out. -->

<section numbered="false" anchor="acknowledgments">
      <name>Acknowledgments</name>
      <t>The authors would like to thank Grenville Armitage, Gorry Fairhurst, Ingemar Johansson, Nicolas Kuhn and Eduard Vasilenko for suggesting improvements to this document.</t>
    </section>
    <section numbered="false" anchor="change-log">
      <name>Change Log</name>
      <ul spacing="normal">
        <li>
          <t>-00 was the first individual submission for feedback by ICCRG.</t>
        </li>
        <li>
          <t>-01 adds Michael Tuexen as new co-author, and adds the following new descriptions:
          </t>
          <ul spacing="normal">
            <li>
              <t>a first version of text for Apple OSes</t>
            </li>
            <li>
              <t>a first version of text for FreeBSD</t>
            </li>
          </ul>
        </li>
        <li>
          <t>-02 adds a reference for Linux pacing, removes a comment on Linux SRTT calculation ("TODO check": it was checked, nothing to change), adds more discussion text:
          </t>
          <ul spacing="normal">
            <li>
              <t>queue dynamics / AQM interactions</t>
            </li>
            <li>
              <t>initial RTT calculation, thanks to Ingemar Johansson</t>
            </li>
            <li>
              <t>mini-bursts and their trade-offs, thanks to Eduard Vasilenko</t>
            </li>
            <li>
              <t>... and now we also have an ACK section for such thanks</t>
            </li>
          </ul>
        </li>
        <li>
          <t>-03:
          </t>
          <ul spacing="normal">
            <li>
              <t>re-structures the parts on general considerations and adds more text in them</t>
            </li>
            <li>
              <t>elaborates on application control</t>
            </li>
          </ul>
        </li>
        <li>
          <t>Adopted version draft-irtf-iccrg-00:
          </t>
          <ul spacing="normal">
            <li>
              <t>adds a reference to the "Understanding the performance of TCP pacing" paper and discusses it in the "Motivations for pacing" section</t>
            </li>
            <li>
              <t>adds an example for the double-loss-after-slow start problem in section 4.1</t>
            </li>
            <li>
              <t>extends the FreeBSD text</t>
            </li>
          </ul>
        </li>
        <li>
          <t>-01:
          </t>
          <ul spacing="normal">
            <li>
              <t>improves the FreeBSD text</t>
            </li>
            <li>
              <t>adds a discussion of the ABC limit in slow start</t>
            </li>
          </ul>
        </li>
        <li>
          <t>-02:
          </t>
          <ul spacing="normal">
            <li>
              <t>Extended the Apple OSes description</t>
            </li>
          </ul>
        </li>
        <li>
          <t>-03:
          </t>
          <ul spacing="normal">
            <li>
              <t>fixing nits, and a bit of text on how DelACKs can harm RTT estimation</t>
            </li>
          </ul>
        </li>
        <li>
          <t>-04:
          </t>
          <ul spacing="normal">
            <li>
              <t>addressing comment from Gorry Fairhurst: primarily editorial, some details in the QUIC section.</t>
            </li>
          </ul>
        </li>
        <li>
          <t>-05:
          </t>
          <ul spacing="normal">
            <li>
              <t>adds RFC 5743 info to abstract and intro. Refer to HTPSCode.</t>
            </li>
          </ul>
        </li>
      </ul>
    </section>
  </back>
  <!-- ##markdown-source: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-->

</rfc>
