Google

jri@google.com

Google

ianswett@google.com

Transport QUIC 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.

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 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.

The words “MUST”, “MUST NOT”, “SHOULD”, and “MAY” are used in this document. It’s not shouting; when they are capitalized, they have the special meaning defined in .

All transmissions in QUIC are sent with a packet-level header, which includes a packet sequence number (referred to below as a packet number). These packet numbers never repeat in the lifetime of a connection, and are monotonically increasing, which makes duplicate detection trivial. This fundamental design decision 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. Retransmittable frames are frames requiring reliable delivery. The most common are STREAM frames, which typically contain application data. Crypto handshake data is sent on stream 0, and uses the reliability machinery of QUIC underneath. ACK frames contain acknowledgment information. QUIC uses a SACK-based scheme, where acks express up to 256 ranges. The ACK frame also includes a receive timestamp for each packet newly acked.

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.

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 sequence number (referred to as the “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 stream offsets encoded within STREAM frames inside of packets that determine delivery order. 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 equivalent 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 up to 256 ACK ranges, opposed to TCP’s 3 SACK ranges. In high loss environments, this speeds recovery.

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 uses a combination of ack information and alarms to detect lost packets. An unacknowledged QUIC packet is marked as lost in one of the following ways: A packet is marked as lost if at least one packet that was sent a threshold number of packets (kReorderingThreshold) after it has been acknowledged. This indicates that the unacknowledged packet is either lost or reordered beyond the specified threshold. This mechanism combines both TCP’s FastRetransmit and FACK mechanisms. If a packet is near the tail, where fewer than kReorderingThreshold packets are sent after it, the sender cannot expect to detect loss based on the previous mechanism. In this case, a sender uses both ack information and an alarm to detect loss. Specifically, when the last sent packet is acknowledged, the sender waits a short period of time to allow for reordering and then marks any unacknowledged packets as lost. This mechanism is based on the Linux implementation of TCP Early Retransmit. If a packet is sent at the tail, there are no packets sent after it, and the sender cannot use ack information to detect its loss. The sender therefore relies on an alarm to detect such tail losses. This mechanism is based on TCP’s Tail Loss Probe. If all else fails, a Retransmission Timeout (RTO) alarm is always set when any retransmittable packet is outstanding. When this alarm fires, all unacknowledged packets are marked as lost. Instead of a packet threshold to tolerate reordering, a QUIC sender may use a time threshold. This allows for senders to be tolerant of short periods of significant reordering. In this mechanism, a QUIC sender marks a packet as lost when a larger packet number is acknowledged and a threshold amount of time has passed since the packet was sent. Handshake packets, which contain STREAM frames for stream 0, are critical to QUIC transport and crypto negotiation, so a separate alarm period is used for them.

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. Maximum number of tail loss probes before an RTO fires. Maximum reordering in packet number space before FACK style loss detection considers a packet lost. Maximum reordering in time space before time based loss detection considers a packet lost. In fraction of an RTT. Minimum time in the future a tail loss probe alarm may be set for. Minimum time in the future an RTO alarm may be set for. The length of the peer’s delayed ack timer. The default RTT used before an RTT sample is taken.

Variables required to implement the congestion control mechanisms are described in this section. Multi-modal alarm used for loss detection. The number of times the handshake packets have been retransmitted without receiving an ack. The number of times a tail loss probe has been sent without receiving an ack. The number of times an rto has been sent without receiving an ack. The last packet number sent prior to the first retransmission timeout. The time the most recent packet was sent. The packet number of the most recently sent packet. The largest packet number acknowledged in an ack frame. The most recent RTT measurement made when receiving an ack for a previously unacked packet. The smoothed RTT of the connection, computed as described in The RTT variance, computed as described in The largest delta between the largest acked retransmittable packet and a packet containing retransmittable frames before it’s declared lost. The reordering window as a fraction of max(smoothed_rtt, latest_rtt). The time at which the next packet will be considered lost based on early transmit or exceeding the reordering window in time. An association of packet numbers to information about them, including a number field indicating the packet number, a time field indicating the time a packet was sent, and a bytes field indicating the packet’s size. sent_packets is ordered by packet number, and packets remain in sent_packets until acknowledged or lost.

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 (UsingTimeLossDetection()) reordering_threshold = infinite time_reordering_fraction = kTimeReorderingFraction else: reordering_threshold = kReorderingThreshold time_reordering_fraction = infinite loss_time = 0 smoothed_rtt = 0 rttvar = 0 largest_sent_before_rto = 0 time_of_last_sent_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: packet_number: The packet number of the sent packet. is_ack_only: A boolean that indicates whether a packet only contains an ACK frame. If true, it is still expected an ack will be received for this packet, but it is not congestion controlled. sent_bytes: The number of bytes sent in the packet, not including UDP or IP overhead, but including QUIC framing overhead. Pseudocode for OnPacketSent follows:

OnPacketSent(packet_number, is_ack_only, sent_bytes): time_of_last_sent_packet = now largest_sent_packet = packet_number sent_packets[packet_number].packet_number = packet_number sent_packets[packet_number].time = now if !is_ack_only: OnPacketSentCC(sent_bytes) sent_packets[packet_number].bytes = sent_bytes SetLossDetectionAlarm()

When an ack 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 if (latest_rtt > ack.ack_delay): latest_rtt -= ack.delay UpdateRtt(latest_rtt) // Find all newly acked packets. for acked_packet in DetermineNewlyAckedPackets(): OnPacketAcked(acked_packet.packet_number) DetectLostPackets(ack.largest_acked_packet) SetLossDetectionAlarm() UpdateRtt(latest_rtt): // Based on {{RFC6298}}. if (smoothed_rtt == 0): smoothed_rtt = latest_rtt rttvar = latest_rtt / 2 else: rttvar = 3/4 * rttvar + 1/4 * abs(smoothed_rtt - latest_rtt) 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 packet number of the newly acked packet, and returns a list of packet numbers that are detected as lost. 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 Pseudocode for OnPacketAcked follows:

OnPacketAcked(acked_packet_number): OnPacketAckedCC(acked_packet_number) // If a packet sent prior to RTO was acked, then the RTO // was spurious. Otherwise, inform congestion control. if (rto_count > 0 && acked_packet_number > largest_sent_before_rto) OnRetransmissionTimeoutVerified() handshake_count = 0 tlp_count = 0 rto_count = 0 sent_packets.remove(acked_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.

The initial flight has no prior RTT sample. A client SHOULD remember the previous RTT it observed when resumption is attempted and use that for an initial RTT value. If no previous RTT is available, the initial RTT defaults to 100ms. Endpoints MUST retransmit handshake frames if not acknowledged within a time limit. This time limit will start as the largest of twice the RTT value and MinTLPTimeout. Each consecutive handshake retransmission doubles the time limit, until an acknowledgement is received. Handshake frames may be cancelled by handshake state transitions. In particular, all non-protected frames SHOULD be no longer be transmitted once packet protection is available. 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 0RTT packets.

Tail loss probes and retransmission timeouts 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.

Early retransmit is implemented with a 1/4 RTT timer. It is part of QUIC’s time based loss detection, but is always enabled, even when only packet reordering loss detection is enabled.

Pseudocode for SetLossDetectionAlarm follows:

SetLossDetectionAlarm(): if (retransmittable packets are not outstanding): loss_detection_alarm.cancel() return if (handshake packets are outstanding): // Handshake retransmission alarm. if (smoothed_rtt == 0): alarm_duration = 2 * kDefaultInitialRtt else: alarm_duration = 2 * smoothed_rtt alarm_duration = max(alarm_duration, kMinTLPTimeout) alarm_duration = alarm_duration * (2 ^ handshake_count) else if (loss_time != 0): // Early retransmit timer or time loss detection. alarm_duration = loss_time - now else if (tlp_count < kMaxTLPs): // Tail Loss Probe if (retransmittable_packets_outstanding == 1): alarm_duration = 1.5 * smoothed_rtt + kDelayedAckTimeout else: alarm_duration = kMinTLPTimeout alarm_duration = max(alarm_duration, 2 * smoothed_rtt) else: // RTO alarm alarm_duration = smoothed_rtt + 4 * rttvar alarm_duration = max(alarm_duration, kMinRTOTimeout) alarm_duration = alarm_duration * (2 ^ rto_count) loss_detection_alarm.set(now + 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. RetransmitAllHandshakePackets() 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 is acknowledged. DetectLostPackets is called every time an ack is received. If the loss detection alarm fires and the loss_time is set, the previous largest acked packet is supplied.

The receiver MUST ignore unprotected packets that ack protected packets. The receiver MUST trust protected acks for unprotected packets, however. Aside from this, loss detection for handshake packets when an ack is processed is identical to other packets.

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 (time_reordering_fraction != infinite): 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 = 9/8 * max(latest_rtt, smoothed_rtt) foreach (unacked < largest_acked.packet_number): time_since_sent = now() - unacked.time_sent packet_delta = largest_acked.packet_number - unacked.packet_number if (time_since_sent > delay_until_lost): lost_packets.insert(unacked) else if (packet_delta > reordering_threshold) 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) foreach (packet in lost_packets) sent_packets.remove(packet.packet_number)

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 congestion control to determine the congestion window and pacing rate. QUIC congestion control is specified in bytes due to finer control and the ease of appropriate byte counting.

QUIC begins every connection in slow start and exits slow start upon loss. QUIC re-enters slow start after a retransmission timeout. While in slow start, QUIC increases the congestion window by the number of acknowledged bytes 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. Because QUIC retransmits frames, not packets, it defines the end of recovery as all packets outstanding at the start of recovery being acknowledged or lost. This is slightly different from TCP’s definition of recovery ending when the lost packet that started recovery is acknowledged. During recovery, the congestion window is not increased or decreased. As such, multiple lost packets only decrease the congestion window once as long as they’re lost before exiting recovery. This causes QUIC to decrease the congestion window multiple times if retransmisions are lost, but limits the reduction to once per round trip.

If recovery sends a tail loss probe, no change is made to the congestion window or pacing rate. 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 or pacing rate until the next acknowledgement arrives. When an ack arrives, if packets prior to the first retransmission timeout are acknowledged, then the congestion window remains the same. If no packets prior to the first retransmission timeout are acknowledged, the retransmission timeout has been validated and the congestion window must be reduced to the minimum congestion window and slow start is begun.

The pacing rate is a function of the mode, the congestion window, and the smoothed rtt. Specifically, the pacing rate is 2 times the congestion window divided by the smoothed RTT during slow start and 1.25 times the congestion window divided by the smoothed RTT during slow start. In order to fairly compete with flows that are not pacing, it is recommended to not pace the first 10 sent packets when exiting quiescence.

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. The default max packet size used for calculating default and minimum congestion windows. Default limit on the amount of outstanding data in bytes. Default minimum congestion window. Reduction in congestion window when a new loss event is detected.

Variables required to implement the congestion control mechanisms are described in this section. The sum of the size in bytes of all sent packets that contain at least one retransmittable or PADDING frame, and have not been acked or declared lost. The size does not include IP or UDP overhead. Packets only containing ack frames do not count towards byte_in_flight to ensure congestion control does not impede congestion feedback. Maximum number of bytes in flight that may be sent. The largest packet number sent when QUIC detects a loss. When a larger packet is acknowledged, QUIC exits recovery. Slow start threshold in bytes. When the congestion window is below ssthresh, the mode is slow start and the window grows by the number of bytes acknowledged.

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

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.

OnPacketAckedCC(acked_packet): // Remove from bytes_in_flight. bytes_in_flight -= acked_packet.bytes if (acked_packet.packet_number < end_of_recovery): // Do not increase congestion window in recovery period. return if (congestion_window < ssthresh): // Slow start. congestion_window += acked_packets.bytes else: // Congestion avoidance. congestion_window += kDefaultMss * acked_packets.bytes / congestion_window

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 recovery epoch if the lost packet is larger // than the end of the previous recovery epoch. if (end_of_recovery < largest_lost_packet.packet_number): end_of_recovery = largest_sent_packet congestion_window *= kLossReductionFactor congestion_window = max(congestion_window, kMinimumWindow) ssthresh = congestion_window

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.

Google Mozilla

RFC Editor’s Note: Please remove this section prior to publication of a final version of this document.

Add more congestion control text (#776)

No significant changes.

No significant changes.

Integrate F-RTO (#544, #409) Add congestion control (#545, #395) Require connection abort if a skipped packet was acknowledged (#415) Simplify RTO calculations (#142, #417)

Overview added to loss detection Changes initial default RTT to 100ms Added time-based loss detection and fixes early retransmit Clarified loss recovery for handshake packets Fixed references and made TCP references informative

Improved description of constants and ACK behavior

Adopted as base for draft-ietf-quic-recovery Updated authors/editors list Added table of contents