Request for Comments: 4653 A. L. N. Reddy
Category: Experimental Texas A&M University
M. Allman
ICIR/ICSI
E. Blanton
Purdue University
August 2006
Improving the Robustness of TCP to Non-Congestion Events
Status of This Memo
This memo defines an Experimental Protocol for the Internet
community. It does not specify an Internet standard of any kind.
Discussion and suggestions for improvement are requested.
Distribution of this memo is unlimited.
Copyright Notice
Copyright (C) The Internet Society (2006).
Abstract
This document specifies Non-Congestion Robustness (NCR) for TCP. In
the absence of explicit congestion notification from the network, TCP
uses loss as an indication of congestion. One of the ways TCP
detects loss is using the arrival of three duplicate acknowledgments.
However, this heuristic is not always correct, notably in the case
when network paths reorder segments (for whatever reason), resulting
in degraded performance. TCP-NCR is designed to mitigate this
degraded performance by increasing the number of duplicate
acknowledgments required to trigger loss recovery, based on the
current state of the connection, in an effort to better disambiguate
true segment loss from segment reordering. This document specifies
the changes to TCP, as well as the costs and benefits of these
modifications.
Table of Contents
1. Introduction ....................................................2
1.1. Terminology ................................................4
2. NCR Description .................................................5
3. Algorithm .......................................................6
3.1. Initialization .............................................8
3.2. Terminating Extended Limited Transmit and
Preventing Bursts ..........................................9
3.3. Extended Limited Transmit .................................10
3.4. Entering Loss Recovery ....................................11
4. Advantages .....................................................12
5. Disadvantages ..................................................12
6. Related Work ...................................................13
7. Security Considerations ........................................14
8. Acknowledgments ................................................14
9. IANA Considerations ............................................14
10. References ....................................................14
10.1. Normative References .....................................14
10.2. Informative References ...................................15
1. Introduction
One strength of TCP [RFC793] lies in its ability to adjust its
sending rate according to the perceived congestion in the network
[Jac88, RFC2581]. In the absence of explicit notification of
congestion from the network, TCP uses segment loss as an indication
of congestion (i.e., assuming queue overflow). TCP receivers send
cumulative acknowledgments (ACKs) indicating the next sequence number
expected from the sender for arriving segments [RFC793]. When
segments arrive out of order, duplicate ACKs are generated. As
specified in [RFC2581], a TCP sender uses the arrival of three
duplicate ACKs as an indication of segment loss. The TCP sender
retransmits the lost segment and reduces the load imposed on the
network, assuming the segment loss was caused by resource contention
within the network path. The TCP sender does not assume loss on the
first or second duplicate ACK, but waits for three duplicate ACKs to
account for minor packet reordering. However, the use of this
constant threshold of duplicate ACKs has several problems that can be
mitigated with a dynamic threshold.
The following is an example of TCP’s behavior:
+ TCP A is the data sender, and TCP B is the data receiver.
+ TCP A sends 10 segments, each consisting of a single data byte
(i.e., transmits bytes 1-10 in segments 1-10).
+ Assume segment 3 is dropped in the network.
+ TCP B cumulatively acknowledges segments 1 and 2, making the
cumulative ACK transmitted to the sender 3 (the next expected
sequence number). (Note: TCP B may generate one or two ACKs,
depending on whether delayed ACKs [RFC1122, RFC2581] are
employed.)
+ The arrival of segments 4-10 at TCP B will each trigger the
transmission of a cumulative ACK for sequence number 3. (Note:
[RFC2581] recommends that delayed ACKs not be used when the ACK
is triggered by an out-of-order segment.)
+ When TCP A receives the third duplicate ACK (or fourth ACK
overall) for sequence number 3, TCP A will retransmit
segment 3 and reduce the sending rate by roughly half (see
[RFC2581] for specifics on the congestion control state
adjustments).
Alternatively, suppose segment 3 was not dropped by the network, but
rather delayed such that segment 3 arrives at TCP B after segment 10.
The above scenario will play out in precisely the same manner
insomuch as a retransmission of segment 3 will be triggered. In
other words, TCP is not capable of disambiguating this reordering
event from a segment loss, resulting in an unnecessary retransmission
and rate reduction.
The following is the specific motivation behind making TCP robust to
reordered segments:
* A number of Internet measurement studies have shown that packet
reordering is not a rare phenomenon [Pax97, BPS99, JIDKT03,
GPL04]. Further, the reordering can be well beyond that required
for fast retransmit to be falsely triggered.
* [BA02, ZKFP03] show the negative performance implications that
packet reordering has on current TCP.
* The requirement imposed by TCP for almost in-order packet
delivery places a constraint on the design of future technology.
Novel routing algorithms, network components, link-layer
retransmission mechanisms, and applications could all be looked
at with a fresh perspective if TCP were to be more robust to
segment reordering. For instance, high-speed packet switches
could cause resequencing of packets if TCP were more robust.
There has been work proposed in the literature explicitly to
ensure that packet ordering is maintained in such switches (e.g.,
[KM02]). Also, link-layer mechanisms that attempt to recover
from packet corruption by retransmitting could be allowed to
reorder packets, and thus increase the chances of local loss
repair rather than rely on TCP to repair the loss (and,
needlessly reduce its sending rate). Additional examples include
multi-path routing, high-delay satellite links, and some of the
schemes proposed for a differentiated services architecture. By
making TCP more robust to non-congestion events, TCP-NCR may open
the design space of the future Internet components.
In this document, we specify a set of TCP sender modifications to
provide Non-Congestion Robustness (NCR) to TCP. In particular, these
changes are built on top of TCP with selective acknowledgments
(SACKs) [RFC2018] and the SACK-based loss recovery scheme given in
[RFC3517], since SACK is widely deployed at this point ([MAF05]
indicates that 68% of web servers and 88% of web clients utilize SACK
as of spring 2004).
Note that the TCP-NCR algorithm provided in this document could be
easily adapted to SCTP [RFC2960] since SCTP uses congestion control
algorithms similar to TCP’s (and thus has the same reordering
robustness issues).
As noted in several places in the remainder of this document, we
consider TCP-NCR experimental in that more experience with the
techniques is required before TCP-NCR should be used on a large scale
on the Internet. We encourage implementation and experimentation
with TCP-NCR in the hopes of gaining an understanding of its
suitability for wide-scale deployment.
The remainder of this document is organized as follows. Section 2
provides a high-level description of the TCP-NCR mechanisms. In
Section 3, we specify the TCP-NCR algorithm. Section 4 provides a
brief overview of the benefits of TCP-NCR, while Section 5 discusses
the drawbacks. Section 6 discusses related work. Section 7
discusses security concerns.
1.1. Terminology
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
"SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this
document are to be interpreted as described in [RFC2119].
Readers should be familiar with the TCP terminology (e.g.,
FlightSize, Pipe) given in [RFC2581] and [RFC3517].
2. NCR Description
As discussed above, in the face of packet reordering, three duplicate
ACKs may not be enough to disambiguate loss from reordering. In this
section we provide a non-normative sketch of TCP-NCR. The detailed
algorithms for implementing Non-Congestion Robustness for TCP are
presented in the next section.
The general idea behind TCP-NCR is to increase the threshold used to
trigger a fast retransmission from the current fixed value of three
duplicate ACKs [RFC2581] to approximately a congestion window of data
having left the network (but not less than the currently standardized
value of three duplicate ACKs). Since cwnd represents the amount of
data a TCP flow can transmit in one round-trip time (RTT), waiting to
receive notice that cwnd bytes have left the network before deciding
whether the root cause is loss or reordering imposes a delay of
roughly one RTT on both the retransmission and the congestion control
response. The appropriate choice for a new value of the threshold is
essentially a trade-off between making the best decision regarding
the cause of the duplicate ACKs and responsiveness. The choice to
trigger a retransmission only after a cwnd’s worth of data is known
to have left the network represents roughly the largest amount of
time a TCP can wait before the (often costly) retransmission timeout
may be triggered. Therefore, the algorithm described in this
document attempts to make the best decision possible at the expense
of timeliness.
Simply increasing the threshold before retransmitting a segment can
make TCP brittle to packet loss or ACK loss since such loss reduces
the number of duplicate ACKs that will arrive at the sender from the
receiver. For instance, if the cwnd is 10 segments and one segment
is lost, a duplicate ACK threshold of 10 will never be met because
duplicate ACKs corresponding to at most 9 segments will arrive at the
sender. To offset the issue of loss, we extend TCP’s Limited
Transmit [RFC3042] scheme to allow for the sending of new data during
the period when the TCP sender is disambiguating loss and reordering.
This new data serves to increase the likelihood that enough duplicate
ACKs arrive at the sender to trigger loss recovery if it is
appropriate.
Note that TCP tightly couples reliability and congestion control:
when a segment is declared lost, a retransmission is triggered, and a
change to the sending rate is also made on the assumption that the
drop is due to resource contention [RFC2581]. Therefore, simply by
changing the retransmission trigger, the congestion control response
is also changed. However, we lack experience on the Internet as to
whether delaying the point that a rate reduction takes place is
appropriate for wide-scale deployment. Therefore, the Extended
Limited Transmit mechanism proposed in this document offers two
variants for experimentation.
The first Extended Limited Transmit variant, Careful Limited
Transmit, calls for the transmission of one previously unsent
segment, in response to duplicate acknowledgments, for every two
segments that are known to have left the network. This effectively
halves the sending rate, since normal TCP operation calls for the
sending of one segment for every segment that has left the network.
Further, the halving starts immediately and is not delayed until a
retransmission is triggered. In the case of packet reordering (i.e.,
not segment loss), the congestion control state is restored to its
previous state when reordering is determined.
The second variant, Aggressive Limited Transmit, calls for
transmitting one previously unsent data segment, in response to
duplicate acknowledgments, for every segment known to have left the
network. With this variant, while waiting to disambiguate the loss
from a reordering event, ACK-clocked transmission continues at
roughly the same rate as before the event started. Retransmission
and the sending rate reduction happen per [RFC2581, RFC3517], albeit
with the delayed threshold described above. Although this approach
delays legitimate rate reductions (possibly slightly and temporarily
aggravating overall congestion on the network), the scheme has the
advantage of not reducing the transmission rate in the face of
segment reordering.
Which of the two Extended Limited Transmit variants is best for use
on the Internet is an open question.
3. Algorithm
The TCP-NCR modifications make two fundamental changes to the way
[RFC3517] currently operates, as follows.
First, the trigger for retransmitting a segment is changed from three
duplicate ACKs [RFC2581, RFC3517] to indications that a congestion
window’s worth of data has left the network. Second, TCP-NCR
decouples initial congestion control decisions from retransmission
decisions, in some cases delaying congestion control changes relative
to TCP’s current behavior as defined in [RFC2581]. The algorithm
provides two alternatives for extending Limited Transmit. The two
variants of extended Limited Transmit are:
Careful Limited Transmit
This variant calls for reducing the sending rate at
approximately the same time [RFC2581] implementations reduce
the congestion window, while at the same time withholding a
retransmission (and the final congestion determination) for
approximately one RTT.
Aggressive Limited Transmit
This variant calls for maintaining the sending rate in the
face of duplicate ACKs until TCP concludes that a segment is
lost and needs to be retransmitted (which TCP-NCR delays by
one RTT when compared with current loss recovery schemes).
A TCP-NCR implementation MUST use either Careful Limited Transmit or
Aggressive Limited Transmit.
A constant MUST be set, depending on which variant of extended
Limited Transmit is used, as follows:
Careful Limited Transmit
LT_F = 2/3
Aggressive Limited Transmit
LT_F = 1/2
This constant reflects the fraction of outstanding data (including
data sent during Extended Limited Transmit) that must be SACKed
before a retransmission is triggered. Since Aggressive Limited
Transmit sends a new segment for every segment known to have left the
network, a total of roughly cwnd segments will be sent during
Aggressive Limited Transmit, and therefore ideally a total of roughly
2*cwnd segments will be outstanding when a retransmission is
triggered. The duplicate ACK threshold is then set to LT_F = 1/2 of
2*cwnd (or about 1 RTT worth of data). The factor is different for
Careful Limited Transmit because the sender only transmits one new
segment for every two segments that are SACKed and therefore will
ideally have a total of 1.5*cwnd segments outstanding when the
retransmission is to be triggered. Hence, the required threshold is
LT_F=2/3 of 1.5*cwnd to delay the retransmission by roughly 1 RTT.
There are situations whereby the sender cannot transmit new data
during Extended Limited Transmit (e.g., lack of data from the
application, receiver’s advertised window limit). These situations
can lead to the problems discussed in the last section when a TCP
does not employ Extended Limited Transmit and is starved for ACKs.
Therefore, TCP-NCR adapts the duplicate ACK threshold on each SACK
arrival to be as robust as possible given the actual amount of data
that has been transmitted, or roughly LT_F times the number of
outstanding segments.
The TCP-NCR modifications specified in this document lend themselves
to incremental deployment. Only the TCP implementation on the sender
side requires modification (assuming both hosts support SACK). The
changes themselves are modest. However, as will be discussed below,
availability of additional buffer space at the receiver will help
maximize the benefits of using TCP-NCR but is not strictly necessary.
The following algorithms depend on the notions provided by [RFC3517],
and we assume the reader is familiar with the terminology given in
[RFC3517]. The TCP-NCR algorithm can be adapted to alternate SACK-
based loss recovery schemes. [BR04, BSRV04] outline non-SACK-based
algorithms; however, we do not specify those algorithms in this
document and do not recommend them due to both the complexity and
security implications of having only a gross understanding of the
number of outstanding segments in the network.
A TCP connection using the Nagle algorithm [RFC896, RFC1122] MAY
employ the TCP-NCR algorithm. If a TCP implementation does implement
TCP-NCR, the implementation MUST follow the various specifications
provided in Sections 3.1 - 3.4. If the Nagle algorithm is not being
used, there is no way to accurately calculate the number of
outstanding segments in the network (and, therefore, no good way to
derive an appropriate duplicate ACK threshold) without adding state
to the TCP sender. A TCP connection that does not employ the Nagle
algorithm SHOULD NOT use TCP-NCR. We envision that NCR could be
adapted to an implementation that carefully tracks the sequence
numbers transmitted in each segment. However, we leave this as
future work.
3.1. Initialization
When entering a period of loss/reordering detection and Extended
Limited Transmit, a TCP-NCR MUST initialize several state variables.
A TCP MUST enter Extended Limited Transmit upon receiving the first
ACK with a SACK block after the reception of an ACK that (a) did not
contain SACK information and (b) did increase the connection’s
cumulative ACK point. The initializations are:
(I.1) The TCP MUST save the current FlightSize.
FlightSizePrev = FlightSize
(I.2) The TCP MUST set a variable for tracking the number of
segments for which an ACK does not trigger a transmission
during Careful Limited Transmit.
Skipped = 0
(Note: Skipped is not used during Aggressive Limited
Transmit.)
(I.3) The TCP MUST set DupThresh (from [RFC3517]) based on the
current FlightSize.
DupThresh = max (LT_F * (FlightSize / SMSS),3)
Note: We keep the lower bound of DupThresh = 3 from
[RFC2581, RFC3517].
In addition to the above steps, the incoming ACK MUST be processed
with the E series of steps in Section 3.3.
3.2. Terminating Extended Limited Transmit and Preventing Bursts
Extended Limited Transmit MUST be terminated at the start of loss
recovery as outlined in Section 3.4.
The arrival of an ACK that advances the cumulative ACK point while in
Extended Limited Transmit, but before loss recovery is triggered,
signals that a series of duplicate ACKs was caused by reordering and
not congestion. Therefore, the receipt of an ACK that extends the
cumulative ACK point MUST terminate Extended Limited Transmit. As
described below (in (T.4)), an ACK that extends the cumulative ACK
point and *also* contains SACK information will also trigger the
beginning of a new Extended Limited Transmit phase.
Upon the termination of Extended Limited Transmit, and especially
when using the Careful variant, TCP-NCR may be in a situation where
the entire cwnd is not being utilized, and therefore TCP-NCR will be
prone to transmitting a burst of segments into the network.
Therefore, to mitigate this bursting when a TCP-NCR in the Extended
Limited Transmit phase receives an ACK that updates the cumulative
ACK point (regardless of whether the ACK contains SACK information),
the following steps MUST be taken:
(T.1) A TCP MUST reset cwnd to:
cwnd = min (FlightSize + SMSS,FlightSizePrev)
This step ensures that cwnd is not grossly larger than the
amount of data outstanding, a situation that would cause a
line rate burst.
(T.2) A TCP MUST set ssthresh to:
ssthresh = FlightSizePrev
This step provides TCP-NCR with a sense of "history". If step
(T.1) reduces cwnd below FlightSizePrev, this step ensures that
TCP-NCR will slow start back to the operating point in effect
before Extended Limited Transmit.
(T.3) A TCP is now permitted to transmit previously unsent data as
allowed by cwnd, FlightSize, application data availability, and
the receiver’s advertised window.
(T.4) When an incoming ACK extends the cumulative ACK point and also
contains SACK information, the initializations in steps (I.2)
and (I.3) from Section 3.1 MUST be taken (but step (I.1) MUST
NOT be executed) to re-start Extended Limited Transmit. In
addition, the series of steps in Section 3.3 (the "E" steps)
MUST be taken.
3.3. Extended Limited Transmit
On each ACK containing SACK information that arrives after TCP-NCR
has entered the Extended Limited Transmit phase (as outlined in
Section 3.1) and before Extended Limited Transmit terminates, the
sender MUST use the following procedure.
(E.1) The SetPipe () procedure from [RFC3517] MUST be used to set
the "pipe" variable (which represents the number of bytes
still considered "in the network"). Note: the current value
of DupThresh MUST be used by SetPipe () to produce an accurate
assessment of the amount of data still considered in the
network.
(E.2) If the comparison in equation (1), below, holds and there are
SMSS bytes of previously unsent data available for
transmission, then the sender MUST transmit one segment of SMSS
bytes.
(pipe + Skipped) <= (FlightSizePrev - SMSS) (1)
If the comparison in equation (1) does not hold or no new data
can be transmitted (due to lack of data from the application
or the advertised window limit), skip to step (E.6).
(E.3) Pipe MUST be incremented by SMSS bytes.
(E.4) If using Careful Limited Transmit, Skipped MUST be incremented
by SMSS bytes to ensure that the next SMSS bytes of SACKed data
processed does not trigger a Limited Transmit transmission
(since the goal of Careful Limited Transmit is to send upon
receipt of every second duplicate ACK).