| QUIC Working Group | J. Iyengar, Editor |
| Internet-Draft | Fastly |
| Intended status: Standards Track | I. Swett, Editor |
| Expires: December 30, 2018 | |
| June 28, 2018 |
This document describes loss detection and congestion control mechanisms for QUIC.¶
Discussion of this draft takes place on the QUIC working group mailing list (quic@ietf.org), which is archived at https://mailarchive.ietf.org/arch/search/?email_list=quic.¶
Working Group information can be found at https://github.com/quicwg; source code and issues list for this draft can be found at https://github.com/quicwg/base-drafts/labels/-recovery.¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as “work in progress”.¶
This Internet-Draft will expire on December 30, 2018.¶
Copyright (c) 2018 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License.¶
QUIC is a new multiplexed and secure transport atop UDP. QUIC builds on decades of transport and security experience, and implements mechanisms that make it attractive as a modern general-purpose transport. The QUIC protocol is described in [QUIC-TRANSPORT].¶
QUIC implements the spirit of known TCP loss recovery mechanisms, described in RFCs, various Internet-drafts, and also those prevalent in the Linux TCP implementation. This document describes QUIC congestion control and loss recovery, and where applicable, attributes the TCP equivalent in RFCs, Internet-drafts, academic papers, and/or TCP implementations.¶
All transmissions in QUIC are sent with a packet-level header, which indicates the encryption level and includes a packet sequence number (referred to below as a packet number). The encryption level indicates the packet number space, as described in [QUIC-TRANSPORT]. Packet numbers never repeat within a packet number space for the lifetime of a connection. Packet numbers monotonically increase within a space, preventing ambiguity.¶
This design obviates the need for disambiguating between transmissions and retransmissions and eliminates significant complexity from QUIC’s interpretation of TCP loss detection mechanisms.¶
Every packet may contain several frames. We outline the frames that are important to the loss detection and congestion control machinery below.¶
Readers familiar with TCP’s loss detection and congestion control will find algorithms here that parallel well-known TCP ones. Protocol differences between QUIC and TCP however contribute to algorithmic differences. We briefly describe these protocol differences below.¶
QUIC uses separate packet number spaces for each encryption level, except 0-RTT and all generations of 1-RTT keys use the same packet number space. Separate packet number spaces ensures acknowledgement of packets sent with one level of encryption will not cause spurious retransmission of packets sent with a different encryption level. Congestion control and RTT measurement are unified across packet number spaces.¶
TCP conflates transmission sequence number at the sender with delivery sequence number at the receiver, which results in retransmissions of the same data carrying the same sequence number, and consequently to problems caused by “retransmission ambiguity”. QUIC separates the two: QUIC uses a packet number for transmissions, and any data that is to be delivered to the receiving application(s) is sent in one or more streams, with delivery order determined by stream offsets encoded within STREAM frames.¶
QUIC’s packet number is strictly increasing, and directly encodes transmission order. A higher QUIC packet number signifies that the packet was sent later, and a lower QUIC packet number signifies that the packet was sent earlier. When a packet containing frames is deemed lost, QUIC rebundles necessary frames in a new packet with a new packet number, removing ambiguity about which packet is acknowledged when an ACK is received. Consequently, more accurate RTT measurements can be made, spurious retransmissions are trivially detected, and mechanisms such as Fast Retransmit can be applied universally, based only on packet number.¶
This design point significantly simplifies loss detection mechanisms for QUIC. Most TCP mechanisms implicitly attempt to infer transmission ordering based on TCP sequence numbers - a non-trivial task, especially when TCP timestamps are not available.¶
QUIC ACKs contain information that is similar to TCP SACK, but QUIC does not allow any acked packet to be reneged, greatly simplifying implementations on both sides and reducing memory pressure on the sender.¶
QUIC supports many ACK ranges, opposed to TCP’s 3 SACK ranges. In high loss environments, this speeds recovery, reduces spurious retransmits, and ensures forward progress without relying on timeouts.¶
QUIC ACKs explicitly encode the delay incurred at the receiver between when a packet is received and when the corresponding ACK is sent. This allows the receiver of the ACK to adjust for receiver delays, specifically the delayed ack timer, when estimating the path RTT. This mechanism also allows a receiver to measure and report the delay from when a packet was received by the OS kernel, which is useful in receivers which may incur delays such as context-switch latency before a userspace QUIC receiver processes a received packet.¶
QUIC senders use both ack information and timeouts to detect lost packets, and this section provides a description of these algorithms. Estimating the network round-trip time (RTT) is critical to these algorithms and is described first.¶
RTT is calculated when an ACK frame arrives by computing the difference between the current time and the time the largest newly acked packet was sent. If no packets are newly acknowledged, RTT cannot be calculated. When RTT is calculated, the ack delay field from the ACK frame SHOULD be subtracted from the RTT as long as the result is larger than the Min RTT. If the result is smaller than the min_rtt, the RTT should be used, but the ack delay field should be ignored.¶
Like TCP, QUIC calculates both smoothed RTT and RTT variance similar to those specified in [RFC6298].¶
Min RTT is the minimum RTT measured over the connection, prior to adjusting by ack delay. Ignoring ack delay for min RTT prevents intentional or unintentional underestimation of min RTT, which in turn prevents underestimating smoothed RTT.¶
Ack-based loss detection implements the spirit of TCP’s Fast Retransmit [RFC5681], Early Retransmit [RFC5827], FACK, and SACK loss recovery [RFC6675]. This section provides an overview of how these algorithms are implemented in QUIC.¶
An unacknowledged packet is marked as lost when an acknowledgment is received for a packet that was sent a threshold number of packets (kReorderingThreshold) after the unacknowledged packet. Receipt of the ack indicates that a later packet was received, while kReorderingThreshold provides some tolerance for reordering of packets in the network.¶
The RECOMMENDED initial value for kReorderingThreshold is 3.¶
We derive this recommendation from TCP loss recovery [RFC5681] [RFC6675]. It is possible for networks to exhibit higher degrees of reordering, causing a sender to detect spurious losses. Detecting spurious losses leads to unnecessary retransmissions and may result in degraded performance due to the actions of the congestion controller upon detecting loss. Implementers MAY use algorithms developed for TCP, such as TCP-NCR [RFC4653], to improve QUIC’s reordering resilience, though care should be taken to map TCP specifics to QUIC correctly. Similarly, using time-based loss detection to deal with reordering, such as in PR-TCP, should be more readily usable in QUIC. Making QUIC deal with such networks is important open research, and implementers are encouraged to explore this space.¶
Unacknowledged packets close to the tail may have fewer than kReorderingThreshold retransmittable packets sent after them. Loss of such packets cannot be detected via Fast Retransmit. To enable ack-based loss detection of such packets, receipt of an acknowledgment for the last outstanding retransmittable packet triggers the Early Retransmit process, as follows.¶
If there are unacknowledged retransmittable packets still pending, they should be marked as lost. To compensate for the reduced reordering resilience, the sender SHOULD set an alarm for a small period of time. If the unacknowledged retransmittable packets are not acknowledged during this time, then these packets MUST be marked as lost.¶
An endpoint SHOULD set the alarm such that a packet is marked as lost no earlier than 1.25 * max(SRTT, latest_RTT) since when it was sent.¶
Using max(SRTT, latest_RTT) protects from the two following cases:¶
The 1.25 multiplier increases reordering resilience. Implementers MAY experiment with using other multipliers, bearing in mind that a lower multiplier reduces reordering resilience and increases spurious retransmissions, and a higher multipler increases loss recovery delay.¶
This mechanism is based on Early Retransmit for TCP [RFC5827]. However, [RFC5827] does not include the alarm described above. Early Retransmit is prone to spurious retransmissions due to its reduced reordering resilence without the alarm. This observation led Linux TCP implementers to implement an alarm for TCP as well, and this document incorporates this advancement.¶
Timer-based loss detection implements a handshake retransmission timer that is optimized for QUIC as well as the spirit of TCP’s Tail Loss Probe and Retransmission Timeout mechanisms.¶
Data in CRYPTO frames is critical to QUIC transport and crypto negotiation, so a more aggressive timeout is used to retransmit it. Below, the term “handshake packet” is used to refer to packets containing CRYPTO frames, not packets with the specific long header packet type Handshake.¶
The initial handshake timeout SHOULD be set to twice the initial RTT.¶
At the beginning, there are no prior RTT samples within a connection. Resumed connections over the same network SHOULD use the previous connection’s final smoothed RTT value as the resumed connection’s initial RTT.¶
If no previous RTT is available, or if the network changes, the initial RTT SHOULD be set to 100ms.¶
When CRYPTO frames are sent, the sender SHOULD set an alarm for the handshake timeout period. When the alarm fires, the sender MUST retransmit all unacknowledged CRYPTO data by calling RetransmitAllUnackedHandshakeData(). On each consecutive firing of the handshake alarm without receiving an acknowledgement for a new packet, the sender SHOULD double the handshake timeout and set an alarm for this period.¶
When CRYPTO frames are outstanding, the TLP and RTO timers are not active unless the CRYPTO frames were sent at 1RTT encryption.¶
When an acknowledgement is received for a handshake packet, the new RTT is computed and the alarm SHOULD be set for twice the newly computed smoothed RTT.¶
A Retry packet causes the content of the client’s Initial packet to be immediately retransmitted along with the token present in the Retry.¶
The Retry indicates that the Initial was received but not processed. It MUST NOT be treated as an acknowledgment for the Initial, but it MAY be used for an RTT measurement.¶
The algorithm described in this section is an adaptation of the Tail Loss Probe algorithm proposed for TCP [TLP].¶
A packet sent at the tail is particularly vulnerable to slow loss detection, since acks of subsequent packets are needed to trigger ack-based detection. To ameliorate this weakness of tail packets, the sender schedules an alarm when the last retransmittable packet before quiescence is transmitted. When this alarm fires, a Tail Loss Probe (TLP) packet is sent to evoke an acknowledgement from the receiver.¶
The alarm duration, or Probe Timeout (PTO), is set based on the following conditions:¶
MaxAckDelay is the maximum ack delay supplied in an incoming ACK frame. MaxAckDelay excludes ack delays that aren’t included in an RTT sample because they’re too large and excludes those which reference an ack-only packet.¶
QUIC diverges from TCP by calculating MaxAckDelay dynamically, instead of assuming a constant delayed ack timeout for all connections. QUIC includes this in all probe timeouts, because it assume the ack delay may come into play, regardless of the number of packets outstanding. TCP’s TLP assumes if at least 2 packets are outstanding, acks will not be delayed.¶
A PTO value of at least 1.5*SRTT ensures that the ACK is overdue. The 1.5 is based on [TLP], but implementations MAY experiment with other constants.¶
To reduce latency, it is RECOMMENDED that the sender set and allow the TLP alarm to fire twice before setting an RTO alarm. In other words, when the TLP alarm fires the first time, a TLP packet is sent, and it is RECOMMENDED that the TLP alarm be scheduled for a second time. When the TLP alarm fires the second time, a second TLP packet is sent, and an RTO alarm SHOULD be scheduled Section 3.3.3.¶
A TLP packet SHOULD carry new data when possible. If new data is unavailable or new data cannot be sent due to flow control, a TLP packet MAY retransmit unacknowledged data to potentially reduce recovery time. Since a TLP alarm is used to send a probe into the network prior to establishing any packet loss, prior unacknowledged packets SHOULD NOT be marked as lost when a TLP alarm fires.¶
A sender may not know that a packet being sent is a tail packet. Consequently, a sender may have to arm or adjust the TLP alarm on every sent retransmittable packet.¶
A Retransmission Timeout (RTO) alarm is the final backstop for loss detection. The algorithm used in QUIC is based on the RTO algorithm for TCP [RFC5681] and is additionally resilient to spurious RTO events [RFC5682].¶
When the last TLP packet is sent, an alarm is scheduled for the RTO period. When this alarm fires, the sender sends two packets, to evoke acknowledgements from the receiver, and restarts the RTO alarm.¶
The sender typically has incurred a high latency penalty by the time an RTO alarm fires, and this penalty increases exponentially in subsequent consecutive RTO events. Sending a single packet on an RTO event therefore makes the connection very sensitive to single packet loss. Sending two packets instead of one significantly increases resilience to packet drop in both directions, thus reducing the probability of consecutive RTO events.¶
QUIC’s RTO algorithm differs from TCP in that the firing of an RTO alarm is not considered a strong enough signal of packet loss, so does not result in an immediate change to congestion window or recovery state. An RTO alarm fires only when there’s a prolonged period of network silence, which could be caused by a change in the underlying network RTT.¶
QUIC also diverges from TCP by including MaxAckDelay in the RTO period. QUIC is able to explicitly model delay at the receiver via the ack delay field in the ACK frame. Since QUIC corrects for this delay in its SRTT and RTTVAR computations, it is necessary to add this delay explicitly in the TLP and RTO computation.¶
When an acknowledgment is received for a packet sent on an RTO event, any unacknowledged packets with lower packet numbers than those acknowledged MUST be marked as lost.¶
A packet sent when an RTO alarm fires MAY carry new data if available or unacknowledged data to potentially reduce recovery time. Since this packet is sent as a probe into the network prior to establishing any packet loss, prior unacknowledged packets SHOULD NOT be marked as lost.¶
A packet sent on an RTO alarm MUST NOT be blocked by the sender’s congestion controller. A sender MUST however count these bytes as additional bytes in flight, since this packet adds network load without establishing packet loss.¶
QUIC SHOULD delay sending acknowledgements in response to packets, but MUST NOT excessively delay acknowledgements of packets containing frames other than ACK or ACN_ECN. Specifically, implementaions MUST attempt to enforce a maximum ack delay to avoid causing the peer spurious timeouts. The RECOMMENDED maximum ack delay in QUIC is 25ms.¶
An acknowledgement MAY be sent for every second full-sized packet, as TCP does [RFC5681], or may be sent less frequently, as long as the delay does not exceed the maximum ack delay. QUIC recovery algorithms do not assume the peer generates an acknowledgement immediately when receiving a second full-sized packet.¶
Out-of-order packets SHOULD be acknowledged more quickly, in order to accelerate loss recovery. The receiver SHOULD send an immediate ACK when it receives a new packet which is not one greater than the largest received packet number.¶
Similarly, packets marked with the ECN Congestion Experienced (CE) codepoint in the IP header SHOULD be acknowledged immediately, to reduce the peer’s response time to congestion events.¶
As an optimization, a receiver MAY process multiple packets before sending any ACK frames in response. In this case they can determine whether an immediate or delayed acknowledgement should be generated after processing incoming packets.¶
In order to quickly complete the handshake and avoid spurious retransmissions due to handshake alarm timeouts, handshake packets SHOULD use a very short ack delay, such as 1ms. ACK frames MAY be sent immediately when the crypto stack indicates all data for that encryption level has been received.¶
When an ACK frame is sent, one or more ranges of acknowledged packets are included. Including older packets reduces the chance of spurious retransmits caused by losing previously sent ACK frames, at the cost of larger ACK frames.¶
ACK frames SHOULD always acknowledge the most recently received packets, and the more out-of-order the packets are, the more important it is to send an updated ACK frame quickly, to prevent the peer from declaring a packet as lost and spuriusly retransmitting the frames it contains.¶
Below is one recommended approach for determining what packets to include in an ACK frame.¶
When a packet containing an ACK frame is sent, the largest acknowledged in that frame may be saved. When a packet containing an ACK frame is acknowledged, the receiver can stop acknowledging packets less than or equal to the largest acknowledged in the sent ACK frame.¶
In cases without ACK frame loss, this algorithm allows for a minimum of 1 RTT of reordering. In cases with ACK frame loss, this approach does not guarantee that every acknowledgement is seen by the sender before it is no longer included in the ACK frame. Packets could be received out of order and all subsequent ACK frames containing them could be lost. In this case, the loss recovery algorithm may cause spurious retransmits, but the sender will continue making forward progress.¶
Constants used in loss recovery are based on a combination of RFCs, papers, and common practice. Some may need to be changed or negotiated in order to better suit a variety of environments.¶
Variables required to implement the congestion control mechanisms are described in this section.¶
At the beginning of the connection, initialize the loss detection variables as follows:¶
loss_detection_alarm.reset()
handshake_count = 0
tlp_count = 0
rto_count = 0
if (kUsingTimeLossDetection)
reordering_threshold = infinite
time_reordering_fraction = kTimeReorderingFraction
else:
reordering_threshold = kReorderingThreshold
time_reordering_fraction = infinite
loss_time = 0
smoothed_rtt = 0
rttvar = 0
min_rtt = infinite
max_ack_delay = 0
largest_sent_before_rto = 0
time_of_last_sent_retransmittable_packet = 0
time_of_last_sent_handshake_packet = 0
largest_sent_packet = 0
After any packet is sent, be it a new transmission or a rebundled transmission, the following OnPacketSent function is called. The parameters to OnPacketSent are as follows:¶
Pseudocode for OnPacketSent follows:¶
OnPacketSent(packet_number, is_ack_only, is_handshake_packet,
sent_bytes):
largest_sent_packet = packet_number
sent_packets[packet_number].packet_number = packet_number
sent_packets[packet_number].time = now
sent_packets[packet_number].ack_only = is_ack_only
if !is_ack_only:
if is_handshake_packet:
time_of_last_sent_handshake_packet = now
time_of_last_sent_retransmittable_packet = now
OnPacketSentCC(sent_bytes)
sent_packets[packet_number].bytes = sent_bytes
SetLossDetectionAlarm()
When an ACK frame is received, it may acknowledge 0 or more packets.¶
Pseudocode for OnAckReceived and UpdateRtt follow:¶
OnAckReceived(ack):
largest_acked_packet = ack.largest_acked
// If the largest acked is newly acked, update the RTT.
if (sent_packets[ack.largest_acked]):
latest_rtt = now - sent_packets[ack.largest_acked].time
UpdateRtt(latest_rtt, ack.ack_delay)
// Find all newly acked packets.
for acked_packet in DetermineNewlyAckedPackets():
OnPacketAcked(acked_packet.packet_number)
DetectLostPackets(ack.largest_acked_packet)
SetLossDetectionAlarm()
// Process ECN information if present.
if (ACK frame contains ECN information):
ProcessECN(ack)
UpdateRtt(latest_rtt, ack_delay):
// min_rtt ignores ack delay.
min_rtt = min(min_rtt, latest_rtt)
// Adjust for ack delay if it's plausible.
if (latest_rtt - min_rtt > ack_delay):
latest_rtt -= ack_delay
// Only save into max ack delay if it's used
// for rtt calculation and is not ack only.
if (!sent_packets[ack.largest_acked].ack_only)
max_ack_delay = max(max_ack_delay, ack_delay)
// Based on {{RFC6298}}.
if (smoothed_rtt == 0):
smoothed_rtt = latest_rtt
rttvar = latest_rtt / 2
else:
rttvar_sample = abs(smoothed_rtt - latest_rtt)
rttvar = 3/4 * rttvar + 1/4 * rttvar_sample
smoothed_rtt = 7/8 * smoothed_rtt + 1/8 * latest_rtt
When a packet is acked for the first time, the following OnPacketAcked function is called. Note that a single ACK frame may newly acknowledge several packets. OnPacketAcked must be called once for each of these newly acked packets.¶
OnPacketAcked takes one parameter, acked_packet, which is the struct of the newly acked packet.¶
If this is the first acknowledgement following RTO, check if the smallest newly acknowledged packet is one sent by the RTO, and if so, inform congestion control of a verified RTO, similar to F-RTO [RFC5682].¶
Pseudocode for OnPacketAcked follows:¶
OnPacketAcked(acked_packet):
if (!acked_packet.is_ack_only):
OnPacketAckedCC(acked_packet)
// If a packet sent prior to RTO was acked, then the RTO
// was spurious. Otherwise, inform congestion control.
if (rto_count > 0 &&
acked_packet.packet_number > largest_sent_before_rto)
OnRetransmissionTimeoutVerified()
handshake_count = 0
tlp_count = 0
rto_count = 0
sent_packets.remove(acked_packet.packet_number)
QUIC loss detection uses a single alarm for all timer-based loss detection. The duration of the alarm is based on the alarm’s mode, which is set in the packet and timer events further below. The function SetLossDetectionAlarm defined below shows how the single timer is set based on the alarm mode.¶
When a connection has unacknowledged handshake data, the handshake alarm is set and when it expires, all unacknowledgedd handshake data is retransmitted.¶
When stateless rejects are in use, the connection is considered immediately closed once a reject is sent, so no timer is set to retransmit the reject.¶
Version negotiation packets are always stateless, and MUST be sent once per handshake packet that uses an unsupported QUIC version, and MAY be sent in response to 0-RTT packets.¶
Tail loss probes [TLP] and retransmission timeouts [RFC6298] are an alarm based mechanism to recover from cases when there are outstanding retransmittable packets, but an acknowledgement has not been received in a timely manner.¶
The TLP and RTO timers are armed when there is not unacknowledged handshake data. The TLP alarm is set until the max number of TLP packets have been sent, and then the RTO timer is set.¶
Pseudocode for SetLossDetectionAlarm follows:¶
SetLossDetectionAlarm():
// Don't arm the alarm if there are no packets with
// retransmittable data in flight.
if (bytes_in_flight == 0):
loss_detection_alarm.cancel()
return
if (handshake packets are outstanding):
// Handshake retransmission alarm.
if (smoothed_rtt == 0):
alarm_duration = 2 * kInitialRtt
else:
alarm_duration = 2 * smoothed_rtt
alarm_duration = max(alarm_duration + max_ack_delay,
kMinTLPTimeout)
alarm_duration = alarm_duration * (2 ^ handshake_count)
loss_detection_alarm.set(
time_of_last_sent_handshake_packet + alarm_duration)
return;
else if (loss_time != 0):
// Early retransmit timer or time loss detection.
alarm_duration = loss_time -
time_of_last_sent_retransmittable_packet
else:
// RTO or TLP alarm
// Calculate RTO duration
alarm_duration =
smoothed_rtt + 4 * rttvar + max_ack_delay
alarm_duration = max(alarm_duration, kMinRTOTimeout)
alarm_duration = alarm_duration * (2 ^ rto_count)
if (tlp_count < kMaxTLPs):
// Tail Loss Probe
tlp_alarm_duration = max(1.5 * smoothed_rtt
+ max_ack_delay, kMinTLPTimeout)
alarm_duration = min(tlp_alarm_duration, alarm_duration)
loss_detection_alarm.set(
time_of_last_sent_retransmittable_packet + alarm_duration)
QUIC uses one loss recovery alarm, which when set, can be in one of several modes. When the alarm fires, the mode determines the action to be performed.¶
Pseudocode for OnLossDetectionAlarm follows:¶
OnLossDetectionAlarm():
if (handshake packets are outstanding):
// Handshake retransmission alarm.
RetransmitAllUnackedHandshakeData()
handshake_count++
else if (loss_time != 0):
// Early retransmit or Time Loss Detection
DetectLostPackets(largest_acked_packet)
else if (tlp_count < kMaxTLPs):
// Tail Loss Probe.
SendOnePacket()
tlp_count++
else:
// RTO.
if (rto_count == 0)
largest_sent_before_rto = largest_sent_packet
SendTwoPackets()
rto_count++
SetLossDetectionAlarm()
Packets in QUIC are only considered lost once a larger packet number in the same packet number space is acknowledged. DetectLostPackets is called every time an ack is received and operates on the sent_packets for that packet number space. If the loss detection alarm fires and the loss_time is set, the previous largest acked packet is supplied.¶
DetectLostPackets takes one parameter, acked, which is the largest acked packet.¶
Pseudocode for DetectLostPackets follows:¶
DetectLostPackets(largest_acked):
loss_time = 0
lost_packets = {}
delay_until_lost = infinite
if (kUsingTimeLossDetection):
delay_until_lost =
(1 + time_reordering_fraction) *
max(latest_rtt, smoothed_rtt)
else if (largest_acked.packet_number == largest_sent_packet):
// Early retransmit alarm.
delay_until_lost = 5/4 * max(latest_rtt, smoothed_rtt)
foreach (unacked < largest_acked.packet_number):
time_since_sent = now() - unacked.time_sent
delta = largest_acked.packet_number - unacked.packet_number
if (time_since_sent > delay_until_lost ||
delta > reordering_threshold):
sent_packets.remove(unacked.packet_number)
if (!unacked.is_ack_only):
lost_packets.insert(unacked)
else if (loss_time == 0 && delay_until_lost != infinite):
loss_time = now() + delay_until_lost - time_since_sent
// Inform the congestion controller of lost packets and
// lets it decide whether to retransmit immediately.
if (!lost_packets.empty()):
OnPacketsLost(lost_packets)
The majority of constants were derived from best common practices among widely deployed TCP implementations on the internet. Exceptions follow.¶
A shorter delayed ack time of 25ms was chosen because longer delayed acks can delay loss recovery and for the small number of connections where less than packet per 25ms is delivered, acking every packet is beneficial to congestion control and loss recovery.¶
The default initial RTT of 100ms was chosen because it is slightly higher than both the median and mean min_rtt typically observed on the public internet.¶
QUIC’s congestion control is based on TCP NewReno [RFC6582] congestion control to determine the congestion window. QUIC congestion control is specified in bytes due to finer control and the ease of appropriate byte counting [RFC3465].¶
QUIC hosts MUST NOT send packets if they would increase bytes_in_flight (defined in Section 4.8.2) beyond the available congestion window, unless the packet is a probe packet sent after the TLP or RTO alarm fires, as described in Section 3.3.2 and Section 3.3.3.¶
If a path has been verified to support ECN, QUIC treats a Congestion Experienced codepoint in the IP header as a signal of congestion. This document specifies an endpoint’s response when its peer receives packets with the Congestion Experienced codepoint. As discussed in [RFC8311], endpoints are permitted to experiment with other response functions.¶
QUIC begins every connection in slow start and exits slow start upon loss or upon increase in the ECN-CE counter. QUIC re-enters slow start anytime the congestion window is less than sshthresh, which typically only occurs after an RTO. While in slow start, QUIC increases the congestion window by the number of bytes acknowledged when each ack is processed.¶
Slow start exits to congestion avoidance. Congestion avoidance in NewReno uses an additive increase multiplicative decrease (AIMD) approach that increases the congestion window by one MSS of bytes per congestion window acknowledged. When a loss is detected, NewReno halves the congestion window and sets the slow start threshold to the new congestion window.¶
Recovery is a period of time beginning with detection of a lost packet or an increase in the ECN-CE counter. Because QUIC retransmits stream data and control frames, not packets, it defines the end of recovery as a packet sent after the start of recovery being acknowledged. This is slightly different from TCP’s definition of recovery, which ends when the lost packet that started recovery is acknowledged.¶
The recovery period limits congestion window reduction to once per round trip. During recovery, the congestion window remains unchanged irrespective of new losses or increases in the ECN-CE counter.¶
A TLP packet MUST NOT be blocked by the sender’s congestion controller. The sender MUST however count these bytes as additional bytes-in-flight, since a TLP adds network load without establishing packet loss.¶
Acknowledgement or loss of tail loss probes are treated like any other packet.¶
When retransmissions are sent due to a retransmission timeout alarm, no change is made to the congestion window until the next acknowledgement arrives. The retransmission timeout is considered spurious when this acknowledgement acknowledges packets sent prior to the first retransmission timeout. The retransmission timeout is considered valid when this acknowledgement acknowledges no packets sent prior to the first retransmission timeout. In this case, the congestion window MUST be reduced to the minimum congestion window and slow start is re-entered.¶
This document does not specify a pacer, but it is RECOMMENDED that a sender pace sending of all retransmittable packets based on input from the congestion controller. For example, a pacer might distribute the congestion window over the SRTT when used with a window-based controller, and a pacer might use the rate estimate of a rate-based controller.¶
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. Packets containing only ACK frames should therefore not be paced, to avoid delaying their delivery to the peer.¶
As an example of a well-known and publicly available implementation of a flow pacer, implementers are referred to the Fair Queue packet scheduler (fq qdisc) in Linux (3.11 onwards).¶
Constants used in congestion control are based on a combination of RFCs, papers, and common practice. Some may need to be changed or negotiated in order to better suit a variety of environments.¶
Variables required to implement the congestion control mechanisms are described in this section.¶
At the beginning of the connection, initialize the congestion control variables as follows:¶
congestion_window = kInitialWindow bytes_in_flight = 0 end_of_recovery = 0 ssthresh = infinite ecn_ce_counter = 0
Whenever a packet is sent, and it contains non-ACK frames, the packet increases bytes_in_flight.¶
OnPacketSentCC(bytes_sent):
bytes_in_flight += bytes_sent
Invoked from loss detection’s OnPacketAcked and is supplied with acked_packet from sent_packets.¶
InRecovery(packet_number):
return packet_number <= end_of_recovery
OnPacketAckedCC(acked_packet):
// Remove from bytes_in_flight.
bytes_in_flight -= acked_packet.bytes
if (InRecovery(acked_packet.packet_number)):
// Do not increase congestion window in recovery period.
return
if (congestion_window < ssthresh):
// Slow start.
congestion_window += acked_packet.bytes
else:
// Congestion avoidance.
congestion_window +=
kInitialMss * acked_packet.bytes / congestion_window
Invoked from ProcessECN and OnPacketLost when a new congestion event is detected. Starts a new recovery period and reduces the congestion window.¶
CongestionEvent(packet_number):
// Start a new congestion event if packet_number
// is larger than the end of the previous recovery epoch.
if (!InRecovery(packet_number)):
end_of_recovery = largest_sent_packet
congestion_window *= kMarkReductionFactor
congestion_window = max(congestion_window, kMinimumWindow)
Invoked when an ACK_ECN frame is received from the peer.¶
ProcessECN(ack):
// If the ECN-CE counter reported by the peer has increased,
// this could be a new congestion event.
if (ack.ce_counter > ecn_ce_counter):
ecn_ce_counter = ack.ce_counter
// Start a new congestion event if the last acknowledged
// packet is past the end of the previous recovery epoch.
CongestionEvent(ack.largest_acked_packet)
Invoked by loss detection from DetectLostPackets when new packets are detected lost.¶
OnPacketsLost(lost_packets):
// Remove lost packets from bytes_in_flight.
for (lost_packet : lost_packets):
bytes_in_flight -= lost_packet.bytes
largest_lost_packet = lost_packets.last()
// Start a new congestion epoch if the last lost packet
// is past the end of the previous recovery epoch.
CongestionEvent(largest_lost_packet.packet_number)
QUIC decreases the congestion window to the minimum value once the retransmission timeout has been verified.¶
OnRetransmissionTimeoutVerified()
congestion_window = kMinimumWindow
This document has no IANA actions. Yet.¶
No significant changes.¶
No significant changes.¶
No significant changes.¶
No significant changes.¶
No significant changes.¶