Convergence of Congestion
Control from Retained StateThales Alenia Spacenicolas.kuhn.ietf@gmail.comOrangeemile.stephan@orange.comUniversity of AberdeenDepartment of EngineeringFraser Noble BuildingAberdeenAB24 3UEUKgorry@erg.abdn.ac.ukUniversity of AberdeenDepartment of EngineeringFraser Noble BuildingAberdeenAB24 3UEUKr.secchi@erg.abdn.ac.ukPrivate Octopus Inc.huitema@huitema.net
Transport
Internet Engineering Task ForceTransport, Congestion Control, QUICThis document specifies a cautious method for IETF transports that
enables fast startup of congestion control for a wide
range of connections.
It reuses a set of computed congestion control parameters that
are based on previously observed path characteristics between
the same pair of transport endpoints. These parameters
are saved, allowing them to be later used to modify the congestion control behavior
of a subsequent connection.It describes assumptions
and defines requirements for how a
sender utilizes these parameters to provide opportunities for a
connection to more rapidly get up to speed and rapidly utilize available
capacity. It discusses how use of Careful Resume impacts the capacity at a
shared network bottleneck and the safe response that is needed after any
indication that the new rate is inappropriate.All Internet transports are required to either
use a Congestion Control (CC) algorithm, or
to constrain their rate of transmission . In 2010,
a survey of alternative CC algorithms , noted that there
are challenges when a CC algorithm operates across an Internet path with a high and/or
varying Bandwidth-Delay Product (BDP). This mechanism targets a
solution for these challenges.A CC algorithm typically takes time to ramp-up the sending rate,
called the "Slow-Start phase", informally known as the time to "Get up
to speed". This defines a time in which a sender
intentionally uses less capacity than might be available, with the
intention to avoid or limit overshoot of the available capacity for the path.
This can increase queuing (latency or jitter) and/or
congestion packet loss for the flow. Any overshoot can have a
detrimental effect on other flows sharing a common bottleneck.
A sender can use a method that observes the rate of acknowledged data,
and seek to avoid an overshoot of the bottleneck capacity (e.g., Hystart++
).
In the extreme case, an overshoot can result in persistent congestion
with unwanted starvation of
other flows (i.e., preventing other flows
from successfully sharing the capacity at a common bottleneck).The present document specifies a CC mechanism,
called Careful Resume, which is expected to
reduce the time to complete a transfer
when the transfer sends significantly more data than allowed by the
Initial congestion Window (IW), and
where the BDP of the path is also significantly
more than the IW.
It introduces an alternative mechanism to select initial CC parameters,
that seek to more rapidly and safely grow the sending rate controlled by
the congestion window (CWND). CC algorithms that are rate-based can make
similar adjustments to their target sending rate.Careful Resume is based on temporal sharing (sometimes known as
caching) of a saved set of CC parameters that relate to previous observations
of the same path. The parameters include:
the saved_cwnd for the path and the minimum Round Trip Time (RTT). These
parameters are saved and used to modify the CC
behavior of a subsequent connection between the
same endpoints. Some congestion control algorithms may use other parameters.
For example, implementations using BBR also retain the
value of the bottleneck bandwidth required to reach the capacity available to
the flow (BBR.max_bw, see ).
When used with the QUIC transport, this provides transport services that resemble
those that could be implemented in TCP, using methods such as TCP Control Block (TCB)
caching.CC parameters are used by Careful Resume for three functions:
Information to confirm whether a saved path corresponds to
the current path.Information about the utilised path capacity to
set CC parameters.Information to check the CC parameters are not too old."Generally, implementations are advised
to be cautious when using saved CC parameters on a new path",
as stated in .
While this statement has been proposed in the context of QUIC standardization, this advice is appropriate for any IETF transport protocol.
Care is therefore needed
to assure safe use and to be robust to changes
in traffic patterns, network routing, and link/node conditions.
There are cases where using the saved parameters of a previous
connection is not appropriate (see ).Whilst a sender could take optimization decisions without considering
the receiver's preference, there are cases where a receiver
could have information that
is not available at the sender, or might benefit from
understanding that Careful Resume might be used.
In these cases, a receiver
could explicitly ask to enable or inhibit Careful Resume
when an application initiates a new connection. Examples where a receiver might request to inhibit use Careful Resume include:
a receiver that can predict the pattern of traffic
(e.g., insight into the volume of data to be sent,
the expected length of a connection, or the requested maximum transfer rate);a receiver with a local indication that a path/local
interface has changed since the CC parameters were saved;knowledge of the current hardware limitations at a receiver;a receiver that can predict additional capacity will be needed
for other concurrent or later flows
(i.e., prefers to activate Careful Resume for a different connection).A related document proposes an extension for QUIC that allows
sender-generated CC parameters to be stored at the receiver
.
This avoids the need for a
sender to retain transport state for each
receiver.
It also allows the receiver to express a preference for whether
a sender ought use Careful Resume.The CWND is one factor that limits the
sending rate of a transport protocol. Other mechanisms also constrain
the maxmimum sending rate. These include the sender pacing rate and the
receiver-advertised window (or flow credit), see .This section provides a set of examples where Careful Resume is expected to improve performance.
Either endpoint can assume the role of a
sender or a receiver. Careful Resume also supports a
bidirectional data transfer,
where both endpoints simultaneously send data
(e.g., remote execution of an application, or a
bidirectional video conference call).In one example, an application uses a series of connections over a path.
Without a new method,
each connection would need to individually
discover appropriate CC parameters, whereas Careful Resume allows the flow
to use a rate based on the previously observed CC parameters.In another example, an application connects after a disruption
had temporarily reduced the path
capacity. When the endpoint
returns to use the path using Careful Resume, the
sending rate can be based on the previously observed CC parameters.
There is particular benefit for
any path with an RTT that is much larger than typical
Internet paths.
In a specific example, an application connected via a satellite access network
could take 9 seconds to complete a 5.3 MB transfer
using standard CC, whereas a sender using Careful Resume
could be reduce this transfer time to 4 seconds. The time to complete a 1 MB transfer could
similarly be reduced by 62 % . This benefit is also
expected for other sizes of transfer and for different path
characteristics when a path has a large BDP.This subsection provides a brief summary of key terms and the
requirements language.The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
"SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and
"OPTIONAL" in this document are to be interpreted as described in
BCP 14
when, and only when, they appear in all
capitals, as shown here. The document uses language drawn
from a range of IETF RFCs.
The following terms are defined:
Careful Resume (CR): The method specified in this document to select
initial CC parameters and to more rapidly and safely
increase the initial sending rate.CC parameters: A set of saved congestion control parameters from
a observing the capacity of an established connection (see ).CWND: The congestion window, or equivalent CC variable limiting
the maximum sending rate;current_endpoint_token: The Endpoint Token of the current receiver;current_rtt: A sample measurement of the current RTT;Endpoint Token: A token identifying a path to a receiver ;flight_size: The current volume of unacknowledged data;jump_cwnd: The resumed CWND, used in the Unvalidated Phase.LifeTime: The time for which the saved CC parameters can be safely re-used.max_jump: The configured maximum jump_cwnd;PipeSize: A measure of the validated available capacity based on the acknowledged data;saved_cwnd: The preserved capacity derived from observation of a
previous connection (see );saved_endpoint_token: The Endpoint Token associated with a set of CC parameters;saved_rtt: The preserved minimum RTT (see ).Unvalidated Packet: A packet sent when the CWND has been increased
beyond the size normally permitted by the congestion control algorithm;
if such a packet is acknowledged, it contributes to the PipeSize, but
if congestion is detected, it triggers entry to the Safe Retreat Phase.This section defines a series of phases that the
congestion controller moves through as a connection
uses Careful Resume. Connect -> Reconnaissance --------------------> Normal
(Normal) | ^
v |
Unvalidated --------------------------+
| | |
| +--> Validating --------------+
| | |
| | |
+---------------+--> Safe Retreat ---+
]]>An established connection in the Normal Phase is permitted to start observing CC parameters.
The key phases of Careful are illustrated in Figure 1.
Examples of the transitions between phases
are provided in .An established connection in the Normal Phase, can save a set of CC parameters for the specific path
to the current endpoint. Each set of CC parameters includes the saved_endpoint_token and the
LifeTime (e.g., as a timestamp after which the parameters must not be used).Observing (saved_cwnd):
The saved_cwnd is a measure of the currently utilised capacity for the connection,
measured as the volume of bytes sent during an RTT. This could be computed
by measuring the volume of data acknowledged in one RTT.
If the measured CWND is less than four times the Initial Window (IW)
a sender can choose to not save the CC parameters, because the additional actions associated with
performing Careful Resume for a small CWND would not justify its use.Observing (saved_rtt): The minimum RTT is saved as the saved_RTT.Implementation notes are provided in . A sender enters the Reconnaissance Phase after connection setup.
In this phase, the CWND is initialised to the IW, and the sender transmits initial data.
The CWND MAY be increased using normal CC as each ACK
confirms delivery of previously unacknowledged data (i.e., the CC is unchanged).The phase seeks to determine if the path is consistent with
a previously observed path (saved as a set of CC parameters).
The following conditions need to be confirmed before the sender
enters the Reconnaissance Phase:Reconnaissance Phase (Endpoint change):
If the current_endpoint_token is not the same as one of the saved_endpoint_tokens,
the sender MUST enter the Normal Phase. (A difference in the Endpoint Token
indicates a the network path was different to one that was observed.)Reconnaissance Phase (Lifetime of saved CC parameters): The CC parameters are temporal.
If the LifeTime of the observed CC parameters is exceeded,
the CC parameters are not used and the sender enters the Normal Phase.The following actions are performed during the Reconnaissance Phase:Reconnaissance Phase (Confirming the RTT): During this phase,
a sender MUST record the minimum RTT for the current connection as the current_rtt.Reconnaissance Phase (Detected congestion): If the sender detects
congestion (e.g., packet loss or ECN-CE marking), the sender MUST enter the Normal Phase
to respond to the detected congestion.
Reconnaissance Phase (Using saved_cwnd):
Only one connection can use a specific set of saved CC parameters.
If another connection has already started to use the saved_cwnd, the sender
MUST enter the Normal Phase.Reconnaissance Phase (Path confirmed):
When a sender has confirmed the RTT and also has
received an acknowledgement for the initial data without reported congestion,
it MAY then enter the Unvalidated Phase.
This transition occurs when a sender has more data than permitted
by the current CWND.
If a sender is rate-limited ,
it might send insufficient data
to be able to validate transmission at the higher rate.
A sender is allowed to remain in the Reconnaissance Phase and to not
transition to the Unvalidated Phase until there is
more data in the transmission buffer
than can be sent using the current CWND.
In some implementations, the decision to enter the
Unvalidated Phase could need coordination with the management
of buffers in the interface to the upper layers.When a path is not confirmed, Careful Resume is not used
and the sender enters the Normal Phase.Implementation notes are provided in .The Unvalidated Phase is designed to enable the CWND
to more rapidly get up to speed by using paced transmission of a tenatively increased CWND.
The following conditions
need to be confirmed before the sender enters the Unvalidated Phase:Unvalidated Phase (Confirming the path on entry):
If the current_rtt is greater than or equal to
(saved_rtt / 2) or the current_rtt is less than or equal to (saved_rtt x 10)
(see ),
the sender MUST enter the Normal Phase
(see trigger rtt_not_validated in ).
The calculation
of a sending rate from a saved_cwnd
is directly impacted by the RTT, therefore a significant change in the RTT
is a strong indication that the previously observed CC
parameters are not be valid for the current path.On entry to the Unvalidated Phase, the sender:Unvalidated Phase (Initialising PipeSize):
The variable PipeSize if initialised to CWND on
entry to the Unvalidated Phase. This records the
CWND before the jump is applied.Unvalidated Phase (Setting the jump_cwnd):
To avoid starving other flows that could have either started
or increased their use of capacity after the Observation Phase,
the jump_cwnd MUST be no more than half of the saved_cwnd.
Hence, jump_cwnd is less than or equal to Min(max_jump,(saved_cwnd/2)). CWND = jump_cwnd. The following actions are performed during the Unvalidated Phase:Unvalidated Phase (Pacing transmission): All packet sent in the
Unvalidated Phase MUST use based on the current_rtt.Unvalidated Phase (Confirming the path during transmission):
If a sender determines that the previous CC parameters
are not valid (due to a detected path change),
the Safe Retreat Phase is entered.
(In the Unvalidated Phase, insufficient time has passed for
a sender to receive feedback validating the the jump in CWND.
Therefore, any detected congestion must have resulted from packets sent
before the Unvalidated Phase.)Unvalidated Phase (Tracking PipeSize):
The variable PipeSize is increased by the volume of data acknowledged
by each received ACK. (This indicates a previously unvalidated packet has been
succesfuly sent over the path.)Unvalidated Phase (Receiving acknowledgement for an unvalidated packet):
The sender enters the Validating Phase when an acknowledgement is
received for the first packet number (or higher) that was sent in the Unvalidated Phase
(see first_unvalidated_packet_acknowledged in ).When the flow is controlled using BBR, Careful Resume is implemented
by setting the pacing rate from the saved congestion control parameters,
with the following precautions:The flag "carefully-resuming" is
added to the BBR state, and initialized
to "False" when the BBR flow starts;Careful Resume is only activated if a BBR flow is in the Startup state;The probing rate is set to 1/2 of the bottleneck bandwidth in
the saved congestion control parameters.The sender starts the Unvalidated Phase at the beginning of a round,
and sets the "carefully-resuming" flags to "True";When the "carefully-resuming" flag is set, the sender sets the BBR pacing rate
to the larger of the nominal pacing rate (BBR.bw times BBRStartupPacingGain)
and the probing rate. The CWND is set to the largest of BBR.bw
and the probing rate, multiplied by BBR.rtt_min times
BBRStartupCwndGain;The "carefully-resuming" flag is reset to False two rounds after
it is set, i.e., after all the packets sent in the first round
of "carefully resuming"
have been received and acknowledged by the peer. At that stage (after the capacity has been validated),
the measured delivery rate is expected to reflect the probing rate.Implementation notes are provided in . The Validating Phase checks that all packets
sent in the Unvalidated Phase were received without inducing congestion.
The CWND remains unvalidated and the sender typically remains in this phase for one RTT.
On entry to the Validating Phase, the sender:Validating Phase (Check flight_size on entry):
On entry to the Validating Phase, if the flight_size is less
equal to the PipeSize, the Normal Phase is entered with the CWND
reset to the PipeSize. (The unvalidated part of the jump_cwnd was not utilised).Validating Phase (Limiting CWND on entry):
On entry to the Validating Phase, the CWND is set to the flight_size.During the Validating Phase, the sender performs the following actions:Validating Phase (Tracking PipeSize):
The PipeSize is increased by the volume of acknowledged data for each
received ACK that indicates a packet was
successfully sent over the path.Validating Phase (Updating CWND): The CWND is updated using the
normal rules for the current congestion controller, this typically
allows CWND to be increased for each received acknowledgement that
indicates a packet has been successfully sent across the path.Validating Phase (Congestion indication):
If a sender determines that congestion was experienced
(e.g., packet loss or ECN-CE marking), Careful Resume
enters the Safe Retreat Phase
(see trigger packet_loss and ECN_CE in ).Validating Phase (Receiving acknowledgement for all unvalidated packets):
The sender enters the Normal Phase when an acknowledgement is
received for the last packet number (or higher)
that was sent in the Unvalidated Phase
(see last_unvalidated_packet_acknowledged in ).When using BBR, validation is performed using the regular BBR rules
for exiting Startup. The measured delivery rate will reflect the
actual capacity of the network. If congestion was experienced and
packet losses were observed, BBR will exit the Startup state
and enter the Drain state while the "carefully-resuming" flag is still
True, which will trigger the "safe retreat" option of the Drain state.This phase is entered when the first loss/ECN-CE marking is
detected for an unvalidated packet.
It drains the path of other unvalidated packets.
(This trigger is the same as used by a QUIC sender to transition
from Slow-Start to Recovery .) On entry to the Safe Retreat Phase, the sender:Safe Retreat Phase (Removing saved information):
The set of saved CC parameters for
the path are deleted, to prevent these
from being used again by other flows.Safe Retreat Phase (Re-initializing CWND):
The CWND MUST be reduced to
no more than (PipeSize/2).
This avoids persistent starvation by allowing capacity for other flows to regain
their share of the total capacity.
The minimum CWND in QUIC is 2 packets (see: section 4.8).Safe Retreat Phase (QUIC recovery): When the CWND is reduced,
a QUIC sender can immediately send a single packet prior to the 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 .) In the Safe Retreat Phase, the sender performs the following actions:Safe Retreat Phase (Tracking PipeSize):
The sender continues to update
the PipeSize after processing each acknowledgement.
(The PipeSize is used to
reset the ssthresh when leaving this phase, it does
not modify CWND.)Safe Retreat Phase (Maintaining CWND):
The CWND MUST NOT be increased in the Safe Retreat Phase.Safe Retreat Phase (Acknowledgement of all unvalidated packets):
The sender enters Normal Phase
when the last packet (or a later packet) sent during the
Unvalidated Phase has been acknowledged, and if required
adjusts the ssthresh (see exit_recovery in ).
The value of ssthresh on leaving the Safe Retreat Phase
MUST NOT be more than the PipeSize.When using BBR, the Safe Retreat Phase is entered if the Drain
state is entered while the "carefully-resuming" flag is still
True, i.e., if less than 2 full rounds have elapsed after
the sender entered the Unvalidated Phase. The delivery rates measured in
these conditions are tainted, because packets sent during the attempt
are still queued at the bottleneck and may have "pushed out"
competing traffic. The delivery rates measured in Drain state
MUST be discarded if the "carefully-resuming" flag is set to True.
This flag is cleared upon exiting the Drain state.Implementation notes are provided in .Unacknowledged packets that were sent in the Unvalidated Phase
can be lost when there is congestion.
Loss recovery commences using the reduced CWND
that was set on entry to the Safe Retreat Phase.Loss Recovery (Receiving acknowledgement for all unvalidated packets):
The sender leaves the Safe Retreat Phase when
the last packet number (or a later packet) sent in the
Unvalidated Phase is acknowledged.
(Note that if the last packet number is not cumulatively acknowledged, then
additional packets might need to be retransmitted.) A sender that experiences a Retransmission Time Out (RTO) expiry
ceases to use Careful Resume.
The sender enters the Normal Phase. If using BBR, the normal
processing of packet losses will cause it to enter the
Drain state while the "carefully-resuming" flag is set to True,
which will force the Safe Retreat mode. As in loss recovery, data sent in the
Unvalidated Phase could be later acknowledged after an RTO event
(see ).In the Normal Phase, the sender transitions to using
the normal CC algorithm (e.g., in congestion avoidance if CWND is more than ssthresh).
(Note that when the sender did not use the entire jump_cwnd the CWND was reduced
on entering the Validating Phase.)Implementation notes are provided in . The Endpoint Token is an implementation-dependent token that allows a sender to identify its own
view of the network path being used to connect to a specific remote endpoint. This internal identifier is used
by Careful Resume
to match the current path with a set of CC parameters associated with a previously observed path.When computing the Endpoint Token, the sender includes information to identify
the path on which it sends, this needs to include:
a unique identifier for its own sending interface (e.g., a globally
assigned address/prefix or other local identifier);when multiple interfaces are in use,
this unique identifier should include an interface identifier
(e.g., an index value or a MAC address to associate the
endpoint with the interface used for sending); an identifier for the destination (e.g., a name or a
destination IP address used to connect to the receiver);The Endpoint Token could include other information such as the sender DSCP, the transport ports, a flow
label, etc and other information (e.g., including PvD information or
information relating to its public-facing IP address).
However, such additional information needs to be set consistently for a resumed connection
to the same remote endpoint.
Although additional information could improve the path
differentiation, it could also reduce the re-usability of the token for resumed connections.This section provides guidance for implementation and use.There are various approaches to measuring the capacity used by a connection.
Congestion controllers, such as CUBIC or Reno, can estimate the
capacity by utilizing the
CWND or flight_size. A different approach could
estimate the same parameters for a rate-based congestion
controller, such as BBR ,
or by observing the rate at which data is acknowledged by the remote endpoint.Implementations are required to calculate a saved_rtt, measuring
the minimum RTT while observing the capacity. For example, this could be the minimum
of a set RTT of measurements measured over the previous 5 minutes.Implementations are expected to include a
LifeTime parameter in the CC parameters that can be used to remove old CC parameters
when no longer needed, or the CC parameters are out of date.There are cases where the current CWND
does not reflect the path capacity. At the end of slow
start, the CWND can be significantly larger than
needed to fully utilize the path (i.e., a CWND
overshoot). It is inappropriate to use an
overshoot in the CWND as a basis for estimating the
capacity. In most cases, the CWND will converge to a stable
value after several more RTTs.
One mitigation could be to set the
saved_cwnd based on
the flight_size, or an averaged CWND.When a sender
is rate-limited, or in the RTT following a burst of
transmission, a sender typically transmits
less data than allowed by the CWND. Such observations could to be discounted when
estimating the saved_cwnd (e.g., when a previous
observation recorded a higher value.)In the Reconnaissance Phase, a sender initiates a connection
and starts sending initial data, while measuring the current_rtt.
The CC is not modified.
A sender therefore needs to limit the initial data,
sent in the first RTT of transmitted data,
to not more than the IW .
This transmission using the IW is
assumed to be a safe starting point for any path to avoid
adding excessive load to a potentially congested path.Careful Resume does not permit multiple concurrent reuse of
the saved CC parameters. When multiple new concurrent connections
are made to a server, each can have a valid saved_endpoint_token,
but the saved_cwnd can once (i.e., if two connections start
simultaneously they cannot both use the saved_cwnd to perform a jump).
This is to prevent a sender from performing multiple jumps in the CWND,
each individually based on the same saved_cwnd, and hence creating an
excessive aggregate load at the bottleneck.The method that is used to prevent re-use of the saved CC parameters
will depend upon the design of the server
(e.g., if all connections from a given client IP arrive at the
same server process, then the server process could use a hash table,
whereas when using some types of load
balancing, a distributed system might be needed to ensure this
invariant when the load balancing hashes
connections by 4-tuple and hence multiple connections from the same
client device are served by different server processes.Path characteristics can change over time for many reasons.
This can result in the previously observed CC parameters
becoming irrelevant. To help confirm the path, the sender compares the
saved_RTT with each of a
series of current_rtt samples.
If the current_rtt sample is less than a half of the
saved_RTT, this is regarded as too small, and is an indicator of a path change.
(This factor of two arises, because the rate should not exceed the observed rate when
the saved_cwnd was measured, because the jump_cwnd is calculated as half the
measured saved_cwnd.)If the current RTT is larger than saved_rtt (when the saved_cwnd was measured),
this results
in a proportionally lower resumed rate, because the transmission using Careful Resume
is paced based on the current_rtt (i.e., a larger RTT sample in the Unvalidated Phase
would reduce the paced sending rate ,and hence is still safe).
If the current_rtt is incorrectly measured as larger
than the actual path RTT, the sender will receive an ACK for an
unvalidated packet before it would have completed the Unvalidated Phase, Careful Resume
uses this
ACK to reset the CWND to reflect the flight_size, and the sender then enters
the Validating Phase.A current_rtt more than ten times the
saved_RTT is indicative of a path change.
(The value of ten was chosen to accommodate both increases in latency from buffering
on a path, and any variation between RTT samples).
A sender also verifies that the initial data was acknowledged.
(i.e., both coukd otherwise could be indicative of persistent congestion).A sender in Reconnaissance Phase reverts to the Normal Phase if congestion is detected.
Some transport protocols implement CC mechanisms that infer potential congestion
from an increase in the current_rtt.
In the Reconnaissance Phase, this indication can occur earlier than congestion
that is reported by
loss or by ECN marking. Designs need to consider if such an indication is
a suitable trigger to revert to the Normal Phase.This section considers the safety
for using saved CC parameters to tentatively update the CWND.
This is designed to mitigate the risk of
adding excessive congestion to an already congested path.A connection must not directly use the previously
saved_cwnd to directly initialize a new flow causing it to resume sending at the same
rate. The jump_cwnd must therefore be no more than half the previously saved_cwnd.The long-term use of the previously observed parameters is not appropriate,
a lifetime therefore needs to be specified during
which the saved CC parameters can be safely re-used. provides guidance on the implementation of
TCP Control Block Interdependence, but does not specify how long a saved
parameter can safely be reused. specifies a method for managing an
unvalidated CWND. This states:
"After a fixed period of time (the non-validated period (NVP)), the sender
adjusts the cwnd (Section 4.4.3). The NVP SHOULD NOT exceed five minutes."
Section 5 of discusses the rationale for
choosing that period.
However, RFC 7661 targets rate-limited connections using normal
CC. Careful Resume includes additional
mechanisms to avoid and mitigate the effects of overshoot, and therefore
this can be used to justify a longer lifetime of the saved_cwnd using
Careful Resume. A QUIC sender must avoid sending a burst of packets greater than IW as a result of a
step-increase in the CWND. This is consistent with ,
. Pacing packets as a function of
the current_rtt, rather than the saved_RTT provides an additional safety during the
Unvalidated Phase, because it avoids a smaller saved_RTT inflating the sending rate.
Pacing also places a limitation on the minimum acceptable current_RTT
to avoid sending at a rate higher than was previously observed. The following example provides a relevant pacing rhythm using the RTT and the
saved_cwnd. The Inter-packet Transmission Time (ITT) is determined
by using the current Maximum Message Size (MMS),
the saved_cwnd and the current_RTT. A safety
margin can be configured to avoid sending more than a maximum
(max_jump):
jump_cwnd = Min(max_jump,saved_cwnd/2)ITT = (current_RTT x MMS)/jump_cwndThis follows the idea presented in ,
and
.
Other sender mitigations have also been suggested to
avoid line-rate bursts (e.g., ).Careful Resume has been designed to be robust to
changes in network conditions
due to variations in the forwarding path, such as reconfiguration of
equipment, or changes in the link conditions. This is mitigated
by path confirmation.Careful Resume has been designed to be robust to changes
in network traffic, including the
arrival of new flows that compete for capacity at a shared bottleneck.
This is mitigated by jumping to no more than a half of
the saved_cwnd and by using pacing.Careful Resume has been designed to avoid unduly suppressing flows
that used the capacity since the available capacity was measured. This is further
mitigated by bounding the duration of the Unvalidated Phase (and the following
Validating Phase), and the conservative design of the Safe Retreat Phase.The purpose of the Validating Phase is to trigger an
entry to the Safe Retreat Phase if the capacity is not validated.When a sender completes the Unvalidated Phase, either by sending a jump_cwnd of data
or after one RTT, it ceases to use the unvalidated CWND. That is, CWND is reset
to the flight_size, and the sender awaits reception of ACKs to validate the
use of this capacity. New packets are sent when previously
sent data is newly acknowledged.
The CWND is increased during the Validating Phase,
based on received ACKs. This allows new data to be sent,
but this does not have any final impact on the CWND
if congestion is subsequently detected.This section considers the safety
after congestion has been detected for unvalidated packets.The Safe Retreat Phase sets a safe CWND value to drain any unvalidated packets
from the path after a packet loss has been detected or ACKs that indicate sent
packets were ECN CE-marked. The CC parameters that were used are invalid,
and are removed.The Safe Retreat reaction differs from a traditional
reaction to detected congestion, because
a jump_cwnd can result in a significantly higher rate than would be allowed by
Slow-Start. This jump could aggressively feed a congested bottleneck,
resulting in overshoot where a disproportionate number of packets
from existing flows are displaced from
the buffer at the congested bottleneck.
For this reason, a sender in the Safe Retreat Phase needs to react to detected congestion by
reducing CWND significantly
below the saved_cwnd.During loss recovery, a receiver can cumulatively acknowledge
data that was previously sent in the Unvalidated Phase in addition
to acknowledging successful retransmission of data.
describes how to appropriately
account for such ACKs.
ACKS received
for unvalidated packets are tracked to
measure the maximum available capacity, called the PipeSize
(The first unvalidated packet can be determined by
recording the sequence number
of the first packet sent in the Unvalidated Phase.)
This calculated PipeSize is later used to reset the ssthresh.
However, note that this is not a safe measure of the currently
available share of the capacity whenever
there was also a significant overshoot at the bottleneck,
and must not be used to reinitialise the CWND.The Proportional Rate Reduction (PRR)
assumes that it is safe to reduce
the rate gradually when in congestion avoidance.
PRR is therefore not appropriate
when there might be significant overshoot in the use of the capacity, which can
be the case when the Safe Retreat Phase is entered.The recovery from loss depends on the design of a transport protocol.
A TCP or SCTP sender is required to retransmit
all lost data .
For QUIC and DCCP, the need for loss recovery
depends on the sender policy for retransmission.
On entry to the Safe Retreat Phase, the CWND can be
significantly reduced, when there was multiple loss,
a sender recovering all lost data could take multiple RTTs to complete.After using Careful Resume, the CC controller returns to the Normal Phase.
The implementation details for different transports depend on the
design of the transport.
In the Normal Phase, a sender is permitted to start
Observing the capacity of the path.The CWND is one factor that limits the
sending rate of a sender. Other mechanisms can also constrain
the maximum sending rate of a transport protocol.
A transport protocol might need to update these mechanisms
to fully utilise the CWND made available by Careful Resume:
A TCP sender is limited by the receiver window (rwnd).
Unless configured at a receiver, the rwnd constrains the rate
of increase for a connection and reduces the benefit of Careful Resume.QUIC this includes flow control mechanisms and mechanisms to prevent amplification
attacks. In particular, a QUIC receiver might need to issue proactive
MAX_DATA frames to increase the flow control limits of a connection
that is started when using Careful Resume to gain the expected benefit.This section provides definitions that enable a
Careful Resume implementation
to generate qlog events when using QUIC.
It introduces an event to report the current phase of a sender,
and an associated description.The event and data structure definitions in this section are
expressed in the Concise Data Definition Language (CDDL)
and its
extensions described in .
The current convention is to use long names for variables.
For example, "CWND" is expanded as "congestion_window"
and "saved_cwnd" is expanded as "saved_congestion_window".Importance: ExtraWhen the CC algorithm changes the Careful Resume Phase
described in of this specification.Definition:The authors would like to thank John Border, Gabriel Montenegro, Patrick McManus,
Ian Swett, Igor Lubashev, Robin Marx, Roland Bless, Franklin Simo, Kazuho Oku, Tong,
Ana Custura, Neal Cardwell, and Joerg Deutschmann for
their fruitful comments on earlier versions of this document.The authors would like to particularly thank Tom Jones for co-authoring
several previous versions of this document.
Ana Custura and Robin Marx developed the qlog support.No current parameters are required to be registered by IANA.This document does not exhibit specific security considerations.
Security considerations for the
interactions with the receiver are discussed in
.Issues in TCP Slow-Start Restart After IdleISIISIISIFeedback from using QUIC's 0-RTT-BDP extension over SATCOM
public accessGoogle QUIC performance over a public SATCOM accessHalfback: Running Short Flows Quickly and Safely The table below is provided to illustrate the operation of Careful Resume.
This table is informative, please refer to the body of the document
for the normative specification. The description is based on a Normal
CC that uses Reno or Cubic. The PipeSize tracks the validated CWND.PS) |CWND=(PS/2) |
|entry:| | |jump_cwnd |{CWND=FS} | |
| | | |=saved_cwnd |else | |
| | | |/2; |{CWND=PS; | |
| | | |CWND |enter |
| | | |=jump_cwnd |Normal} | |
+------+---------+---------+------------+-----------+------------+
|CWND: |When in |CWND |CWND is not |CWND can |CWND is not |
| |observe, |increases|increased |increase |increased |
| |measure |using SS | |using SS | |
| |saved | | | | |
| |_cwnd | | | | |
+------+---------+---------+------------+-----------+------------+
|PS: | - | - | PS+=ACked |
+------+---------+---------+------------+-----------+------------+
|RTT: |Measure |Measure | - | - | - |
| |saved_rtt|current | | | |
| | |_rtt | | | |
+------+---------+---------+------------+-----------+------------+
|If |Normal |Normal | Enter | - |
|loss |CC |CC; | Safe | |
|or | |CR is not| Retreat | |
|ECNCE:| |allowed | | |
+------+---------+---------+------------+-----------+------------+
|Next |Observing|If ( |If (FS=CWND |If (ACK |If (ACK |
|Phase:|(as |FS=CWND, |or >1 RTT |>= last |>= last |
| |needed) |Lifetime,|has passed |unvalidated|unvalidated |
| | |and RTT |or ACK for |packet), |packet), |
| | |confirmed|>= last |enter |{ssthresh=PS|
| | |), enter |unvalidated |Normal |and enter |
| | |Unvalidat|packet), | |Normal} |
| | |ing else |enter | | |
| | |enter |Validating | | |
| | |Normal | | | |
+------+---------+---------+------------+-----------+------------+
]]>The following abbreviations are used
SS = Slow-Start FS = flight_size; PS = PipeSize; ACK = acknowledgement.
The PipeSize tracks the validated part of the cwnd. It is set to the CWND
on entry to the Unvalidated Phase and
is updated as each additional packet is acknowledged.Note:
For an implementation that keeps track of transmitted data in terms of packets:
In the Unvalidated Phase, the first unvalidated packet corresponds to the highest sent packet recorded on entry to this phase.
In the Validating Phase and Safe Retreat Phase, this corresponds to the last unvalidated packet. It is also the highest sent packet number recorded on entry to this phase.The remaining subsections provide informative examples of use.Note: Although the QLOG variables are expressed in bytes,
to simplify the description, these examples
are described in term of packet numbers.In the first example of using Careful Resume,
the sender starts by sending IW packets, assumed to be 10 packets, in the Reconnaissance Phase,
and then continues in a subsequent RTT to send more packets until the sender
becomes CWND-limited (i.e., flight_size = CWND).The sender in the Reconaissance Phase then
confirms the RTT and other conditions for using Careful Resume.
In this example, this is confirmed when the sender has 29 packets in flight.The sender then enters the Unvalidated Phase.
(This path confirmation could have happened earlier if data had been available to send.) The
sender initialises the PipeSize to the CWND
(at this time this is the same as the flight_size, i.e., 29 packets)
and then sets the CWND to 150 packets (based upon half of the previously
observed saved_cwnd of 300 packets).The sender now sends 121 unvalidated packets (the unused portion of the current CWND).
Each time a packet is sent, the sender checks whether 1 RTT has passed since entering the
Unvalidated Phase (otherwise, the Validating Phase is entered). This check triggers only
for cases where the sender is rate-limited, see the following example.The PipeSize increases after each ACK is received.When the first unvalidated packet is acknowledged (packet number 30) the sender
enters the Validating Phase. (This transition would also occur if the flight_size increased to equal CWND.)
During this phase, the CWND can be increased for each ACK that acknowledges an unvalidated packet, because
this indicates that the packet was indeed validated.When an ACK is received for the last packet that was sent in the Unvalidated Phase,
the sender completes using Careful Resume. It then enters the Normal Phase. If CWND is less than ssthresh,
a Reno or Cubic sender in the Normal Phase is permitted to use Slow-Start to grow the CWND towards
the ssthresh, and will then enter congestion avoidance.A rate-limited sender will not fully utilize the available CWND when using Careful Resume,
and CWND is therefore reset on entry to the Validating Phase, as described below.The sender starts by sending IW packets (10) in the Reconnaissance Phase.
It commences as described in the first example, transitioning to the Unvalidated Phase.
This sets the CWND to 150 packets, and the PipeSize to the flight_size (i.e., 29 packets).The sender then becomes rate-limited because it only sends 50 unvalidated packets. After about one RTT (detected by using local timestamps or by receiving an ACK for the first
unvalidated packet), the sender will still not have fully used the CWND. It then enters
the Validating Phase and resets the CWND to the current flight_size, (i.e., 50 packets).
During this phase, the CWND can be increased for each received ACK
that validates reception of an unvalidated packet.
The PipeSize also increases with each ACK received, to reflect the discovered capacity.When an ACK is received for the last packet sent in the Unvalidated Phase,
the sender has completed using Careful Resume. It then enters the Normal Phase,
as in the example with no loss.When a packet is lost in the Reconnaissance Phase, the sender will enter the Normal Phase
and recovers this using the normal method. (There is no change to the CC method, because
the sender has discovered a potential capacity limit and is not
allowed to continue to use Careful Resume.)As in the first example, the sender enters the Unvalidated Phase and sets the CWND to 150 packets
with the PipeSize initialized to the flight_size (i.e., 29 packets).The sender now sends 121 unvalidated packets (the remaining unused CWND).
This example considers the case when one of the unvalidated packet is lost, which
we choose to be packet 64 (the 35th packet sent in the Unvalidated Phase).ACKs confirm the first 34 unvalidated packets are received without loss.
The PipeSize at this point is equal to 63 (29 + 34) packets.The loss is then detected (by a timer or by receiving three ACKs that do not cover packet number 35),
the sender then enters the Safe Retreat Phase because the window was not validated.
The PipeSize at this point is equal to 66 (29 + 34) packets. Assuming that the IW was 10 packets,
the CWND is reset to Max(10,PS/2) = Max(10,66/2) = 33 packets.
This CWND is used during the Safe Retreat Phase, because congestion was detected and the
sender still does not yet know if the remaining unvalidated packets will be
successfully acknowledged. A conservative CWND calculation ensures the sender drains
the path after this potentially severe congestion event.
There is no further increase in CWND in this phase.The sender continues to receive ACKs for the remaining 86 (121-35) unvalidated packets.
Recall that the 35th unvalidated packet was lost and had packet number 64 (29+35).
The PipeSize tracks the
capacity discovered by acknowledgments for the unvalidated packets and
continues to be further increased for each received ACK acknowledges new data.
Although the PipeSize cannot be used to safety initialise the CWND (because it was measured when the sender
had aggressively created overload), the estimated PipeSize
(which, in this case, is 121-1 = 120 packets) can be used to set the ssthresh on exit
from Safe Retreat, since it does indicate an upper limit to the current capacity.At the point where all packets sent in the Unvalidated Phase have been either acknowledged
or have been declared lost, the sender updates ssthresh and enters the Normal Phase.
Because CWND will now now be less than ssthresh, a sender in the Normal Phase is permitted to use
Slow-Start to grow the CWND towards the ssthresh,
after which it will enter congestion avoidance.Previous individual submissions were discussed in TSVWG and QUIC.
WG -00 included clarifications and restructuring to form the 1st WG draft.WG -01 included review comments and suggestions from John Border,
and follows the setting of the TSVWG milestone
with an intended status of "Proposed Standard".WG -02 includes steps to complete the spec. In particular, consideration of rate-limited
senders; selection of reasoned parameters; specification of the Safe Retreat Phase; and
improvements to the consistency throughout. Added the Validating Phase.WG -03, explain entry to Validating Phase, editorial tidy.WG -04, update based on review comments from Kazuho Oku.WG-05, update based on review comments from Neal Cardwell. WG feedback from IETF-118.
Reviewed the requirements v. guidelines; clarified that CC is not changed in recon., but
the recon. info is used to steer the next phase; clarified saved_cwnd can be computed
from ACK rate; use jump once; that real server platforms are complex.
Clarified lifetime for saved CC params. Incorporates comments from Tong.WG-06, SR updated following Hackathon comments from Kazuho Oku, and rework of use of PipeSize.
Added an informative summary of actions, on suggestion by Tong.
Added examples based on text by Ana Custura. WG-07, Use "rate-limited" uniformly instead of application and data limited.Updated to exit early when the unvalidated CWND not utilised,
detected in tests by Q Misell. Change
pipe_size to be PipeSize.WG-08, Updated CDDL, and made constraints to Observing into guidance, they say what
makes sense - but do not need to be followed for conformance. Updated table in the appendix to align
with text.WG-09, Cleaning text to separate guidelines and specification
and adjust wording to improve clarity based on questions received during implementation.
WG-10, CH developed text to explain expected operation with BBR.
This also fixed some typos introduced in previous edits. Fix XML and fix CDDL bugs for submission.