RFC 9002 - QUIC Loss Detection and Congestion Control

6. Loss Detection

QUIC senders use acknowledgments to detect lost packets and a PTO to ensure acknowledgments are received; see Section 6.2. This section provides a description of these algorithms.¶

If a packet is lost, the QUIC transport needs to recover from that loss, such as by retransmitting the data, sending an updated frame, or discarding the frame. For more information, see Section 13.3 of [QUIC-TRANSPORT].¶

Loss detection is separate per packet number space, unlike RTT measurement and congestion control, because RTT and congestion control are properties of the path, whereas loss detection also relies upon key availability.¶

6.1. Acknowledgment-Based Detection

Acknowledgment-based loss detection implements the spirit of TCP's Fast Retransmit [RFC5681], Early Retransmit [RFC5827], Forward Acknowledgment [FACK], SACK loss recovery [RFC6675], and RACK-TLP [RFC8985]. This section provides an overview of how these algorithms are implemented in QUIC.¶

A packet is declared lost if it meets all of the following conditions:¶

The packet is unacknowledged, in flight, and was sent prior to an acknowledged packet.
The packet was sent kPacketThreshold packets before an acknowledged packet (Section 6.1.1), or it was sent long enough in the past (Section 6.1.2).

The acknowledgment indicates that a packet sent later was delivered, and the packet and time thresholds provide some tolerance for packet reordering.¶

Spuriously declaring packets as lost leads to unnecessary retransmissions and may result in degraded performance due to the actions of the congestion controller upon detecting loss. Implementations can detect spurious retransmissions and increase the packet or time reordering threshold to reduce future spurious retransmissions and loss events. Implementations with adaptive time thresholds MAY choose to start with smaller initial reordering thresholds to minimize recovery latency.¶

6.1.1. Packet Threshold

The RECOMMENDED initial value for the packet reordering threshold (kPacketThreshold) is 3, based on best practices for TCP loss detection [RFC5681] [RFC6675]. In order to remain similar to TCP, implementations SHOULD NOT use a packet threshold less than 3; see [RFC5681].¶

Some networks may exhibit higher degrees of packet reordering, causing a sender to detect spurious losses. Additionally, packet reordering could be more common with QUIC than TCP because network elements that could observe and reorder TCP packets cannot do that for QUIC and also because QUIC packet numbers are encrypted. Algorithms that increase the reordering threshold after spuriously detecting losses, such as RACK [RFC8985], have proven to be useful in TCP and are expected to be at least as useful in QUIC.¶

6.1.2. Time Threshold

Once a later packet within the same packet number space has been acknowledged, an endpoint SHOULD declare an earlier packet lost if it was sent a threshold amount of time in the past. To avoid declaring packets as lost too early, this time threshold MUST be set to at least the local timer granularity, as indicated by the kGranularity constant. The time threshold is:¶

max(kTimeThreshold * max(smoothed_rtt, latest_rtt), kGranularity)

If packets sent prior to the largest acknowledged packet cannot yet be declared lost, then a timer SHOULD be set for the remaining time.¶

Using max(smoothed_rtt, latest_rtt) protects from the two following cases:¶

the latest RTT sample is lower than the smoothed RTT, perhaps due to reordering where the acknowledgment encountered a shorter path;
the latest RTT sample is higher than the smoothed RTT, perhaps due to a sustained increase in the actual RTT, but the smoothed RTT has not yet caught up.

The RECOMMENDED time threshold (kTimeThreshold), expressed as an RTT multiplier, is 9/8. The RECOMMENDED value of the timer granularity (kGranularity) is 1 millisecond.¶

Implementations MAY experiment with absolute thresholds, thresholds from previous connections, adaptive thresholds, or the including of RTT variation. Smaller thresholds reduce reordering resilience and increase spurious retransmissions, and larger thresholds increase loss detection delay.¶

6.2. Probe Timeout

A Probe Timeout (PTO) triggers the sending of one or two probe datagrams when ack-eliciting packets are not acknowledged within the expected period of time or the server may not have validated the client's address. A PTO enables a connection to recover from loss of tail packets or acknowledgments.¶

As with loss detection, the PTO is per packet number space. That is, a PTO value is computed per packet number space.¶

A PTO timer expiration event does not indicate packet loss and MUST NOT cause prior unacknowledged packets to be marked as lost. When an acknowledgment is received that newly acknowledges packets, loss detection proceeds as dictated by the packet and time threshold mechanisms; see Section 6.1.¶

The PTO algorithm used in QUIC implements the reliability functions of Tail Loss Probe [RFC8985], RTO [RFC5681], and F-RTO algorithms for TCP [RFC5682]. The timeout computation is based on TCP's RTO period [RFC6298].¶

6.2.1. Computing PTO

When an ack-eliciting packet is transmitted, the sender schedules a timer for the PTO period as follows:¶

PTO = smoothed_rtt + max(4*rttvar, kGranularity) + max_ack_delay

The PTO period is the amount of time that a sender ought to wait for an acknowledgment of a sent packet. This time period includes the estimated network RTT (smoothed_rtt), the variation in the estimate (4*rttvar), and max_ack_delay, to account for the maximum time by which a receiver might delay sending an acknowledgment.¶

When the PTO is armed for Initial or Handshake packet number spaces, the max_ack_delay in the PTO period computation is set to 0, since the peer is expected to not delay these packets intentionally; see Section 13.2.1 of [QUIC-TRANSPORT].¶

The PTO period MUST be at least kGranularity to avoid the timer expiring immediately.¶

When ack-eliciting packets in multiple packet number spaces are in flight, the timer MUST be set to the earlier value of the Initial and Handshake packet number spaces.¶

An endpoint MUST NOT set its PTO timer for the Application Data packet number space until the handshake is confirmed. Doing so prevents the endpoint from retransmitting information in packets when either the peer does not yet have the keys to process them or the endpoint does not yet have the keys to process their acknowledgments. For example, this can happen when a client sends 0-RTT packets to the server; it does so without knowing whether the server will be able to decrypt them. Similarly, this can happen when a server sends 1-RTT packets before confirming that the client has verified the server's certificate and can therefore read these 1-RTT packets.¶

A sender SHOULD restart its PTO timer every time an ack-eliciting packet is sent or acknowledged, or when Initial or Handshake keys are discarded (Section 4.9 of [QUIC-TLS]). This ensures the PTO is always set based on the latest estimate of the RTT and for the correct packet across packet number spaces.¶

When a PTO timer expires, the PTO backoff MUST be increased, resulting in the PTO period being set to twice its current value. The PTO backoff factor is reset when an acknowledgment is received, except in the following case. A server might take longer to respond to packets during the handshake than otherwise. To protect such a server from repeated client probes, the PTO backoff is not reset at a client that is not yet certain that the server has finished validating the client's address. That is, a client does not reset the PTO backoff factor on receiving acknowledgments in Initial packets.¶

This exponential reduction in the sender's rate is important because consecutive PTOs might be caused by loss of packets or acknowledgments due to severe congestion. Even when there are ack-eliciting packets in flight in multiple packet number spaces, the exponential increase in PTO occurs across all spaces to prevent excess load on the network. For example, a timeout in the Initial packet number space doubles the length of the timeout in the Handshake packet number space.¶

The total length of time over which consecutive PTOs expire is limited by the idle timeout.¶

The PTO timer MUST NOT be set if a timer is set for time threshold loss detection; see Section 6.1.2. A timer that is set for time threshold loss detection will expire earlier than the PTO timer in most cases and is less likely to spuriously retransmit data.¶

6.2.2. Handshakes and New Paths

Resumed connections over the same network MAY use the previous connection's final smoothed RTT value as the resumed connection's initial RTT. When no previous RTT is available, the initial RTT SHOULD be set to 333 milliseconds. This results in handshakes starting with a PTO of 1 second, as recommended for TCP's initial RTO; see Section 2 of [RFC6298].¶

A connection MAY use the delay between sending a PATH_CHALLENGE and receiving a PATH_RESPONSE to set the initial RTT (see kInitialRtt in Appendix A.2) for a new path, but the delay SHOULD NOT be considered an RTT sample.¶

When the Initial keys and Handshake keys are discarded (see Section 6.4), any Initial packets and Handshake packets can no longer be acknowledged, so they are removed from bytes in flight. When Initial or Handshake keys are discarded, the PTO and loss detection timers MUST be reset, because discarding keys indicates forward progress and the loss detection timer might have been set for a now-discarded packet number space.¶

6.2.2.1. Before Address Validation

Until the server has validated the client's address on the path, the amount of data it can send is limited to three times the amount of data received, as specified in Section 8.1 of [QUIC-TRANSPORT]. If no additional data can be sent, the server's PTO timer MUST NOT be armed until datagrams have been received from the client because packets sent on PTO count against the anti-amplification limit.¶

When the server receives a datagram from the client, the amplification limit is increased and the server resets the PTO timer. If the PTO timer is then set to a time in the past, it is executed immediately. Doing so avoids sending new 1-RTT packets prior to packets critical to the completion of the handshake. In particular, this can happen when 0-RTT is accepted but the server fails to validate the client's address.¶

Since the server could be blocked until more datagrams are received from the client, it is the client's responsibility to send packets to unblock the server until it is certain that the server has finished its address validation (see Section 8 of [QUIC-TRANSPORT]). That is, the client MUST set the PTO timer if the client has not received an acknowledgment for any of its Handshake packets and the handshake is not confirmed (see Section 4.1.2 of [QUIC-TLS]), even if there are no packets in flight. When the PTO fires, the client MUST send a Handshake packet if it has Handshake keys, otherwise it MUST send an Initial packet in a UDP datagram with a payload of at least 1200 bytes.¶

6.2.3. Speeding up Handshake Completion

When a server receives an Initial packet containing duplicate CRYPTO data, it can assume the client did not receive all of the server's CRYPTO data sent in Initial packets, or the client's estimated RTT is too small. When a client receives Handshake or 1-RTT packets prior to obtaining Handshake keys, it may assume some or all of the server's Initial packets were lost.¶

To speed up handshake completion under these conditions, an endpoint MAY, for a limited number of times per connection, send a packet containing unacknowledged CRYPTO data earlier than the PTO expiry, subject to the address validation limits in Section 8.1 of [QUIC-TRANSPORT]. Doing so at most once for each connection is adequate to quickly recover from a single packet loss. An endpoint that always retransmits packets in response to receiving packets that it cannot process risks creating an infinite exchange of packets.¶

Endpoints can also use coalesced packets (see Section 12.2 of [QUIC-TRANSPORT]) to ensure that each datagram elicits at least one acknowledgment. For example, a client can coalesce an Initial packet containing PING and PADDING frames with a 0-RTT data packet, and a server can coalesce an Initial packet containing a PING frame with one or more packets in its first flight.¶

6.2.4. Sending Probe Packets

When a PTO timer expires, a sender MUST send at least one ack-eliciting packet in the packet number space as a probe. An endpoint MAY send up to two full-sized datagrams containing ack-eliciting packets to avoid an expensive consecutive PTO expiration due to a single lost datagram or to transmit data from multiple packet number spaces. All probe packets sent on a PTO MUST be ack-eliciting.¶

In addition to sending data in the packet number space for which the timer expired, the sender SHOULD send ack-eliciting packets from other packet number spaces with in-flight data, coalescing packets if possible. This is particularly valuable when the server has both Initial and Handshake data in flight or when the client has both Handshake and Application Data in flight because the peer might only have receive keys for one of the two packet number spaces.¶

If the sender wants to elicit a faster acknowledgment on PTO, it can skip a packet number to eliminate the acknowledgment delay.¶

An endpoint SHOULD include new data in packets that are sent on PTO expiration. Previously sent data MAY be sent if no new data can be sent. Implementations MAY use alternative strategies for determining the content of probe packets, including sending new or retransmitted data based on the application's priorities.¶

It is possible the sender has no new or previously sent data to send. As an example, consider the following sequence of events: new application data is sent in a STREAM frame, deemed lost, then retransmitted in a new packet, and then the original transmission is acknowledged. When there is no data to send, the sender SHOULD send a PING or other ack-eliciting frame in a single packet, rearming the PTO timer.¶

Alternatively, instead of sending an ack-eliciting packet, the sender MAY mark any packets still in flight as lost. Doing so avoids sending an additional packet but increases the risk that loss is declared too aggressively, resulting in an unnecessary rate reduction by the congestion controller.¶

Consecutive PTO periods increase exponentially, and as a result, connection recovery latency increases exponentially as packets continue to be dropped in the network. Sending two packets on PTO expiration increases resilience to packet drops, thus reducing the probability of consecutive PTO events.¶

When the PTO timer expires multiple times and new data cannot be sent, implementations must choose between sending the same payload every time or sending different payloads. Sending the same payload may be simpler and ensures the highest priority frames arrive first. Sending different payloads each time reduces the chances of spurious retransmission.¶

6.3. Handling Retry Packets

A Retry packet causes a client to send another Initial packet, effectively restarting the connection process. A Retry packet indicates that the Initial packet was received but not processed. A Retry packet cannot be treated as an acknowledgment because it does not indicate that a packet was processed or specify the packet number.¶

Clients that receive a Retry packet reset congestion control and loss recovery state, including resetting any pending timers. Other connection state, in particular cryptographic handshake messages, is retained; see Section 17.2.5 of [QUIC-TRANSPORT].¶

The client MAY compute an RTT estimate to the server as the time period from when the first Initial packet was sent to when a Retry or a Version Negotiation packet is received. The client MAY use this value in place of its default for the initial RTT estimate.¶

6.4. Discarding Keys and Packet State

When Initial and Handshake packet protection keys are discarded (see Section 4.9 of [QUIC-TLS]), all packets that were sent with those keys can no longer be acknowledged because their acknowledgments cannot be processed. The sender MUST discard all recovery state associated with those packets and MUST remove them from the count of bytes in flight.¶

Endpoints stop sending and receiving Initial packets once they start exchanging Handshake packets; see Section 17.2.2.1 of [QUIC-TRANSPORT]. At this point, recovery state for all in-flight Initial packets is discarded.¶

When 0-RTT is rejected, recovery state for all in-flight 0-RTT packets is discarded.¶

If a server accepts 0-RTT, but does not buffer 0-RTT packets that arrive before Initial packets, early 0-RTT packets will be declared lost, but that is expected to be infrequent.¶

It is expected that keys are discarded at some time after the packets encrypted with them are either acknowledged or declared lost. However, Initial and Handshake secrets are discarded as soon as Handshake and 1-RTT keys are proven to be available to both client and server; see Section 4.9.1 of [QUIC-TLS].¶

7. Congestion Control

This document specifies a sender-side congestion controller for QUIC similar to TCP NewReno [RFC6582].¶

The signals QUIC provides for congestion control are generic and are designed to support different sender-side algorithms. A sender can unilaterally choose a different algorithm to use, such as CUBIC [RFC8312].¶

If a sender uses a different controller than that specified in this document, the chosen controller MUST conform to the congestion control guidelines specified in Section 3.1 of [RFC8085].¶

Similar to TCP, packets containing only ACK frames do not count toward bytes in flight and are not congestion controlled. Unlike TCP, QUIC can detect the loss of these packets and MAY use that information to adjust the congestion controller or the rate of ACK-only packets being sent, but this document does not describe a mechanism for doing so.¶

The congestion controller is per path, so packets sent on other paths do not alter the current path's congestion controller, as described in Section 9.4 of [QUIC-TRANSPORT].¶

The algorithm in this document specifies and uses the controller's congestion window in bytes.¶

An endpoint MUST NOT send a packet if it would cause bytes_in_flight (see Appendix B.2) to be larger than the congestion window, unless the packet is sent on a PTO timer expiration (see Section 6.2) or when entering recovery (see Section 7.3.2).¶

7.1. Explicit Congestion Notification

If a path has been validated to support Explicit Congestion Notification (ECN) [RFC3168] [RFC8311], QUIC treats a Congestion Experienced (CE) codepoint in the IP header as a signal of congestion. This document specifies an endpoint's response when the peer-reported ECN-CE count increases; see Section 13.4.2 of [QUIC-TRANSPORT].¶

7.2. Initial and Minimum Congestion Window

QUIC begins every connection in slow start with the congestion window set to an initial value. Endpoints SHOULD use an initial congestion window of ten times the maximum datagram size (max_datagram_size), while limiting the window to the larger of 14,720 bytes or twice the maximum datagram size. This follows the analysis and recommendations in [RFC6928], increasing the byte limit to account for the smaller 8-byte overhead of UDP compared to the 20-byte overhead for TCP.¶

If the maximum datagram size changes during the connection, the initial congestion window SHOULD be recalculated with the new size. If the maximum datagram size is decreased in order to complete the handshake, the congestion window SHOULD be set to the new initial congestion window.¶

Prior to validating the client's address, the server can be further limited by the anti-amplification limit as specified in Section 8.1 of [QUIC-TRANSPORT]. Though the anti-amplification limit can prevent the congestion window from being fully utilized and therefore slow down the increase in congestion window, it does not directly affect the congestion window.¶

The minimum congestion window is the smallest value the congestion window can attain in response to loss, an increase in the peer-reported ECN-CE count, or persistent congestion. The RECOMMENDED value is 2 * max_datagram_size.¶

7.3. Congestion Control States

The NewReno congestion controller described in this document has three distinct states, as shown in Figure 1.¶

                 New path or      +------------+
            persistent congestion |   Slow     |
        (O)---------------------->|   Start    |
                                  +------------+
                                        |
                                Loss or |
                        ECN-CE increase |
                                        v
 +------------+     Loss or       +------------+
 | Congestion |  ECN-CE increase  |  Recovery  |
 | Avoidance  |------------------>|   Period   |
 +------------+                   +------------+
           ^                            |
           |                            |
           +----------------------------+
              Acknowledgment of packet
                sent during recovery

Figure 1: Congestion Control States and Transitions

These states and the transitions between them are described in subsequent sections.¶

7.3.1. Slow Start

A NewReno sender is in slow start any time the congestion window is below the slow start threshold. A sender begins in slow start because the slow start threshold is initialized to an infinite value.¶

While a sender is in slow start, the congestion window increases by the number of bytes acknowledged when each acknowledgment is processed. This results in exponential growth of the congestion window.¶

The sender MUST exit slow start and enter a recovery period when a packet is lost or when the ECN-CE count reported by its peer increases.¶

A sender reenters slow start any time the congestion window is less than the slow start threshold, which only occurs after persistent congestion is declared.¶

7.3.2. Recovery

A NewReno sender enters a recovery period when it detects the loss of a packet or when the ECN-CE count reported by its peer increases. A sender that is already in a recovery period stays in it and does not reenter it.¶

On entering a recovery period, a sender MUST set the slow start threshold to half the value of the congestion window when loss is detected. The congestion window MUST be set to the reduced value of the slow start threshold before exiting the recovery period.¶

Implementations MAY reduce the congestion window immediately upon entering a recovery period or use other mechanisms, such as Proportional Rate Reduction [PRR], to reduce the congestion window more gradually. If the congestion window is reduced immediately, a single packet can be sent prior to reduction. This speeds up loss recovery if the data in the lost packet is retransmitted and is similar to TCP as described in Section 5 of [RFC6675].¶

The recovery period aims to limit congestion window reduction to once per round trip. Therefore, during a recovery period, the congestion window does not change in response to new losses or increases in the ECN-CE count.¶

A recovery period ends and the sender enters congestion avoidance when a packet sent during the recovery period is acknowledged. This is slightly different from TCP's definition of recovery, which ends when the lost segment that started recovery is acknowledged [RFC5681].¶

7.3.3. Congestion Avoidance

A NewReno sender is in congestion avoidance any time the congestion window is at or above the slow start threshold and not in a recovery period.¶

A sender in congestion avoidance uses an Additive Increase Multiplicative Decrease (AIMD) approach that MUST limit the increase to the congestion window to at most one maximum datagram size for each congestion window that is acknowledged.¶

The sender exits congestion avoidance and enters a recovery period when a packet is lost or when the ECN-CE count reported by its peer increases.¶

7.4. Ignoring Loss of Undecryptable Packets

During the handshake, some packet protection keys might not be available when a packet arrives, and the receiver can choose to drop the packet. In particular, Handshake and 0-RTT packets cannot be processed until the Initial packets arrive, and 1-RTT packets cannot be processed until the handshake completes. Endpoints MAY ignore the loss of Handshake, 0-RTT, and 1-RTT packets that might have arrived before the peer had packet protection keys to process those packets. Endpoints MUST NOT ignore the loss of packets that were sent after the earliest acknowledged packet in a given packet number space.¶

7.5. Probe Timeout

Probe packets MUST NOT be blocked by the congestion controller. A sender MUST however count these packets as being additionally in flight, since these packets add network load without establishing packet loss. Note that sending probe packets might cause the sender's bytes in flight to exceed the congestion window until an acknowledgment is received that establishes loss or delivery of packets.¶

7.6. Persistent Congestion

When a sender establishes loss of all packets sent over a long enough duration, the network is considered to be experiencing persistent congestion.¶

7.6.1. Duration

The persistent congestion duration is computed as follows:¶

(smoothed_rtt + max(4*rttvar, kGranularity) + max_ack_delay) *
    kPersistentCongestionThreshold

Unlike the PTO computation in Section 6.2, this duration includes the max_ack_delay irrespective of the packet number spaces in which losses are established.¶

This duration allows a sender to send as many packets before establishing persistent congestion, including some in response to PTO expiration, as TCP does with Tail Loss Probes [RFC8985] and an RTO [RFC5681].¶

Larger values of kPersistentCongestionThreshold cause the sender to become less responsive to persistent congestion in the network, which can result in aggressive sending into a congested network. Too small a value can result in a sender declaring persistent congestion unnecessarily, resulting in reduced throughput for the sender.¶

The RECOMMENDED value for kPersistentCongestionThreshold is 3, which results in behavior that is approximately equivalent to a TCP sender declaring an RTO after two TLPs.¶

This design does not use consecutive PTO events to establish persistent congestion, since application patterns impact PTO expiration. For example, a sender that sends small amounts of data with silence periods between them restarts the PTO timer every time it sends, potentially preventing the PTO timer from expiring for a long period of time, even when no acknowledgments are being received. The use of a duration enables a sender to establish persistent congestion without depending on PTO expiration.¶

7.6.2. Establishing Persistent Congestion

A sender establishes persistent congestion after the receipt of an acknowledgment if two packets that are ack-eliciting are declared lost, and:¶

across all packet number spaces, none of the packets sent between the send times of these two packets are acknowledged;
the duration between the send times of these two packets exceeds the persistent congestion duration (Section 7.6.1); and
a prior RTT sample existed when these two packets were sent.

These two packets MUST be ack-eliciting, since a receiver is required to acknowledge only ack-eliciting packets within its maximum acknowledgment delay; see Section 13.2 of [QUIC-TRANSPORT].¶

The persistent congestion period SHOULD NOT start until there is at least one RTT sample. Before the first RTT sample, a sender arms its PTO timer based on the initial RTT (Section 6.2.2), which could be substantially larger than the actual RTT. Requiring a prior RTT sample prevents a sender from establishing persistent congestion with potentially too few probes.¶

Since network congestion is not affected by packet number spaces, persistent congestion SHOULD consider packets sent across packet number spaces. A sender that does not have state for all packet number spaces or an implementation that cannot compare send times across packet number spaces MAY use state for just the packet number space that was acknowledged. This might result in erroneously declaring persistent congestion, but it will not lead to a failure to detect persistent congestion.¶

When persistent congestion is declared, the sender's congestion window MUST be reduced to the minimum congestion window (kMinimumWindow), similar to a TCP sender's response on an RTO [RFC5681].¶

7.6.3. Example

The following example illustrates how a sender might establish persistent congestion. Assume:¶

smoothed_rtt + max(4*rttvar, kGranularity) + max_ack_delay = 2
kPersistentCongestionThreshold = 3

Consider the following sequence of events:¶

Table 1
Time	Action
t=0	Send packet #1 (application data)
t=1	Send packet #2 (application data)
t=1.2	Receive acknowledgment of #1
t=2	Send packet #3 (application data)
t=3	Send packet #4 (application data)
t=4	Send packet #5 (application data)
t=5	Send packet #6 (application data)
t=6	Send packet #7 (application data)
t=8	Send packet #8 (PTO 1)
t=12	Send packet #9 (PTO 2)
t=12.2	Receive acknowledgment of #9

Packets 2 through 8 are declared lost when the acknowledgment for packet 9 is received at t = 12.2.¶

The congestion period is calculated as the time between the oldest and newest lost packets: 8 - 1 = 7. The persistent congestion duration is 2 * 3 = 6. Because the threshold was reached and because none of the packets between the oldest and the newest lost packets were acknowledged, the network is considered to have experienced persistent congestion.¶

While this example shows PTO expiration, they are not required for persistent congestion to be established.¶

7.7. Pacing

A sender SHOULD pace sending of all in-flight packets based on input from the congestion controller.¶

Sending multiple packets into the network without any delay between them creates a packet burst that might cause short-term congestion and losses. Senders MUST either use pacing or limit such bursts. Senders SHOULD limit bursts to the initial congestion window; see Section 7.2. A sender with knowledge that the network path to the receiver can absorb larger bursts MAY use a higher limit.¶

An implementation should take care to architect its congestion controller to work well with a pacer. For instance, a pacer might wrap the congestion controller and control the availability of the congestion window, or a pacer might pace out packets handed to it by the congestion controller.¶

Timely delivery of ACK frames is important for efficient loss recovery. To avoid delaying their delivery to the peer, packets containing only ACK frames SHOULD therefore not be paced.¶

Endpoints can implement pacing as they choose. A perfectly paced sender spreads packets exactly evenly over time. For a window-based congestion controller, such as the one in this document, that rate can be computed by averaging the congestion window over the RTT. Expressed as a rate in units of bytes per time, where congestion_window is in bytes:¶

rate = N * congestion_window / smoothed_rtt

Or expressed as an inter-packet interval in units of time:¶

interval = ( smoothed_rtt * packet_size / congestion_window ) / N

Using a value for N that is small, but at least 1 (for example, 1.25) ensures that variations in RTT do not result in underutilization of the congestion window.¶

Practical considerations, such as packetization, scheduling delays, and computational efficiency, can cause a sender to deviate from this rate over time periods that are much shorter than an RTT.¶

One possible implementation strategy for pacing uses a leaky bucket algorithm, where the capacity of the "bucket" is limited to the maximum burst size and the rate the "bucket" fills is determined by the above function.¶

7.8. Underutilizing the Congestion Window

When bytes in flight is smaller than the congestion window and sending is not pacing limited, the congestion window is underutilized. This can happen due to insufficient application data or flow control limits. When this occurs, the congestion window SHOULD NOT be increased in either slow start or congestion avoidance.¶

A sender that paces packets (see Section 7.7) might delay sending packets and not fully utilize the congestion window due to this delay. A sender SHOULD NOT consider itself application limited if it would have fully utilized the congestion window without pacing delay.¶

A sender MAY implement alternative mechanisms to update its congestion window after periods of underutilization, such as those proposed for TCP in [RFC7661].¶