Internet Engineering Task Force (IETF) Y. Cheng
Request for Comments: 8985 N. Cardwell
Category: Standards Track N. Dukkipati
ISSN: 2070-1721 P. Jha
Google, Inc.
February 2021
The RACK-TLP Loss Detection Algorithm for TCP
Abstract
This document presents the RACK-TLP loss detection algorithm for TCP.
RACK-TLP uses per-segment transmit timestamps and selective
acknowledgments (SACKs) and has two parts. Recent Acknowledgment
(RACK) starts fast recovery quickly using time-based inferences
derived from acknowledgment (ACK) feedback, and Tail Loss Probe (TLP)
leverages RACK and sends a probe packet to trigger ACK feedback to
avoid retransmission timeout (RTO) events. Compared to the widely
used duplicate acknowledgment (DupAck) threshold approach, RACK-TLP
detects losses more efficiently when there are application-limited
flights of data, lost retransmissions, or data packet reordering
events. It is intended to be an alternative to the DupAck threshold
approach.
Status of This Memo
This is an Internet Standards Track document.
This document is a product of the Internet Engineering Task Force
(IETF). It represents the consensus of the IETF community. It has
received public review and has been approved for publication by the
Internet Engineering Steering Group (IESG). Further information on
Internet Standards is available in Section 2 of RFC 7841.
Information about the current status of this document, any errata,
and how to provide feedback on it may be obtained at
https://www.rfc-editor.org/info/rfc8985.
Copyright Notice
Copyright (c) 2021 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.
Table of Contents
1. Introduction
1.1. Background
1.2. Motivation
2. Terminology
3. RACK-TLP High-Level Design
3.1. RACK: Time-Based Loss Inferences from ACKs
3.2. TLP: Sending One Segment to Probe Losses Quickly with RACK
3.3. RACK-TLP: Reordering Resilience with a Time Threshold
3.3.1. Reordering Design Rationale
3.3.2. Reordering Window Adaptation
3.4. An Example of RACK-TLP in Action: Fast Recovery
3.5. An Example of RACK-TLP in Action: RTO
3.6. Design Summary
4. Requirements
5. Definitions
5.1. Terms
5.2. Per-Segment Variables
5.3. Per-Connection Variables
5.4. Per-Connection Timers
6. RACK Algorithm Details
6.1. Upon Transmitting a Data Segment
6.2. Upon Receiving an ACK
6.3. Upon RTO Expiration
7. TLP Algorithm Details
7.1. Initializing State
7.2. Scheduling a Loss Probe
7.3. Sending a Loss Probe upon PTO Expiration
7.4. Detecting Losses Using the ACK of the Loss Probe
7.4.1. General Case: Detecting Packet Losses Using RACK
7.4.2. Special Case: Detecting a Single Loss Repaired by the
Loss Probe
8. Managing RACK-TLP Timers
9. Discussion
9.1. Advantages and Disadvantages
9.2. Relationships with Other Loss Recovery Algorithms
9.3. Interaction with Congestion Control
9.4. TLP Recovery Detection with Delayed ACKs
9.5. RACK-TLP for Other Transport Protocols
10. Security Considerations
11. IANA Considerations
12. References
12.1. Normative References
12.2. Informative References
Acknowledgments
Authors' Addresses
1. Introduction
This document presents RACK-TLP, a TCP loss detection algorithm that
improves upon the widely implemented duplicate acknowledgment
(DupAck) counting approach described in [RFC5681] and [RFC6675]; it
is RECOMMENDED as an alternative to that earlier approach. RACK-TLP
has two parts. Recent Acknowledgment (RACK) detects losses quickly
using time-based inferences derived from ACK feedback. Tail Loss
Probe (TLP) triggers ACK feedback by quickly sending a probe segment
to avoid retransmission timeout (RTO) events.
1.1. Background
In traditional TCP loss recovery algorithms [RFC5681] [RFC6675], a
sender starts fast recovery when the number of DupAcks received
reaches a threshold (DupThresh) that defaults to 3 (this approach is
referred to as "DupAck counting" in the rest of the document). The
sender also halves the congestion window during the recovery. The
rationale behind the partial window reduction is that congestion does
not seem severe since ACK clocking is still maintained. The time
elapsed in fast recovery can be just one round trip, e.g., if the
sender uses SACK-based recovery [RFC6675] and the number of lost
segments is small.
If fast recovery is not triggered or is triggered but fails to repair
all the losses, then the sender resorts to RTO recovery. The RTO
timer interval is conservatively the smoothed RTT (SRTT) plus four
times the RTT variation, and is lower bounded to 1 second [RFC6298].
Upon RTO timer expiration, the sender retransmits the first
unacknowledged segment and resets the congestion window to the loss
window value (by default, 1 full-sized segment [RFC5681]). The
rationale behind the congestion window reset is that an entire flight
of data and the ACK clock were lost, so this deserves a cautious
response. The sender then retransmits the rest of the data following
the slow start algorithm [RFC5681]. The time elapsed in RTO recovery
is one RTO interval plus the number of round trips needed to repair
all the losses.
1.2. Motivation
Fast recovery is the preferred form of loss recovery because it can
potentially recover all losses in the timescale of a single round
trip, with only a fractional congestion window reduction. RTO
recovery and congestion window reset should ideally be the last
resort and should ideally be used only when the entire flight is
lost. However, in addition to losing an entire flight of data, the
following situations can unnecessarily resort to RTO recovery with
traditional TCP loss recovery algorithms [RFC5681] [RFC6675]:
1. Packet drops for short flows or at the end of an application data
flight. When the sender is limited by the application (e.g.,
structured request/response traffic), segments lost at the end of
the application data transfer often can only be recovered by RTO.
Consider an example where only the last segment in a flight of
100 segments is lost. Lacking any DupAck, the sender RTO
expires, reduces the congestion window to 1, and raises the
congestion window to just 2 after the loss repair is
acknowledged. In contrast, any single segment loss occurring
between the first and the 97th segment would result in fast
recovery, which would only cut the window in half.
2. Lost retransmissions. Heavy congestion or traffic policers can
cause retransmissions to be lost. Lost retransmissions cause a
resort to RTO recovery since DupAck counting does not detect the
loss of the retransmissions. Then the slow start after RTO
recovery could cause burst losses again, which severely degrades
performance [POLICER16].
3. Packet reordering. In this document, "reordering" refers to the
events where segments are delivered at the TCP receiver in a
chronological order different from their chronological
transmission order. Link-layer protocols (e.g., 802.11 block
ACK), link bonding, or routers' internal load balancing (e.g.,
ECMP) can deliver TCP segments out of order. The degree of such
reordering is usually within the order of the path round-trip
time. If the reordering degree is beyond DupThresh, DupAck
counting can cause a spurious fast recovery and unnecessary
congestion window reduction. To mitigate the issue, Non-
Congestion Robustness (NCR) for TCP [RFC4653] increases the
DupThresh from the current fixed value of three duplicate ACKs
[RFC5681] to approximate a congestion window of data having left
the network.
2. Terminology
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 [RFC2119] [RFC8174] when, and only when, they appear in all
capitals, as shown here.
3. RACK-TLP High-Level Design
RACK-TLP allows senders to recover losses more effectively in all
three scenarios described in the previous section. There are two
design principles behind RACK-TLP. The first principle is to detect
losses via ACK events as much as possible, to repair losses at round-
trip timescales. The second principle is to gently probe the network
to solicit additional ACK feedback, to avoid RTO expiration and
subsequent congestion window reset. At a high level, the two
principles are implemented in RACK and TLP, respectively.
3.1. RACK: Time-Based Loss Inferences from ACKs
The rationale behind RACK is that if a segment is delivered out of
order, then the segments sent chronologically before that were either
lost or reordered. This concept is not fundamentally different from
those described in [RFC5681], [RFC6675], or [FACK]. RACK's key
innovation is using per-segment transmission timestamps and widely
deployed SACK [RFC2018] options to conduct time-based inferences
instead of inferring losses by counting ACKs or SACKed sequences.
Time-based inferences are more robust than DupAck counting approaches
because they do not depend on flight size and thus are effective for
application-limited traffic.
Conceptually, RACK keeps a virtual timer for every data segment sent
(including retransmissions). Each timer expires dynamically based on
the latest RTT measurements plus an additional delay budget to
accommodate potential packet reordering (called the "reordering
window"). When a segment's timer expires, RACK marks the
corresponding segment as lost for retransmission.
In reality, as an algorithm, RACK does not arm a timer for every
segment sent because it's not necessary. Instead, the sender records
the most recent transmission time of every data segment sent,
including retransmissions. For each ACK received, the sender
calculates the latest RTT measurement (if eligible) and adjusts the
expiration time of every segment sent but not yet delivered. If a
segment has expired, RACK marks it as lost.
Since the time-based logic of RACK applies equally to retransmissions
and original transmissions, it can detect lost retransmissions as
well. If a segment has been retransmitted but its most recent
(re)transmission timestamp has expired, then, after a reordering
window, it's marked as lost.
3.2. TLP: Sending One Segment to Probe Losses Quickly with RACK
RACK infers losses from ACK feedback; however, in some cases, ACKs
are sparse, particularly when the inflight is small or when the
losses are high. In some challenging cases, the last few segments in
a flight are lost. With the operations described in [RFC5681] or
[RFC6675], the sender's RTO would expire and reset the congestion
window when, in reality, most of the flight has been delivered.
Consider an example where a sender with a large congestion window
transmits 100 new data segments after an application write and only
the last three segments are lost. Without RACK-TLP, the RTO expires,
the sender retransmits the first unacknowledged segment, and the
congestion window slow starts from 1. After all the retransmits are
acknowledged, the congestion window is increased to 4. The total
delivery time for this application transfer is three RTTs plus one
RTO, a steep cost given that only a tiny fraction of the flight was
lost. If instead the losses had occurred three segments sooner in
the flight, then fast recovery would have recovered all losses within
one round trip and would have avoided resetting the congestion
window.
Fast recovery would be preferable in such scenarios; TLP is designed
to trigger the feedback RACK needed to enable that. After the last
(100th) segment was originally sent, TLP sends the next available
(new) segment or retransmits the last (highest-sequenced) segment in
two round trips to probe the network, hence the name "Tail Loss
Probe". The successful delivery of the probe would solicit an ACK.
RACK uses this ACK to detect that the 98th and 99th segments were
lost, trigger fast recovery, and retransmit both successfully. The
total recovery time is four RTTs, and the congestion window is only
partially reduced instead of being fully reset. If the probe was
also lost, then the sender would invoke RTO recovery, resetting the
congestion window.
3.3. RACK-TLP: Reordering Resilience with a Time Threshold
3.3.1. Reordering Design Rationale
Upon receiving an ACK indicating a SACKed segment, a sender cannot
tell immediately whether that was a result of reordering or loss. It
can only distinguish between the two in hindsight if the missing
sequence ranges are filled in later without retransmission. Thus, a
loss detection algorithm needs to budget some wait time -- a
reordering window -- to try to disambiguate packet reordering from
packet loss.
The reordering window in the DupAck counting approach is implicitly
defined as the elapsed time to receive DupThresh SACKed segments or
duplicate acknowledgments. This approach is effective if the network
reordering degree (in sequence distance) is smaller than DupThresh
and at least DupThresh segments after the loss is acknowledged. For
cases where the reordering degree is larger than the default
DupThresh of 3 packets, one alternative is to dynamically adapt
DupThresh based on the FlightSize (e.g., the sender adjusts the
DupThresh to half of the FlightSize). However, this does not work
well with the following two types of reordering:
1. Application-limited flights where the last non-full-sized segment
is delivered first and then the remaining full-sized segments in
the flight are delivered in order. This reordering pattern can
occur when segments traverse parallel forwarding paths. In such
scenarios, the degree of reordering in packet distance is one
segment less than the flight size.
2. A flight of segments that are delivered partially out of order.
One cause for this pattern is wireless link-layer retransmissions
with an inadequate reordering buffer at the receiver. In such
scenarios, the wireless sender sends the data packets in order
initially, but some are lost and then recovered by link-layer
retransmissions; the wireless receiver delivers the TCP data
packets in the order they are received due to the inadequate
reordering buffer. The random wireless transmission errors in
such scenarios cause the reordering degree, expressed in packet
distance, to have highly variable values up to the flight size.
In the above two cases, the degree of reordering in packet distance
is highly variable. This makes the DupAck counting approach
ineffective, including dynamic adaptation variants as in [RFC4653].
Instead, the degree of reordering in time difference in such cases is
usually within a single round-trip time. This is because the packets
either traverse disjoint paths with similar propagation delays or are
repaired quickly by the local access technology. Hence, using a time
threshold instead of a packet threshold strikes a middle ground,
allowing a bounded degree of reordering resilience while still
allowing fast recovery. This is the rationale behind the RACK-TLP
reordering resilience design.
Specifically, RACK-TLP introduces a new dynamic reordering window
parameter in time units, and the sender considers a data segment S
lost if both of these conditions are met:
1. Another data segment sent later than S has been delivered.
2. S has not been delivered after the estimated round-trip time plus
the reordering window.
Note that condition (1) implies at least one round trip of time has
elapsed since S has been sent.
3.3.2. Reordering Window Adaptation
The RACK reordering window adapts to the measured duration of
reordering events within reasonable and specific bounds to
disincentivize excessive reordering. More specifically, the sender
sets the reordering window as follows:
1. The reordering window SHOULD be set to zero if no reordering has
been observed on the connection so far, and either (a) three
segments have been SACKed since the last recovery or (b) the
sender is already in fast or RTO recovery. Otherwise, the
reordering window SHOULD start from a small fraction of the
round-trip time or zero if no round-trip time estimate is
available.
2. The RACK reordering window SHOULD adaptively increase (using the
algorithm in "Step 4: Update RACK reordering window" below) if
the sender receives a Duplicate Selective Acknowledgment (DSACK)
option [RFC2883]. Receiving a DSACK suggests the sender made a
spurious retransmission, which may have been due to the
reordering window being too small.
3. The RACK reordering window MUST be bounded, and this bound SHOULD
be SRTT.
Rules 2 and 3 are required to adapt to reordering caused by dynamics
such as the prolonged link-layer loss recovery episodes described
earlier. Each increase in the reordering window requires a new round
trip where the sender receives a DSACK; thus, depending on the extent
of reordering, it may take multiple round trips to fully adapt.
For short flows, the low initial reordering window helps recover
losses quickly, at the risk of spurious retransmissions. The
rationale is that spurious retransmissions for short flows are not
expected to produce excessive additional network traffic. For long
flows, the design tolerates reordering within a round trip. This
handles reordering in small timescales (reordering within the round-
trip time of the shortest path).
However, the fact that the initial reordering window is low and the
reordering window's adaptive growth is bounded means that there will
continue to be a cost to reordering that disincentivizes excessive
reordering.
3.4. An Example of RACK-TLP in Action: Fast Recovery
The following example in Figure 1 illustrates the RACK-TLP algorithm
in action:
Event TCP DATA SENDER TCP DATA RECEIVER
_____ ____________________________________________________________
1. Send P0, P1, P2, P3 -->
[P1, P2, P3 dropped by network]
2. <-- Receive P0, ACK P0
3a. 2RTTs after (2), TLP timer fires
3b. TLP: retransmits P3 -->
4. <-- Receive P3, SACK P3
5a. Receive SACK for P3
5b. RACK: marks P1, P2 lost
5c. Retransmit P1, P2 -->
[P1 retransmission dropped by network]
6. <-- Receive P2, SACK P2 & P3
7a. RACK: marks P1 retransmission lost
7b. Retransmit P1 -->
8. <-- Receive P1, ACK P3
Figure 1: RACK-TLP Protocol Example
Figure 1 illustrates a sender sending four segments (P0, P1, P2, P3)
and losing the last three segments. After two round trips, TLP sends
a loss probe, retransmitting the last segment, P3, to solicit SACK
feedback and restore the ACK clock (Event 3). The delivery of P3
enables RACK to infer (Event 5b) that P1 and P2 were likely lost
because they were sent before P3. The sender then retransmits P1 and
P2. Unfortunately, the retransmission of P1 is lost again. However,
the delivery of the retransmission of P2 allows RACK to infer that
the retransmission of P1 was likely lost (Event 7a); hence, P1 should
be retransmitted (Event 7b). Note that [RFC5681] mandates a
principle that loss in two successive windows of data or the loss of
a retransmission must be taken as two indications of congestion and
therefore results in two separate congestion control reactions.
3.5. An Example of RACK-TLP in Action: RTO
In addition to enhancing fast recovery, RACK improves the accuracy of
RTO recovery by reducing spurious retransmissions.
Without RACK, upon RTO timer expiration, the sender marks all the
unacknowledged segments as lost. This approach can lead to spurious
retransmissions. For example, consider a simple case where one
segment was sent with an RTO of 1 second and then the application
writes more data, causing a second and third segment to be sent right
before the RTO of the first segment expires. Suppose none of the
segments were lost. Without RACK, if there is a spurious RTO, then
the sender marks all three segments as lost and retransmits the first
segment. If the ACK for the original copy of the first segment
arrives right after the spurious RTO retransmission, then the sender
continues slow start and spuriously retransmits the second and third
segments since it (erroneously) presumed they are lost.
With RACK, upon RTO timer expiration, the only segment automatically
marked as lost is the first segment (since it was sent an RTO ago);
for all the other segments, RACK only marks the segment as lost if at
least one round trip has elapsed since the segment was transmitted.
Consider the previous example scenario, but this time with RACK.
With RACK, when the RTO expires, the sender only marks the first
segment as lost and retransmits that segment. The other two very
recently sent segments are not marked as lost because they were sent
less than one round trip ago and there were no ACKs providing
evidence that they were lost. Upon receiving the ACK for the RTO
retransmission, the RACK sender would not yet retransmit the second
or third segment, but rather would re-arm the RTO timer and wait for
a new RTO interval to elapse before marking the second or third
segment as lost.
3.6. Design Summary
To summarize, RACK-TLP aims to adapt to small time-varying degrees of
reordering, quickly recover most losses within one to two round
trips, and avoid costly RTO recoveries. In the presence of
reordering, the adaptation algorithm can impose sometimes needless
delays when it waits to disambiguate loss from reordering, but the
penalty for waiting is bounded to one round trip, and such delays are
confined to flows long enough to have observed reordering.
4. Requirements
The reader is expected to be familiar with the definitions given in
the TCP congestion control [RFC5681], selective acknowledgment
[RFC2018], and loss recovery [RFC6675] RFCs. RACK-TLP has the
following requirements:
1. The connection MUST use selective acknowledgment (SACK) options
[RFC2018], and the sender MUST keep SACK scoreboard information
on a per-connection basis ("SACK scoreboard" has the same meaning
here as in [RFC6675], Section 3).
2. For each data segment sent, the sender MUST store its most recent
transmission time with a timestamp whose granularity is finer
than 1/4 of the minimum RTT of the connection. At the time of
writing, microsecond resolution is suitable for intra-data center
traffic, and millisecond granularity or finer is suitable for the
Internet. Note that RACK-TLP can be implemented with TSO (TCP
Segmentation Offload) support by having multiple segments in a
TSO aggregate share the same timestamp.
3. RACK DSACK-based reordering window adaptation is RECOMMENDED but
is not required.
4. TLP requires RACK.
5. Definitions
The reader is expected to be familiar with the variables SND.UNA,
SND.NXT, SEG.ACK, and SEG.SEQ in [RFC793]; Sender Maximum Segment
Size (SMSS) and FlightSize in [RFC5681]; DupThresh in [RFC6675]; and
RTO and SRTT in [RFC6298]. A RACK-TLP implementation uses several
new terms and needs to store new per-segment and per-connection
state, described below.
5.1. Terms
These terms are used to explain the variables and algorithms below:
RACK.segment
Among all the segments that have been either selectively or
cumulatively acknowledged, the term "RACK.segment" denotes the
segment that was sent most recently (including retransmissions).
RACK.ack_ts
Denotes the time when the full sequence range of RACK.segment was
selectively or cumulatively acknowledged.
5.2. Per-Segment Variables
These variables indicate the status of the most recent transmission
of a data segment:
Segment.lost
True if the most recent (re)transmission of the segment has been
marked as lost and needs to be retransmitted. False otherwise.
Segment.retransmitted
True if the segment has ever been retransmitted. False otherwise.
Segment.xmit_ts
The time of the last transmission of a data segment, including
retransmissions, if any, with a clock granularity specified in the
"Requirements" section. A maximum value INFINITE_TS indicates an
invalid timestamp that represents that the segment is not
currently in flight.
Segment.end_seq
The next sequence number after the last sequence number of the
data segment.
5.3. Per-Connection Variables
RACK.xmit_ts
The latest transmission timestamp of RACK.segment.
RACK.end_seq
The Segment.end_seq of RACK.segment.
RACK.segs_sacked
Returns the total number of segments selectively acknowledged in
the SACK scoreboard.
RACK.fack
The highest selectively or cumulatively acknowledged sequence
(i.e., forward acknowledgment).
RACK.min_RTT
The estimated minimum round-trip time (RTT) of the connection.
RACK.rtt
The RTT of the most recently delivered segment on the connection
(either cumulatively acknowledged or selectively acknowledged)
that was not marked as invalid as a possible spurious
retransmission.
RACK.reordering_seen
Indicates whether the sender has detected data segment reordering
event(s).
RACK.reo_wnd
A reordering window computed in the unit of time used for
recording segment transmission times. It is used to defer the
moment at which RACK marks a segment as lost.
RACK.dsack_round
Indicates if a DSACK option has been received in the latest round
trip.
RACK.reo_wnd_mult
The multiplier applied to adjust RACK.reo_wnd.
RACK.reo_wnd_persist
The number of loss recoveries before resetting RACK.reo_wnd.
TLP.is_retrans
A boolean indicating whether there is an unacknowledged TLP
retransmission.
TLP.end_seq
The value of SND.NXT at the time of sending a TLP probe.
TLP.max_ack_delay:
The sender's budget for the maximum delayed ACK interval.
5.4. Per-Connection Timers
RACK reordering timer
A timer that allows RACK to wait for reordering to resolve in
order to try to disambiguate reordering from loss when some
segments are marked as SACKed.
TLP PTO
A timer event indicating that an ACK is overdue and the sender
should transmit a TLP segment to solicit SACK or ACK feedback.
These timers augment the existing timers maintained by a sender,
including the RTO timer [RFC6298]. A RACK-TLP sender arms one of
these three timers -- RACK reordering timer, TLP PTO timer, or RTO
timer -- when it has unacknowledged segments in flight. The
implementation can simplify managing all three timers by multiplexing
a single timer among them with an additional variable to indicate the
event to invoke upon the next timer expiration.
6. RACK Algorithm Details
6.1. Upon Transmitting a Data Segment
Upon transmitting a new segment or retransmitting an old segment,
record the time in Segment.xmit_ts and set Segment.lost to FALSE.
Upon retransmitting a segment, set Segment.retransmitted to TRUE.
RACK_transmit_new_data(Segment):
Segment.xmit_ts = Now()
Segment.lost = FALSE
RACK_retransmit_data(Segment):
Segment.retransmitted = TRUE
Segment.xmit_ts = Now()
Segment.lost = FALSE
6.2. Upon Receiving an ACK
Step 1: Update RACK.min_RTT.
Use the RTT measurements obtained via [RFC6298] or [RFC7323] to
update the estimated minimum RTT in RACK.min_RTT. The sender SHOULD
track a windowed min-filtered estimate of recent RTT measurements
that can adapt when migrating to significantly longer paths rather
than tracking a simple global minimum of all RTT measurements.
Step 2: Update the state for the most recently sent segment that has
been delivered.
In this step, RACK updates the state that tracks the most recently
sent segment that has been delivered: RACK.segment. RACK maintains
its latest transmission timestamp in RACK.xmit_ts and its highest
sequence number in RACK.end_seq. These two variables are used in
later steps to estimate if some segments not yet delivered were
likely lost. Given the information provided in an ACK, each segment
cumulatively ACKed or SACKed is marked as delivered in the
scoreboard. Because an ACK can also acknowledge retransmitted data
segments and because retransmissions can be spurious, the sender
needs to take care to avoid spurious inferences. For example, if the
sender were to use timing information from a spurious retransmission,
the RACK.rtt could be vastly underestimated.
To avoid spurious inferences, ignore a segment as invalid if any of
its sequence range has been retransmitted before and if either of two
conditions is true:
1. The Timestamp Echo Reply field (TSecr) of the ACK's timestamp
option [RFC7323], if available, indicates the ACK was not
acknowledging the last retransmission of the segment.
2. The segment was last retransmitted less than RACK.min_rtt ago.
The second check is a heuristic when the TCP Timestamp option is not
available or when the round-trip time is less than the TCP Timestamp
clock granularity.
Among all the segments newly ACKed or SACKed by this ACK that pass
the checks above, update the RACK.rtt to be the RTT sample calculated
using this ACK. Furthermore, record the most recent Segment.xmit_ts
in RACK.xmit_ts if it is ahead of RACK.xmit_ts. If Segment.xmit_ts
equals RACK.xmit_ts (e.g., due to clock granularity limits), then
compare Segment.end_seq and RACK.end_seq to break the tie when
deciding whether to update the RACK.segment's associated state.
Step 2 may be summarized in pseudocode as:
RACK_sent_after(t1, seq1, t2, seq2):
If t1 > t2:
Return true
Else if t1 == t2 AND seq1 > seq2:
Return true
Else:
Return false
RACK_update():
For each Segment newly acknowledged, cumulatively or selectively,
in ascending order of Segment.xmit_ts:
rtt = Now() - Segment.xmit_ts
If Segment.retransmitted is TRUE:
If ACK.ts_option.echo_reply < Segment.xmit_ts:
Continue
If rtt < RACK.min_rtt:
Continue
RACK.rtt = rtt
If RACK_sent_after(Segment.xmit_ts, Segment.end_seq
RACK.xmit_ts, RACK.end_seq):
RACK.xmit_ts = Segment.xmit_ts
RACK.end_seq = Segment.end_seq
Step 3: Detect data segment reordering.
To detect reordering, the sender looks for original data segments
being delivered out of order. To detect such cases, the sender
tracks the highest sequence selectively or cumulatively acknowledged
in the RACK.fack variable. ".fack" stands for the most "Forward ACK"
(this term is adopted from [FACK]). If a never-retransmitted segment
that's below RACK.fack is (selectively or cumulatively) acknowledged,
it has been delivered out of order. The sender sets
RACK.reordering_seen to TRUE if such a segment is identified.
RACK_detect_reordering():
For each Segment newly acknowledged, cumulatively or selectively,
in ascending order of Segment.end_seq:
If Segment.end_seq > RACK.fack:
RACK.fack = Segment.end_seq
Else if Segment.end_seq < RACK.fack AND
Segment.retransmitted is FALSE:
RACK.reordering_seen = TRUE
Step 4: Update the RACK reordering window.
The RACK reordering window, RACK.reo_wnd, serves as an adaptive
allowance for settling time before marking a segment as lost. This
step documents a detailed algorithm that follows the principles
outlined in the "Reordering Window Adaptation" section.
If no reordering has been observed based on the previous step, then
one way the sender can enter fast recovery is when the number of
SACKed segments matches or exceeds DupThresh (similar to [RFC6675]).
Furthermore, when no reordering has been observed, the RACK.reo_wnd
is set to 0 both upon entering and during fast recovery or RTO
recovery.
Otherwise, if some reordering has been observed, then RACK does not
trigger fast recovery based on DupThresh.
Whether or not reordering has been observed, RACK uses the reordering
window to assess whether any segments can be marked as lost. As a
consequence, the sender also enters fast recovery when there are any
number of SACKed segments, as long as the reorder window has passed
for some non-SACKed segments.
When the reordering window is not set to 0, it starts with a
conservative RACK.reo_wnd of RACK.min_RTT/4. This value was chosen
because Linux TCP used the same factor in its implementation to delay
Early Retransmit [RFC5827] to reduce spurious loss detections in the
presence of reordering, and experience showed this worked reasonably
well [DMCG11].
However, the reordering detection in the previous step, Step 3, has a
self-reinforcing drawback when the reordering window is too small to
cope with the actual reordering. When that happens, RACK could
spuriously mark reordered segments as lost, causing them to be
retransmitted. In turn, the retransmissions can prevent the
necessary conditions for Step 3 to detect reordering since this
mechanism requires ACKs or SACKs only for segments that have never
been retransmitted. In some cases, such scenarios can persist,
causing RACK to continue to spuriously mark segments as lost without
realizing the reordering window is too small.
To avoid the issue above, RACK dynamically adapts to higher degrees
of reordering using DSACK options from the receiver. Receiving an
ACK with a DSACK option indicates a possible spurious retransmission,
suggesting that RACK.reo_wnd may be too small. The RACK.reo_wnd
increases linearly for every round trip in which the sender receives
some DSACK option so that after N round trips in which a DSACK is
received, the RACK.reo_wnd becomes (N+1) * min_RTT / 4, with an
upper-bound of SRTT.
If the reordering is temporary, then a large adapted reordering
window would unnecessarily delay loss recovery later. Therefore,
RACK persists using the inflated RACK.reo_wnd for up to 16 loss
recoveries, after which it resets RACK.reo_wnd to its starting value,
min_RTT / 4. The downside of resetting the reordering window is the
risk of triggering spurious fast recovery episodes if the reordering
remains high. The rationale for this approach is to bound such
spurious recoveries to approximately once every 16 recoveries (less
than 7%).
To track the linear scaling factor for the adaptive reordering
window, RACK uses the variable RACK.reo_wnd_mult, which is
initialized to 1 and adapts with the observed reordering.
The following pseudocode implements the above algorithm for updating
the RACK reordering window:
RACK_update_reo_wnd():
/* DSACK-based reordering window adaptation */
If RACK.dsack_round is not None AND
SND.UNA >= RACK.dsack_round:
RACK.dsack_round = None
/* Grow the reordering window per round that sees DSACK.
Reset the window after 16 DSACK-free recoveries */
If RACK.dsack_round is None AND
any DSACK option is present on latest received ACK:
RACK.dsack_round = SND.NXT
RACK.reo_wnd_mult += 1
RACK.reo_wnd_persist = 16
Else if exiting Fast or RTO recovery:
RACK.reo_wnd_persist -= 1
If RACK.reo_wnd_persist <= 0:
RACK.reo_wnd_mult = 1
If RACK.reordering_seen is FALSE:
If in Fast or RTO recovery:
Return 0
Else if RACK.segs_sacked >= DupThresh:
Return 0
Return min(RACK.reo_wnd_mult * RACK.min_RTT / 4, SRTT)
Step 5: Detect losses.
For each segment that has not been SACKed, RACK considers that
segment lost if another segment that was sent later has been
delivered and the reordering window has passed. RACK considers the
reordering window to have passed if the RACK.segment was sent a
sufficient time after the segment in question, if a sufficient time
has elapsed since the RACK.segment was S/ACKed, or some combination
of the two. More precisely, RACK marks a segment as lost if:
RACK.xmit_ts >= Segment.xmit_ts
AND
RACK.xmit_ts - Segment.xmit_ts + (now - RACK.ack_ts) >= RACK.reo_wnd
Solving this second condition for "now", the moment at which a
segment is marked as lost, yields:
now >= Segment.xmit_ts + RACK.reo_wnd + (RACK.ack_ts - RACK.xmit_ts)
Then (RACK.ack_ts - RACK.xmit_ts) is the round-trip time of the most
recently (re)transmitted segment that's been delivered. When
segments are delivered in order, the most recently (re)transmitted
segment that's been delivered is also the most recently delivered;
hence, RACK.rtt == RACK.ack_ts - RACK.xmit_ts. But if segments were
reordered, then the segment delivered most recently was sent before
the most recently (re)transmitted segment. Hence, RACK.rtt >
(RACK.ack_ts - RACK.xmit_ts).
Since RACK.RTT >= (RACK.ack_ts - RACK.xmit_ts), the previous equation
reduces to saying that the sender can declare a segment lost when:
now >= Segment.xmit_ts + RACK.reo_wnd + RACK.rtt
In turn, that is equivalent to stating that a RACK sender should
declare a segment lost when:
Segment.xmit_ts + RACK.rtt + RACK.reo_wnd - now <= 0
Note that if the value on the left-hand side is positive, it
represents the remaining wait time before the segment is deemed lost.
But this risks a timeout (RTO) if no more ACKs come back (e.g., due
to losses or application-limited transmissions) to trigger the
marking. For timely loss detection, it is RECOMMENDED that the
sender install a reordering timer. This timer expires at the
earliest moment when RACK would conclude that all the unacknowledged
segments within the reordering window were lost.
The following pseudocode implements the algorithm above. When an ACK
is received or the RACK reordering timer expires, call
RACK_detect_loss_and_arm_timer(). The algorithm breaks timestamp
ties by using the TCP sequence space since high-speed networks often
have multiple segments with identical timestamps.
RACK_detect_loss():
timeout = 0
RACK.reo_wnd = RACK_update_reo_wnd()
For each segment, Segment, not acknowledged yet:
If RACK_sent_after(RACK.xmit_ts, RACK.end_seq,
Segment.xmit_ts, Segment.end_seq):
remaining = Segment.xmit_ts + RACK.rtt +
RACK.reo_wnd - Now()
If remaining <= 0:
Segment.lost = TRUE
Segment.xmit_ts = INFINITE_TS
Else:
timeout = max(remaining, timeout)
Return timeout
RACK_detect_loss_and_arm_timer():
timeout = RACK_detect_loss()
If timeout != 0
Arm the RACK timer to call
RACK_detect_loss_and_arm_timer() after timeout
As an optimization, an implementation can choose to check only
segments that have been sent before RACK.xmit_ts. This can be more
efficient than scanning the entire SACK scoreboard, especially when
there are many segments in flight. The implementation can use a
separate doubly linked list ordered by Segment.xmit_ts, insert a
segment at the tail of the list when it is (re)transmitted, and
remove a segment from the list when it is delivered or marked as
lost. In Linux TCP, this optimization improved CPU usage by orders
of magnitude during some fast recovery episodes on high-speed WAN
networks.
6.3. Upon RTO Expiration
Upon RTO timer expiration, RACK marks the first outstanding segment
as lost (since it was sent an RTO ago); for all the other segments,
RACK only marks the segment as lost if the time elapsed since the
segment was transmitted is at least the sum of the recent RTT and the
reordering window.
RACK_mark_losses_on_RTO():
For each segment, Segment, not acknowledged yet:
If SEG.SEQ == SND.UNA OR
Segment.xmit_ts + RACK.rtt + RACK.reo_wnd - Now() <= 0:
Segment.lost = TRUE
7. TLP Algorithm Details
7.1. Initializing State
Reset TLP.is_retrans and TLP.end_seq when initiating a connection,
fast recovery, or RTO recovery.
TLP_init():
TLP.end_seq = None
TLP.is_retrans = false
7.2. Scheduling a Loss Probe
The sender schedules a loss probe timeout (PTO) to transmit a segment
during the normal transmission process. The sender SHOULD start or
restart a loss probe PTO timer after transmitting new data (that was
not itself a loss probe) or upon receiving an ACK that cumulatively
acknowledges new data unless it is already in fast recovery, RTO
recovery, or segments have been SACKed (i.e., RACK.segs_sacked is not
zero). These conditions are excluded because they are addressed by
similar mechanisms, like Limited Transmit [RFC3042], the RACK
reordering timer, and Forward RTO-Recovery (F-RTO) [RFC5682].
The sender calculates the PTO interval by taking into account a
number of factors.
First, the default PTO interval is 2*SRTT. By that time, it is
prudent to declare that an ACK is overdue since under normal
circumstances, i.e., no losses, an ACK typically arrives in one SRTT.
Choosing the PTO to be exactly an SRTT would risk causing spurious
probes given that network and end-host delay variance can cause an
ACK to be delayed beyond the SRTT. Hence, the PTO is conservatively
chosen to be the next integral multiple of SRTT.
Second, when there is no SRTT estimate available, the PTO SHOULD be 1
second. This conservative value corresponds to the RTO value when no
SRTT is available, per [RFC6298].
Third, when the FlightSize is one segment, the sender MAY inflate the
PTO by TLP.max_ack_delay to accommodate a potentially delayed
acknowledgment and reduce the risk of spurious retransmissions. The
actual value of TLP.max_ack_delay is implementation specific.
Finally, if the time at which an RTO would fire (here denoted as
"TCP_RTO_expiration()") is sooner than the computed time for the PTO,
then the sender schedules a TLP to be sent at that RTO time.
Summarizing these considerations in pseudocode form, a sender SHOULD
use the following logic to select the duration of a PTO:
TLP_calc_PTO():
If SRTT is available:
PTO = 2 * SRTT
If FlightSize is one segment:
PTO += TLP.max_ack_delay
Else:
PTO = 1 sec
If Now() + PTO > TCP_RTO_expiration():
PTO = TCP_RTO_expiration() - Now()
7.3. Sending a Loss Probe upon PTO Expiration
When the PTO timer expires, the sender MUST check whether both of the
following conditions are met before sending a loss probe:
1. First, there is no other previous loss probe still in flight.
This ensures that, at any given time, the sender has at most one
additional packet in flight beyond the congestion window limit.
This invariant is maintained using the state variable
TLP.end_seq, which indicates the latest unacknowledged TLP loss
probe's ending sequence. It is reset when the loss probe has
been acknowledged or is deemed lost or irrelevant.
2. Second, the sender has obtained an RTT measurement since the last
loss probe transmission or the start of the connection, whichever
was later. This condition ensures that loss probe
retransmissions do not prevent taking the RTT samples necessary
to adapt SRTT to an increase in path RTT.
If either one of these two conditions is not met, then the sender
MUST skip sending a loss probe and MUST proceed to re-arm the RTO
timer, as specified at the end of this section.
If both conditions are met, then the sender SHOULD transmit a
previously unsent data segment, if one exists and the receive window
allows, and increment the FlightSize accordingly. Note that the
FlightSize could be one packet greater than the congestion window
temporarily until the next ACK arrives.
If such an unsent segment is not available, then the sender SHOULD
retransmit the highest-sequence segment sent so far and set
TLP.is_retrans to true. This segment is chosen to deal with the
retransmission ambiguity problem in TCP. Suppose a sender sends N
segments and then retransmits the last segment (segment N) as a loss
probe, after which the sender receives a SACK for segment N. As long
as the sender waits for the RACK reordering window to expire, it
doesn't matter if that SACK was for the original transmission of
segment N or the TLP retransmission; in either case, the arrival of
the SACK for segment N provides evidence that the N-1 segments
preceding segment N were likely lost.
In a case where there is only one original outstanding segment of
data (N=1), the same logic (trivially) applies: an ACK for a single
outstanding segment tells the sender that the N-1=0 segments
preceding that segment were lost. Furthermore, whether there are N>1
or N=1 outstanding segments, there is a question about whether the
original last segment or its TLP retransmission were lost; the sender
estimates whether there was such a loss using TLP recovery detection
(see below).
The sender MUST follow the RACK transmission procedures in the "Upon
Transmitting a Data Segment" section upon sending either a
retransmission or a new data loss probe. This is critical for
detecting losses using the ACK for the loss probe.
After attempting to send a loss probe, regardless of whether a loss
probe was sent, the sender MUST re-arm the RTO timer, not the PTO
timer, if the FlightSize is not zero. This ensures RTO recovery
remains the last resort if TLP fails. The following pseudocode
summarizes the operations.
TLP_send_probe():
If TLP.end_seq is None and
Sender has taken a new RTT sample since last probe or
the start of connection:
TLP.is_retrans = false
Segment = send buffer segment starting at SND.NXT
If Segment exists and fits the peer receive window limit:
/* Transmit the lowest-sequence unsent Segment */
Transmit Segment
RACK_transmit_data(Segment)
TLP.end_seq = SND.NXT
Increase FlightSize by Segment length
Else:
/* Retransmit the highest-sequence Segment sent */
Segment = send buffer segment ending at SND.NXT
Transmit Segment
RACK_retransmit_data(Segment)
TLP.end_seq = SND.NXT
TLP.is_retrans = true
If FlightSize is not zero:
Rearm RTO timer to fire at timeout = now + RTO
7.4. Detecting Losses Using the ACK of the Loss Probe
When there is packet loss in a flight ending with a loss probe, the
feedback solicited by a loss probe will reveal one of two scenarios,
depending on the pattern of losses.
7.4.1. General Case: Detecting Packet Losses Using RACK
If the loss probe and the ACK that acknowledges the probe are
delivered successfully, RACK-TLP uses this ACK -- just as it would
with any other ACK -- to detect if any segments sent prior to the
probe were dropped. RACK would typically infer that any
unacknowledged data segments sent before the loss probe were lost,
since they were sent sufficiently far in the past (where at least one
PTO has elapsed, plus one round trip for the loss probe to be ACKed).
More specifically, RACK_detect_loss() (Step 5) would mark those
earlier segments as lost. Then the sender would trigger a fast
recovery to recover those losses.
7.4.2. Special Case: Detecting a Single Loss Repaired by the Loss Probe
If the TLP retransmission repairs all the lost in-flight sequence
ranges (i.e., only the last segment in the flight was lost), the ACK
for the loss probe appears to be a regular cumulative ACK, which
would not normally trigger the congestion control response to this
packet loss event. The following TLP recovery detection mechanism
examines ACKs to detect this special case to make congestion control
respond properly [RFC5681].
After a TLP retransmission, the sender checks for this special case
of a single loss that is recovered by the loss probe itself. To
accomplish this, the sender checks for a duplicate ACK or DSACK
indicating that both the original segment and TLP retransmission
arrived at the receiver, which means there was no loss. If the TLP
sender does not receive such an indication, then it MUST assume that
the original data segment, the TLP retransmission, or a corresponding
ACK was lost for congestion control purposes.
If the TLP retransmission is spurious, a receiver that uses DSACK
would return an ACK that covers TLP.end_seq with a DSACK option (Case
1). If the receiver does not support DSACK, it would return a DupAck
without any SACK option (Case 2). If the sender receives an ACK
matching either case, then the sender estimates that the receiver
received both the original data segment and the TLP probe
retransmission. The sender considers the TLP episode to be done and
records that fact by setting TLP.end_seq to None.
Upon receiving an ACK that covers some sequence number after
TLP.end_seq, the sender should have received any ACKs for the
original segment and TLP probe retransmission segment. At that time,
if the TLP.end_seq is still set and thus indicates that the TLP probe
retransmission remains unacknowledged, then the sender should presume
that at least one of its data segments was lost. The sender then
SHOULD invoke a congestion control response equivalent to a fast
recovery.
More precisely, on each ACK, the sender executes the following:
TLP_process_ack(ACK):
If TLP.end_seq is not None AND ACK's ack. number >= TLP.end_seq:
If not TLP.is_retrans:
TLP.end_seq = None /* TLP of new data delivered */
Else if ACK has a DSACK option matching TLP.end_seq:
TLP.end_seq = None /* Case 1, above */
Else If ACK's ack. number > TLP.end_seq:
TLP.end_seq = None /* Repaired the single loss */
(Invoke congestion control to react to
the loss event the probe has repaired)
Else If ACK is a DupAck without any SACK option:
TLP.end_seq = None /* Case 2, above */
8. Managing RACK-TLP Timers
The RACK reordering timer, the TLP PTO timer, the RTO, and Zero
Window Probe (ZWP) timer [RFC793] are mutually exclusive and are used
in different scenarios. When arming a RACK reordering timer or TLP
PTO timer, the sender SHOULD cancel any other pending timers. An
implementation is expected to have one timer with an additional state
variable indicating the type of the timer.
9. Discussion
9.1. Advantages and Disadvantages
The biggest advantage of RACK-TLP is that every data segment, whether
it is an original data transmission or a retransmission, can be used
to detect losses of the segments sent chronologically prior to it.
This enables RACK-TLP to use fast recovery in cases with application-
limited flights of data, lost retransmissions, or data segment
reordering events. Consider the following examples:
1. Packet drops at the end of an application data flight: Consider a
sender that transmits an application-limited flight of three data
segments (P1, P2, P3), and P1 and P3 are lost. Suppose the
transmission of each segment is at least RACK.reo_wnd after the
transmission of the previous segment. RACK will mark P1 as lost
when the SACK of P2 is received, and this will trigger the
retransmission of P1 as R1. When R1 is cumulatively
acknowledged, RACK will mark P3 as lost, and the sender will
retransmit P3 as R3. This example illustrates how RACK is able
to repair certain drops at the tail of a transaction without an
RTO recovery. Notice that neither the conventional duplicate ACK
threshold [RFC5681], nor the loss recovery algorithm [RFC6675],
nor the Forward Acknowledgment [FACK] algorithm can detect such
losses because of the required segment or sequence count.
2. Lost retransmission: Consider a flight of three data segments
(P1, P2, P3) that are sent; P1 and P2 are dropped. Suppose the
transmission of each segment is at least RACK.reo_wnd after the
transmission of the previous segment. When P3 is SACKed, RACK
will mark P1 and P2 as lost, and they will be retransmitted as R1
and R2. Suppose R1 is lost again but R2 is SACKed; RACK will
mark R1 as lost and trigger retransmission again. Again, neither
the conventional three-duplicate ACK threshold approach, nor the
loss recovery algorithm [RFC6675], nor the Forward Acknowledgment
[FACK] algorithm can detect such losses. And such a lost
retransmission can happen when TCP is being rate-limited,
particularly by token bucket policers with a large bucket depth
and low rate limit; in such cases, retransmissions are often lost
repeatedly because standard congestion control requires multiple
round trips to reduce the rate below the policed rate.
3. Packet reordering: Consider a simple reordering event where a
flight of segments are sent as (P1, P2, P3). P1 and P2 carry a
full payload of Maximum Sender Size (MSS) octets, but P3 has only
a 1-octet payload. Suppose the sender has detected reordering
previously and thus RACK.reo_wnd is min_RTT/4. Now P3 is
reordered and delivered first, before P1 and P2. As long as P1
and P2 are delivered within min_RTT/4, RACK will not consider P1
and P2 lost. But if P1 and P2 are delivered outside the
reordering window, then RACK will still spuriously mark P1 and P2
as lost.
The examples above show that RACK-TLP is particularly useful when the
sender is limited by the application, which can happen with
interactive or request/response traffic. Similarly, RACK still works
when the sender is limited by the receive window, which can happen
with applications that use the receive window to throttle the sender.
RACK-TLP works more efficiently with TCP Segmentation Offload (TSO)
compared to DupAck counting. RACK always marks the entire TSO
aggregate as lost because the segments in the same TSO aggregate have
the same transmission timestamp. By contrast, the algorithms based
on sequence counting (e.g., [RFC6675], [RFC5681]) may mark only a
subset of segments in the TSO aggregate as lost, forcing the stack to
perform expensive fragmentation of the TSO aggregate or to
selectively tag individual segments as lost in the scoreboard.
The main drawback of RACK-TLP is the additional state required
compared to DupAck counting. RACK requires the sender to record the
transmission time of each segment sent at a clock granularity that is
finer than 1/4 of the minimum RTT of the connection. TCP
implementations that already record this for RTT estimation do not
require any new per-packet state. But implementations that are not
yet recording segment transmission times will need to add per-packet
internal state (expected to be either 4 or 8 octets per segment or
TSO aggregate) to track transmission times. In contrast, the loss
detection approach described in [RFC6675] does not require any per-
packet state beyond the SACK scoreboard; this is particularly useful
on ultra-low RTT networks where the RTT may be less than the sender
TCP clock granularity (e.g., inside data centers). Another
disadvantage is that the reordering timer may expire prematurely
(like any other retransmission timer) and cause higher spurious
retransmissions, especially if DSACK is not supported.
9.2. Relationships with Other Loss Recovery Algorithms
The primary motivation of RACK-TLP is to provide a general
alternative to some of the standard loss recovery algorithms
[RFC5681] [RFC6675] [RFC5827] [RFC4653]. In particular, the SACK
loss recovery algorithm for TCP [RFC6675] is not designed to handle
lost retransmissions, so its NextSeg() does not work for lost
retransmissions, and it does not specify the corresponding required
additional congestion response. Therefore, the algorithm [RFC6675]
MUST NOT be used with RACK-TLP; instead, a modified recovery
algorithm that carefully addresses such a case is needed.
The Early Retransmit mechanism [RFC5827] and NCR for TCP [RFC4653]
dynamically adjust the duplicate ACK threshold based on the current
or previous flight sizes. RACK-TLP takes a different approach by
using a time-based reordering window. RACK-TLP can be seen as an
extended Early Retransmit [RFC5827] without a FlightSize limit but
with an additional reordering window. [FACK] considers an original
segment to be lost when its sequence range is sufficiently far below
the highest SACKed sequence. In some sense, RACK-TLP can be seen as
a generalized form of FACK that operates in time space instead of
sequence space, enabling it to better handle reordering, application-
limited traffic, and lost retransmissions.
RACK-TLP is compatible with the standard RTO [RFC6298], RTO Restart
[RFC7765], F-RTO [RFC5682], and Eifel algorithms [RFC3522]. This is
because RACK-TLP only detects loss by using ACK events. It neither
changes the RTO timer calculation nor detects spurious RTOs. RACK-
TLP slightly changes the behavior of [RFC6298] by preceding the RTO
with a TLP and reducing potential spurious retransmissions after RTO.
9.3. Interaction with Congestion Control
RACK-TLP intentionally decouples loss detection from congestion
control. RACK-TLP only detects losses; it does not modify the
congestion control algorithm [RFC5681] [RFC6937]. A segment marked
as lost by RACK-TLP MUST NOT be retransmitted until congestion
control deems this appropriate. As mentioned in the paragraph
following Figure 1 (Section 3.4, Paragraph 3), [RFC5681] mandates a
principle that loss in two successive windows of data or the loss of
a retransmission must be taken as two indications of congestion and
therefore trigger two separate reactions. The Proportional Rate
Reduction (PRR) algorithm [RFC6937] is RECOMMENDED for the specific
congestion control actions taken upon the losses detected by RACK-
TLP. In the absence of PRR [RFC6937], when RACK-TLP detects a lost
retransmission, the congestion control MUST trigger an additional
congestion response per the aforementioned principle in [RFC5681].
If multiple original transmissions or retransmissions were lost in a
window, the congestion control specified in [RFC5681] only reacts
once per window. The congestion control implementer is advised to
carefully consider this subtle situation introduced by RACK-TLP.
The only exception -- the only way in which RACK-TLP modulates the
congestion control algorithm -- is that one outstanding loss probe
can be sent even if the congestion window is fully used. However,
this temporary overcommit is accounted for and credited in the in-
flight data tracked for congestion control, so that congestion
control will erase the overcommit upon the next ACK.
If packet losses happen after reordering has been observed, RACK-TLP
may take longer to detect losses than the pure DupAck counting
approach. In this case, TCP may continue to increase the congestion
window upon receiving ACKs during this time, making the sender more
aggressive.
The following simple example compares how RACK-TLP and non-RACK-TLP
loss detection interact with congestion control: suppose a sender has
a congestion window (cwnd) of 20 segments on a SACK-enabled
connection. It sends 10 data segments, and all of them are lost.
Without RACK-TLP, the sender would time out, reset cwnd to 1, and
retransmit the first segment. It would take four round trips (1 + 2
+ 4 + 3 = 10) to retransmit all the 10 lost segments using slow
start. The recovery latency would be RTO + 4*RTT, with an ending
cwnd of 4 segments due to congestion window validation.
With RACK-TLP, a sender would send the TLP after 2*RTT and get a
DupAck, enabling RACK to detect the losses and trigger fast recovery.
If the sender implements Proportional Rate Reduction [RFC6937], it
would slow start to retransmit the remaining 9 lost segments since
the number of segments in flight (0) is lower than the slow start
threshold (10). The slow start would again take four round trips (1
+ 2 + 4 + 3 = 10) to retransmit all the lost segments. The recovery
latency would be 2*RTT + 4*RTT, with an ending cwnd set to the slow-
start threshold of 10 segments.
The difference in recovery latency (RTO + 4*RTT vs 6*RTT) can be
significant if the RTT is much smaller than the minimum RTO (1 second
in [RFC6298]) or if the RTT is large. The former case can happen in
local area networks, data center networks, or content distribution
networks with deep deployments. The latter case can happen in
developing regions with highly congested and/or high-latency
networks.
9.4. TLP Recovery Detection with Delayed ACKs
Delayed or stretched ACKs complicate the detection of repairs done by
TLP since, with such ACKs, the sender takes a longer time to receive
fewer ACKs than would normally be expected. To mitigate this
complication, before sending a TLP loss probe retransmission, the
sender should attempt to wait long enough that the receiver has sent
any delayed ACKs that it is withholding. The sender algorithm
described above features such a delay in the form of
TLP.max_ack_delay. Furthermore, if the receiver supports DSACK,
then, in the case of a delayed ACK, the sender's TLP recovery
detection mechanism (see above) can use the DSACK information to
infer that the original and TLP retransmission both arrived at the
receiver.
If there is ACK loss or a delayed ACK without a DSACK, then this
algorithm is conservative because the sender will reduce the
congestion window when, in fact, there was no packet loss. In
practice, this is acceptable and potentially even desirable: if there
is reverse path congestion, then reducing the congestion window can
be prudent.
9.5. RACK-TLP for Other Transport Protocols
RACK-TLP can be implemented in other transport protocols (e.g.,
[QUIC-LR]). The [SPROUT] loss detection algorithm was also
independently designed to use a 10 ms reordering window to improve
its loss detection similar to RACK.
10. Security Considerations
RACK-TLP algorithm behavior is based on information conveyed in SACK
options, so it has security considerations similar to those described
in the Security Considerations section of [RFC6675].
Additionally, RACK-TLP has a lower risk profile than the loss
recovery algorithm [RFC6675] because it is not vulnerable to ACK-
splitting attacks [SCWA99]: for an MSS-sized segment sent, the
receiver or the attacker might send MSS ACKs that selectively or
cumulatively acknowledge one additional byte per ACK. This would not
fool RACK. In such a scenario, RACK.xmit_ts would not advance
because all the sequence ranges within the segment were transmitted
at the same time and thus carry the same transmission timestamp. In
other words, SACKing only one byte of a segment or SACKing the
segment in entirety have the same effect with RACK.
11. IANA Considerations
This document has no IANA actions.
12. References
12.1. Normative References
[RFC2018] Mathis, M., Mahdavi, J., Floyd, S., and A. Romanow, "TCP
Selective Acknowledgment Options", RFC 2018,
DOI 10.17487/RFC2018, October 1996,
<https://www.rfc-editor.org/info/rfc2018>.
[RFC2119] Bradner, S., "Key words for use in RFCs to Indicate
Requirement Levels", BCP 14, RFC 2119,
DOI 10.17487/RFC2119, March 1997,
<https://www.rfc-editor.org/info/rfc2119>.
[RFC2883] Floyd, S., Mahdavi, J., Mathis, M., and M. Podolsky, "An
Extension to the Selective Acknowledgement (SACK) Option
for TCP", RFC 2883, DOI 10.17487/RFC2883, July 2000,
<https://www.rfc-editor.org/info/rfc2883>.
[RFC5681] Allman, M., Paxson, V., and E. Blanton, "TCP Congestion
Control", RFC 5681, DOI 10.17487/RFC5681, September 2009,
<https://www.rfc-editor.org/info/rfc5681>.
[RFC6298] Paxson, V., Allman, M., Chu, J., and M. Sargent,
"Computing TCP's Retransmission Timer", RFC 6298,
DOI 10.17487/RFC6298, June 2011,
<https://www.rfc-editor.org/info/rfc6298>.
[RFC6675] Blanton, E., Allman, M., Wang, L., Jarvinen, I., Kojo, M.,
and Y. Nishida, "A Conservative Loss Recovery Algorithm
Based on Selective Acknowledgment (SACK) for TCP",
RFC 6675, DOI 10.17487/RFC6675, August 2012,
<https://www.rfc-editor.org/info/rfc6675>.
[RFC7323] Borman, D., Braden, B., Jacobson, V., and R.
Scheffenegger, Ed., "TCP Extensions for High Performance",
RFC 7323, DOI 10.17487/RFC7323, September 2014,
<https://www.rfc-editor.org/info/rfc7323>.
[RFC793] Postel, J., "Transmission Control Protocol", STD 7,
RFC 793, DOI 10.17487/RFC0793, September 1981,
<https://www.rfc-editor.org/info/rfc793>.
[RFC8174] Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC
2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174,
May 2017, <https://www.rfc-editor.org/info/rfc8174>.
12.2. Informative References
[DMCG11] Dukkipati, N., Matthis, M., Cheng, Y., and M. Ghobadi,
"Proportional Rate Reduction for TCP", Proceedings of the
2011 ACM SIGCOMM Conference on Internet Measurement
Conference pp. 155-170, DOI 10.1145/2068816.2068832,
November 2011, <https://doi.org/10.1145/2068816.2068832>.
[FACK] Mathis, M. and J. Mahdavi, "Forward acknowledgement:
refining TCP congestion control", ACM SIGCOMM Computer
Communication Review Volume 26, Issue 4,
DOI 10.1145/248157.248181, August 1996,
<https://doi.org/10.1145/248157.248181>.
[POLICER16]
Flach, T., Papageorge, P., Terzis, A., Pedrosa, L., Cheng,
Y., Karim, T., Katz-Bassett, E., and R. Govindan, "An
Internet-Wide Analysis of Traffic Policing", Proceedings
of the 2016 ACM SIGCOMM Conference pp. 468-482,
DOI 10.1145/2934872.2934873, August 2016,
<https://doi.org/10.1145/2934872.2934873>.
[QUIC-LR] Iyengar, J. and I. Swett, "QUIC Loss Detection and
Congestion Control", Work in Progress, Internet-Draft,
draft-ietf-quic-recovery-34, 14 January 2021,
<https://tools.ietf.org/html/draft-ietf-quic-recovery-34>.
[RFC3042] Allman, M., Balakrishnan, H., and S. Floyd, "Enhancing
TCP's Loss Recovery Using Limited Transmit", RFC 3042,
DOI 10.17487/RFC3042, January 2001,
<https://www.rfc-editor.org/info/rfc3042>.
[RFC3522] Ludwig, R. and M. Meyer, "The Eifel Detection Algorithm
for TCP", RFC 3522, DOI 10.17487/RFC3522, April 2003,
<https://www.rfc-editor.org/info/rfc3522>.
[RFC4653] Bhandarkar, S., Reddy, A. L. N., Allman, M., and E.
Blanton, "Improving the Robustness of TCP to Non-
Congestion Events", RFC 4653, DOI 10.17487/RFC4653, August
2006, <https://www.rfc-editor.org/info/rfc4653>.
[RFC5682] Sarolahti, P., Kojo, M., Yamamoto, K., and M. Hata,
"Forward RTO-Recovery (F-RTO): An Algorithm for Detecting
Spurious Retransmission Timeouts with TCP", RFC 5682,
DOI 10.17487/RFC5682, September 2009,
<https://www.rfc-editor.org/info/rfc5682>.
[RFC5827] Allman, M., Avrachenkov, K., Ayesta, U., Blanton, J., and
P. Hurtig, "Early Retransmit for TCP and Stream Control
Transmission Protocol (SCTP)", RFC 5827,
DOI 10.17487/RFC5827, May 2010,
<https://www.rfc-editor.org/info/rfc5827>.
[RFC6937] Mathis, M., Dukkipati, N., and Y. Cheng, "Proportional
Rate Reduction for TCP", RFC 6937, DOI 10.17487/RFC6937,
May 2013, <https://www.rfc-editor.org/info/rfc6937>.
[RFC7765] Hurtig, P., Brunstrom, A., Petlund, A., and M. Welzl, "TCP
and Stream Control Transmission Protocol (SCTP) RTO
Restart", RFC 7765, DOI 10.17487/RFC7765, February 2016,
<https://www.rfc-editor.org/info/rfc7765>.
[SCWA99] Savage, S., Cardwell, N., Wetherall, D., and T. Anderson,
"TCP congestion control with a misbehaving receiver", ACM
Computer Communication Review 29(5),
DOI 10.1145/505696.505704, October 1999,
<https://doi.org/10.1145/505696.505704>.
[SPROUT] Winstein, K., Sivaraman, A., and H. Balakrishnan,
"Stochastic Forecasts Achieve High Throughput and Low
Delay over Cellular Networks", 10th USENIX Symposium on
Networked Systems Design and Implementation (NSDI '13)",
2013.
Acknowledgments
The authors thank Matt Mathis for his insights in FACK and Michael
Welzl for his per-packet timer idea that inspired this work. Eric
Dumazet, Randy Stewart, Van Jacobson, Ian Swett, Rick Jones, Jana
Iyengar, Hiren Panchasara, Praveen Balasubramanian, Yoshifumi
Nishida, Bob Briscoe, Felix Weinrank, Michael Tüxen, Martin Duke,
Ilpo Jarvinen, Theresa Enghardt, Mirja Kühlewind, Gorry Fairhurst,
Markku Kojo, and Yi Huang contributed to this document or the
implementations in Linux, FreeBSD, Windows, and QUIC.
Authors' Addresses
Yuchung Cheng
Google, Inc.
Email: ycheng@google.com
Neal Cardwell
Google, Inc.
Email: ncardwell@google.com
Nandita Dukkipati
Google, Inc.
Email: nanditad@google.com
Priyaranjan Jha
Google, Inc.