]> University of Oslo

PO Box 1080 Blindern 0316 Oslo Norway michawe@ifi.uio.no http://welzl.at/

MTI Systems

25111 Country Club Blvd, Suite 295 North Olmsted, OH 44070 United States of America wes@mti-systems.com

Apple Inc.

One Apple Park Way Cupertino, California 95014 United States of America vidhi_goel@apple.com

Münster University of Applied Sciences

Stegerwaldstrasse 39 48565 Steinfurt Germany tuexen@fh-muenster.de

IRTF Internet Congestion Control next generation unicorn sparkling distributed ledger 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. About This Document The latest revision of this draft can be found at . Status information for this document may be found at . Discussion of this document takes place on the Internet Congestion Control Research Group mailing list (), which is archived at . Subscribe at . Source for this draft and an issue tracker can be found at .

Introduction 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. 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 to send the same cwnd sized data but spread it across a round-trip time more evenly. 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 being back-to-back), loss can be more likely to happen than in a case where there are time gaps between packets (e.g., when they are spread out 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. 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 .

Conventions and Definitions The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 when, and only when, they appear in all capitals, as shown here.

Pacing: general considerations and consequences

More likely to saturate a bottleneck We can distinguish between two reasons for packet losses that are due to congestion at a bottleneck with a DropTail (FIFO) queue:

1. 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.
2. The bottleneck is fully saturated. The queue is full, and packets drain from it more slowly than new packets arrive.

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 to happen 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. 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. 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. 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. 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).

Backing off after the increase The two loss types explained in 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 >= 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 > 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. Since pacing makes loss type 2 more likely, beta < 0.5 may be a better choice after slow start overshoot when pacing is used.

Able to work with smaller queues The probability of loss type 1 in 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.

Queue dynamics 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 . Paced traffic, on the other hand, can cause a bottleneck queue to grow more slowly and steadily, incuring 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 to drop packets or mark them using Explicit Congestion Notification (ECN). 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.

Getting good RTT estimates 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. Thus, using this sample to pace the initial window can cause the pacing rate to become unnecessarily low. This may be the reason why the Linux TCP implementation does not pace the first 10 packets (see ). 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) .

Mini-bursts and their trade-offs Generally, hardware can perform better on large blocks of data than on multiple small data blocks (fewer copy operations). This plays a role for mechanisms such as TCP Segment Offload (TSO), which, however, only matter at high data rates. 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. 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 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.

Implementation examples

Linux TCP The following description is based on Linux kernel version 6.8.9. 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. 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. 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 , and SRTT is the smoothed round-trip time . 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). 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) 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. This description is based on the longer Linux pacing analysis text in .

Apple OSes Pacing was added to Apple OS as a private API in iOS 17 and macOS 14. In its current form, an application or transport protocol computes and sets the desired transmit timestamp on a per packet basis and sends it to the pacing module in AQM. The packets are queued in the AQM until the current time becomes greater than or equal to corresponding packet's transmit timestamp. There is an upper limit of 3 seconds for how long the AQM will hold a queued packet before sending it. The above simplicity in the kernel allows upper layer protocols or applications to set a transmit timestamp in a manner that is suitable for them. For example, a stream based protocol like TCP might pace packets differently than a video conferencing app.

FreeBSD FreeBSD has the infrastructure to support multiple TCP stacks. Each TCP stack has a tcp_output() function, which handles most of the sending of TCP segments. The default stack does not support pacing and its tcp_output() may be called whenever

1. a TCP segment is received or
2. a TCP timer (like the retransmission or delayed ACK timer) runs off or
3. the application provides new data to be sent

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. The RACK and BBR stacks both support pacing by leveraging the TCP High Precision Timer System (HPTS) , which is a kernel loadable module available in FreeBSD 14 and higher. The tcp_output() 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. The HPTS is optimized for handling a large number of TCP connections and the tcp_output() function of the RACK stack is also optimized for being called more often than the tcp_output() function of the default stack. This allows to use TSO in combination with TCP pacing. This subsystem underpins recently published research by Netflix and Stanford into application-informed pacing at scale .

QUIC BBR implementations Pacing capability is expected in QUIC senders. While standard QUIC congestion control 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 (). BBR congestion control implementations are common in QUIC stacks, and pacing is integral to BBR, so this document focuses on it. Pacing in QUIC stacks commonly involves:

1. Access to lower-level (e.g. OS and hardware) capabilities needed for effective pacing.
2. Managing additional timers related to pacing, along with those already needed for retransmission, and other events.
3. Details of the actual pacing algorithm (e.g. granularity of bursts allowed, etc.).

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. 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 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 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. 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. 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. 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).

Security Considerations While congestion control designs, including aspects such as pacing, could result in unwanted competing traffic, they do not directly result in new security considerations. 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.

IANA Considerations This document has no IANA actions.

References Normative References Google Google Google Google Google 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 [RFC8312], 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 [RFC793] and QUIC [RFC9000]. This document specifies version 2 of the BBR algorithm, also sometimes referred to as BBRv2 or bbr2. 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. 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. Informative References This document describes loss detection and congestion control mechanisms for QUIC. 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. 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] 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]

Acknowledgments The authors would like to thank Grenville Armitage, Ingemar Johannson and Eduard Vasilenko for suggesting improvements to this document.

Change Log

• -00 was the first individual submission for feedback by ICCRG.
• -01 adds Michael Tuexen as new co-author, and adds the following new descriptions:
  • a first version of text for Apple OSes
  • a first version of text for FreeBSD
• -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:
  • queue dynamics / AQM interactions
  • initial RTT calculation, thanks to Ingemar Johannson
  • mini-bursts and their trade-offs, thanks to Eduard Vasilenko
  • ... and now we also have an ACK section for such thanks