Forward RTO-Recovery (F-RTO): An Algorithm for Detecting Spurious Retransmission Timeouts with TCP

Network Working Group                                       P. Sarolahti
Request for Comments: 5682                         Nokia Research Center
Updates: 4138                                                    M. Kojo
Category: Standards Track                         University of Helsinki
                                                             K. Yamamoto
                                                                 M. Hata
                                                              NTT Docomo
                                                          September 2009


        Forward RTO-Recovery (F-RTO): An Algorithm for Detecting
               Spurious Retransmission Timeouts with TCP

Abstract

   The purpose of this document is to move the F-RTO (Forward
   RTO-Recovery) functionality for TCP in RFC 4138 from
   Experimental to Standards Track status.  The F-RTO support for Stream
   Control Transmission Protocol (SCTP) in RFC 4138 remains with
   Experimental status.  See Appendix B for the differences between this
   document and RFC 4138.

   Spurious retransmission timeouts cause suboptimal TCP performance
   because they often result in unnecessary retransmission of the last
   window of data.  This document describes the F-RTO detection
   algorithm for detecting spurious TCP retransmission timeouts.  F-RTO
   is a TCP sender-only algorithm that does not require any TCP options
   to operate.  After retransmitting the first unacknowledged segment
   triggered by a timeout, the F-RTO algorithm of the TCP sender
   monitors the incoming acknowledgments to determine whether the
   timeout was spurious.  It then decides whether to send new segments
   or retransmit unacknowledged segments.  The algorithm effectively
   helps to avoid additional unnecessary retransmissions and thereby
   improves TCP performance in the case of a spurious timeout.

Status of This Memo

   This document specifies an Internet standards track protocol for the
   Internet community, and requests discussion and suggestions for
   improvements.  Please refer to the current edition of the "Internet
   Official Protocol Standards" (STD 1) for the standardization state
   and status of this protocol.  Distribution of this memo is unlimited.

Copyright and License Notice

   Copyright (c) 2009 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
   (http://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 BSD License.

Table of Contents

   1. Introduction ....................................................3
      1.1. Conventions and Terminology ................................5
   2. Basic F-RTO Algorithm ...........................................5
      2.1. The Algorithm ..............................................5
      2.2. Discussion .................................................7
   3. SACK-Enhanced Version of the F-RTO Algorithm ....................9
      3.1. The Algorithm ..............................................9
      3.2. Discussion ................................................11
   4. Taking Actions after Detecting Spurious RTO ....................11
   5. Evaluation of RFC 4138 .........................................12
   6. Security Considerations ........................................13
   7. Acknowledgments ................................................14
   Appendix A. Discussion of Window-Limited Cases ....................15
   Appendix B. Changes since RFC 4138 ................................16
   References ........................................................16
      Normative References ...........................................16
      Informative References .........................................17

1.  Introduction

   The Transmission Control Protocol (TCP) [Pos81] has two methods for
   triggering retransmissions.  First, the TCP sender relies on incoming
   duplicate acknowledgments (ACKs), which indicate that the receiver is
   missing some of the data.  After a required number of successive
   duplicate ACKs have arrived at the sender, it retransmits the first
   unacknowledged segment [APB09] and continues with a loss recovery
   algorithm such as NewReno [FHG04] or SACK-based (Selective
   Acknowledgment) loss recovery [BAFW03].  Second, the TCP sender
   maintains a retransmission timer that triggers retransmission of
   segments, if they have not been acknowledged before the
   retransmission timeout (RTO) occurs.  When the retransmission timeout
   occurs, the TCP sender enters the RTO recovery where the congestion
   window is initialized to one segment and unacknowledged segments are
   retransmitted using the slow-start algorithm.  The retransmission
   timer is adjusted dynamically, based on the measured round-trip times
   [PA00].

   It has been pointed out that the retransmission timer can expire
   spuriously and cause unnecessary retransmissions when no segments
   have been lost [LK00, GL02, LM03].  After a spurious retransmission
   timeout, the late acknowledgments of the original segments arrive at
   the sender, usually triggering unnecessary retransmissions of a whole
   window of segments during the RTO recovery.  Furthermore, after a
   spurious retransmission timeout, a conventional TCP sender increases
   the congestion window on each late acknowledgment in slow start.
   This injects a large number of data segments into the network within
   one round-trip time, thus violating the packet conservation principle
   [Jac88].

   There are a number of potential reasons for spurious retransmission
   timeouts.  First, some mobile networking technologies involve sudden
   delay spikes on transmission because of actions taken during a hand-
   off.  Second, a hand-off may take place from a low latency path to a
   high latency path, suddenly increasing the round-trip time beyond the
   current RTO value.  Third, on a low-bandwidth link the arrival of
   competing traffic (possibly with higher priority), or some other
   change in available bandwidth, can cause a sudden increase of the
   round-trip time.  This may trigger a spurious retransmission timeout.
   A persistently reliable link layer can also cause a sudden delay when
   a data frame and several retransmissions of it are lost for some
   reason.  This document does not distinguish between the different
   causes of such a delay spike.  Rather, it discusses the spurious
   retransmission timeouts caused by a delay spike in general.

   This document describes the F-RTO detection algorithm for TCP.  It is
   based on the detection mechanism of the "Forward RTO-Recovery"
   (F-RTO) algorithm [SKR03] that is used for detecting spurious
   retransmission timeouts and thus avoids unnecessary retransmissions
   following the retransmission timeout.  When the timeout is not
   spurious, the F-RTO algorithm reverts back to the conventional RTO
   recovery algorithm, and therefore has similar behavior and
   performance.  In contrast to alternative algorithms proposed for
   detecting unnecessary retransmissions (Eifel [LK00, LM03] and DSACK-
   based (Duplicate SACK) algorithms [BA04]), F-RTO does not require any
   TCP options for its operation, and it can be implemented by modifying
   only the TCP sender.  The Eifel algorithm uses TCP timestamps [BBJ92]
   for detecting a spurious timeout upon arrival of the first
   acknowledgment after the retransmission.  The DSACK-based algorithms
   require that the TCP Selective Acknowledgment Option [MMFR96], with
   the DSACK extension [FMMP00], is in use.  With DSACK, the TCP
   receiver can report if it has received a duplicate segment, enabling
   the sender to detect afterwards whether it has retransmitted segments
   unnecessarily.  The F-RTO algorithm only attempts to detect and avoid
   unnecessary retransmissions after an RTO.  Eifel and DSACK can also
   be used for detecting unnecessary retransmissions caused by other
   events, such as packet reordering.

   When the retransmission timer expires, the F-RTO sender retransmits
   the first unacknowledged segment as usual [APB09].  Deviating from
   the normal operation after a timeout, it then tries to transmit new,
   previously unsent data for the first acknowledgment that arrives
   after the timeout, given that the acknowledgment advances the window.
   If the second acknowledgment that arrives after the timeout advances
   the window (i.e., acknowledges data that was not retransmitted), the
   F-RTO sender declares the timeout spurious and exits the RTO
   recovery.  However, if either of these two acknowledgments is a
   duplicate ACK, there will not be sufficient evidence of a spurious
   timeout.  Therefore, the F-RTO sender retransmits the unacknowledged
   segments in slow start similar to the traditional algorithm.  With a
   SACK-enhanced version of the F-RTO algorithm, spurious timeouts may
   be detected even if duplicate ACKs arrive after an RTO
   retransmission.

   This document specifies the F-RTO algorithm for TCP only, replacing
   the F-RTO functionality with TCP in RFC 4138 [SK05] and moving it
   from Experimental to Standards Track status.  The algorithm can also
   be applied to the Stream Control Transmission Protocol (SCTP) [Ste07]
   that has acknowledgment and packet retransmission concepts similar to
   TCP.  The considerations on applying F-RTO to SCTP are discussed in
   RFC 4138, but the F-RTO support for SCTP remains with Experimental
   status.

   This document is organized as follows.  Section 2 describes the basic
   F-RTO algorithm, and the SACK-enhanced F-RTO algorithm is given in
   Section 3.  Section 4 discusses the possible actions to be taken
   after detecting a spurious RTO.  Section 5 summarizes the experience
   with F-RTO implementations and the experimental results, and Section
   6 discusses the security considerations.

1.1.  Conventions and 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 BCP 14, RFC 2119
   [RFC2119] and indicate requirement levels for protocols.

2.  Basic F-RTO Algorithm

   A timeout is considered spurious if it would have been avoided had
   the sender waited longer for an acknowledgment to arrive [LM03].
   F-RTO affects the TCP sender behavior only after a retransmission
   timeout.  Otherwise, the TCP behavior remains the same.  When the
   retransmission timer expires, the F-RTO algorithm monitors incoming
   acknowledgments, and if the TCP sender gets an acknowledgment for a
   segment that was not retransmitted due to the timeout, the F-RTO
   algorithm declares a timeout spurious.  The actions taken in response
   to a spurious timeout are not specified in this document, but we
   discuss some alternatives in Section 4.  This section introduces the
   algorithm and then discusses the different steps of the algorithm in
   more detail.

   Following the practice used with the Eifel Detection algorithm
   [LM03], we use the "SpuriousRecovery" variable to indicate whether
   the retransmission is declared spurious by the sender.  This variable
   can be used as an input for a corresponding response algorithm.  With
   F-RTO, the value of SpuriousRecovery can be either SPUR_TO
   (indicating a spurious retransmission timeout) or FALSE (indicating
   that the timeout is not declared spurious and the TCP sender should
   follow the conventional RTO recovery algorithm).  In addition, we use
   the "recover" variable specified in the NewReno algorithm [FHG04].

2.1.  The Algorithm

   A TCP sender implementing the basic F-RTO algorithm MUST take the
   following steps after the retransmission timer expires.  If the
   retransmission timer expires again during the execution of the F-RTO
   algorithm, the TCP sender MUST re-start the algorithm processing from
   step 1.  If the sender implements some loss recovery algorithm other
   than Reno or NewReno [FHG04], the F-RTO algorithm SHOULD NOT be
   entered when earlier fast recovery is underway.

   The F-RTO algorithm takes different actions based on whether an
   incoming acknowledgment advances the cumulative acknowledgment point
   for a received in-order segment, or whether it is a duplicate
   acknowledgment to indicate an out-of-order segment.  Duplicate
   acknowledgment is defined in [APB09].  The F-RTO algorithm does not
   specify actions for receiving a segment that neither acknowledges new
   data nor is a duplicate acknowledgment.  The TCP sender SHOULD ignore
   such segments and wait for a segment that either acknowledges new
   data or is a duplicate acknowledgment.

   1) When the retransmission timer expires, retransmit the first
      unacknowledged segment and set SpuriousRecovery to FALSE.  If the
      TCP sender is already in RTO recovery AND "recover" is larger than
      or equal to SND.UNA (the oldest unacknowledged sequence number
      [Pos81]), do not enter step 2 of this algorithm.  Instead, store
      the highest sequence number transmitted so far in variable
      "recover" and continue with slow-start retransmissions following
      the conventional RTO recovery algorithm.

   2) When the first acknowledgment after the RTO retransmission arrives
      at the TCP sender, store the highest sequence number transmitted
      so far in variable "recover".  The TCP sender chooses one of the
      following actions, depending on whether the ACK advances the
      window or whether it is a duplicate ACK.

      a) If the acknowledgment is a duplicate ACK, OR the Acknowledgment
         field covers "recover" but not more than "recover", OR the
         acknowledgment does not acknowledge all of the data that was
         retransmitted in step 1, revert to the conventional RTO
         recovery and continue by retransmitting unacknowledged data in
         slow start.  Do not enter step 3 of this algorithm.  The
         SpuriousRecovery variable remains as FALSE.

      b) Else, if the acknowledgment advances the window AND the
         Acknowledgment field does not cover "recover", transmit up to
         two new (previously unsent) segments and enter step 3 of this
         algorithm.  If the TCP sender does not have enough unsent data,
         it can send only one segment.  In addition, the TCP sender MAY
         override the Nagle algorithm [Nag84] and immediately send a
         segment if needed.  Note that sending two segments in this step
         is allowed by TCP congestion control requirements [APB09]: an
         F-RTO TCP sender simply chooses different segments to transmit.

         If the TCP sender does not have any new data to send, or the
         advertised window prohibits new transmissions, the recommended
         action is to skip step 3 of this algorithm and continue with
         slow-start retransmissions, following the conventional RTO
         recovery algorithm.  However, alternative ways of handling the
         window-limited cases that could result in better performance
         are discussed in Appendix A.

   3) When the second acknowledgment after the RTO retransmission
      arrives at the TCP sender, the TCP sender either declares the
      timeout spurious, or starts retransmitting the unacknowledged
      segments.


      a) If the acknowledgment is a duplicate ACK, set the congestion
         window to no more than 3 * MSS (where MSS indicates Maximum
         Segment Size), and continue with the slow-start algorithm
         retransmitting unacknowledged segments.  The congestion window
         can be set to 3 * MSS, because two round-trip times have
         elapsed since the RTO, and a conventional TCP sender would have
         increased cwnd to 3 during the same time.  Leave
         SpuriousRecovery set to FALSE.

      b) If the acknowledgment advances the window (i.e., if it
         acknowledges data that was not retransmitted after the
         timeout), declare the timeout spurious, set SpuriousRecovery to
         SPUR_TO, and set the value of the "recover" variable to SND.UNA
         (the oldest unacknowledged sequence number [Pos81]).

2.2.  Discussion

   The F-RTO sender takes cautious actions when it receives duplicate
   acknowledgments after a retransmission timeout.  Because duplicate
   ACKs may indicate that segments have been lost, reliably detecting a
   spurious timeout is difficult due to the lack of additional
   information.  Therefore, it is prudent to follow the conventional TCP
   recovery in those cases.

   The condition in step 1 prevents the execution of the F-RTO algorithm
   in case a previous RTO recovery is underway when the retransmission
   timer expires, except in case the retransmission timer expires
   multiple times for the same segment.  If the retransmission timer
   expires during an earlier RTO-based loss recovery, acknowledgments
   for retransmitted segments may falsely lead the TCP sender to declare
   the timeout spurious.

   If the first acknowledgment after the RTO retransmission covers the
   "recover" point at algorithm step (2a), there is not enough evidence
   that a non-retransmitted segment has arrived at the receiver after
   the timeout.  This is a common case when a fast retransmission is
   lost and has been retransmitted again after an RTO, while the rest of
   the unacknowledged segments were successfully delivered to the TCP
   receiver before the retransmission timeout.  Therefore, the timeout
   cannot be declared spurious in this case.

   If the first acknowledgment after the RTO retransmission does not
   acknowledge all of the data that was retransmitted in step 1, the TCP
   sender reverts to the conventional RTO recovery.  Otherwise, a
   malicious receiver acknowledging partial segments could cause the
   sender to declare the timeout spurious in a case where data was lost.

   The TCP sender is allowed to send two new segments in algorithm
   branch (2b) because the conventional TCP sender would transmit two
   segments when the first new ACK arrives after the RTO retransmission.
   If sending new data is not possible in algorithm branch (2b), or if
   the receiver window limits the transmission, the TCP sender has to
   send something in order to prevent the TCP transfer from stalling.
   If no segments were sent, the pipe between sender and receiver might
   run out of segments, and no further acknowledgments would arrive.
   Therefore, in the window-limited case, the recommendation is to
   revert to the conventional RTO recovery with slow-start
   retransmissions.  Appendix A discusses some alternative solutions for
   window-limited situations.

   If the retransmission timeout is declared spurious, the TCP sender
   sets the value of the "recover" variable to SND.UNA in order to allow
   fast retransmit [FHG04].  The "recover" variable was proposed for
   avoiding unnecessary, multiple fast retransmits when the
   retransmission timer expires during fast recovery with NewReno TCP.
   Because the F-RTO sender retransmits only the segment that triggered
   the timeout, the problem of unnecessary multiple fast retransmits
   [FHG04] cannot occur.  Therefore, if three duplicate ACKs arrive at
   the sender after the timeout, they probably indicate a packet loss,
   and thus fast retransmit should be used to allow efficient recovery.
   If there are not enough duplicate ACKs arriving at the sender after a
   packet loss, the retransmission timer expires again and the sender
   enters step 1 of this algorithm.

   When the timeout is declared spurious, the TCP sender cannot detect
   whether the unnecessary RTO retransmission was lost.  In principle,
   the loss of the RTO retransmission should be taken as a congestion
   signal.  Thus, there is a small possibility that the F-RTO sender
   will violate the congestion control rules, if it chooses to fully
   revert congestion control parameters after detecting a spurious
   timeout.  The Eifel Detection algorithm has a similar property, while
   the DSACK option can be used to detect whether the retransmitted
   segment was successfully delivered to the receiver.

   The F-RTO algorithm has a side effect on the TCP round-trip time
   measurement.  Because the TCP sender can avoid most of the
   unnecessary retransmissions after detecting a spurious timeout, the
   sender is able to take round-trip time samples on the delayed
   segments.  If the regular RTO recovery was used without TCP
   timestamps, this would not be possible due to the retransmission
   ambiguity.  As a result, the RTO is likely to have more accurate and
   larger values with F-RTO than with the regular TCP after a spurious
   timeout that was triggered due to delayed segments.  We believe this
   is an advantage in networks that are prone to delay spikes.

   There are some situations where the F-RTO algorithm may not avoid
   unnecessary retransmissions after a spurious timeout.  If packet
   reordering or packet duplication occurs on the segment that triggered
   the spurious timeout, the F-RTO algorithm may not detect the spurious
   timeout due to incoming duplicate ACKs.  Additionally, if a spurious
   timeout occurs during fast recovery, the F-RTO algorithm often cannot
   detect the spurious timeout because the segments that were
   transmitted before the fast recovery trigger duplicate ACKs.
   However, we consider these cases rare, and note that in cases where
   F-RTO fails to detect the spurious timeout, it retransmits the
   unacknowledged segments in slow start, and thus performs the same as
   the regular RTO recovery.

3.  SACK-Enhanced Version of the F-RTO Algorithm

   This section describes an alternative version of the F-RTO algorithm
   that uses the TCP Selective Acknowledgment Option [MMFR96].  By using
   the SACK option, the TCP sender detects spurious timeouts in most of
   the cases when packet reordering or packet duplication is present.
   If the SACK information acknowledges new data that was not
   transmitted after the RTO retransmission, the sender may declare the
   timeout spurious, even when duplicate ACKs follow the RTO.

3.1.  The Algorithm

   Given that the TCP Selective Acknowledgment Option [MMFR96] is
   enabled for a TCP connection, a TCP sender MAY apply the SACK-
   enhanced F-RTO algorithm.  If the sender applies the SACK-enhanced
   F-RTO algorithm, it MUST follow the steps below.  This algorithm
   SHOULD NOT be applied if the TCP sender is already in loss recovery
   when a retransmission timeout occurs.

   The steps of the SACK-enhanced version of the F-RTO algorithm are as
   follows.  If the retransmission timer expires again during the
   execution of the SACK-enhanced F-RTO algorithm, the TCP sender MUST
   re-start the algorithm processing from step 1.

   1) When the retransmission timer expires, retransmit the first
      unacknowledged segment and set SpuriousRecovery to FALSE.
      Following the recommendation in the SACK specification [MMFR96],
      reset the SACK scoreboard.  If "RecoveryPoint" is larger than or
      equal to SND.UNA, do not enter step 2 of this algorithm.  Instead,
      set variable "RecoveryPoint" to indicate the highest sequence
      number transmitted so far and continue with slow-start
      retransmissions following the conventional RTO recovery algorithm.

   2) Wait until the acknowledgment of the data retransmitted due to the
      timeout arrives at the sender.  If duplicate ACKs arrive before
      the cumulative acknowledgment for retransmitted data, adjust the
      scoreboard according to the incoming SACK information.  Stay in
      step 2 and wait for the next new acknowledgment.  If the
      retransmission timeout expires again, go to step 1 of the
      algorithm.  When a new acknowledgment arrives, set variable
      "RecoveryPoint" to indicate the highest sequence number
      transmitted so far.

      a) If the Cumulative Acknowledgment field covers "RecoveryPoint"
         but not more than "RecoveryPoint", revert to the conventional
         RTO recovery and set the congestion window to no more than 2 *
         MSS, like a regular TCP would do.  Do not enter step 3 of this
         algorithm.

      b) Else, if the Cumulative Acknowledgment field does not cover
         "RecoveryPoint" but is larger than SND.UNA, transmit up to two
         new (previously unsent) segments and proceed to step 3.  If the
         TCP sender is not able to transmit any previously unsent data
         -- either due to receiver window limitation or because it does
         not have any new data to send -- the recommended action is to
         refrain from entering step 3 of this algorithm.  Rather,
         continue with slow-start retransmissions following the
         conventional RTO recovery algorithm.

         It is also possible to apply some of the alternatives for
         handling window-limited cases discussed in Appendix A.

   3) The next acknowledgment arrives at the sender.  Either a duplicate
      ACK or a new cumulative ACK (advancing the window) applies in this
      step.  Other types of ACKs are ignored without any action.

      a) If the Cumulative Acknowledgment field or the SACK information
         covers more than "RecoveryPoint", set the congestion window to
         no more than 3 * MSS and proceed with the conventional RTO
         recovery, retransmitting unacknowledged segments.  Take this
         branch also when the acknowledgment is a duplicate ACK and it
         does not acknowledge any new, previously unacknowledged data
         below "RecoveryPoint" in the SACK information.  Leave
         SpuriousRecovery set to FALSE.

      b) If the Cumulative Acknowledgment field or a SACK information in
         the ACK does not cover more than "RecoveryPoint" AND it
         acknowledges data that was not acknowledged earlier (either
         with cumulative acknowledgment or using SACK information),
         declare the timeout spurious and set SpuriousRecovery to
         SPUR_TO.  The retransmission timeout can be declared spurious,
         because the segment acknowledged with this ACK was transmitted
         before the timeout.

   If there are unacknowledged holes between the received SACK
   information, those segments are retransmitted similarly to the
   conventional SACK recovery algorithm [BAFW03].  If the algorithm
   exits with SpuriousRecovery set to SPUR_TO, "RecoveryPoint" is set to
   SND.UNA, thus allowing fast recovery on incoming duplicate
   acknowledgments.

3.2.  Discussion

   The SACK-enhanced algorithm works on the same principle as the basic
   algorithm, but by utilizing the additional information from the SACK
   option.  When a genuine retransmission timeout occurs during a steady
   state of a connection, it can be assumed that there are no segments
   left in the pipe.  Otherwise, the acknowledgments triggered by these
   segments would have triggered the SACK loss recovery or transmission
   of new segments.  Therefore, if the F-RTO sender receives
   acknowledgments for segments transmitted before the retransmission
   timeout in response to the two new segments sent at the algorithm
   step 2, the normal operation of TCP has been just delayed, and the
   retransmission timeout is considered spurious.  Note that this
   reasoning works only when the TCP sender is not in loss recovery at
   the time the retransmission timeout occurs.  The condition in step 1
   checking that "RecoveryPoint" is larger than or equal to SND.UNA
   prevents the execution of the F-RTO algorithm in case a previous loss
   recovery, either RTO recovery or SACK loss recovery, is underway when
   the retransmission timer expires.  It, however, allows the execution
   of the F-RTO algorithm, if the retransmission timer expires multiple
   times for the same segment.

4.  Taking Actions after Detecting Spurious RTO

   Upon a retransmission timeout, a conventional TCP sender assumes that
   outstanding segments are lost and starts retransmitting the
   unacknowledged segments.  When the retransmission timeout is detected
   to be spurious, the TCP sender should not continue retransmitting
   based on the timeout.  For example, if the sender was in congestion
   avoidance phase transmitting new, previously unsent segments, it
   should continue transmitting previously unsent segments in congestion
   avoidance.

   There are currently two alternatives specified for a spurious timeout
   response algorithm, the Eifel Response Algorithm [LG05], and an
   algorithm for adapting the retransmission timeout after a spurious
   RTO [BBA06].  If no specific response algorithm is implemented, the
   TCP SHOULD respond to spurious timeout conservatively, applying the
   TCP congestion control specification [APB09].  Different response
   algorithms for spurious retransmission timeouts have been analyzed in
   some research papers [GL03, Sar03] and IETF documents [SL03].

5.  Evaluation of RFC 4138

   F-RTO was first specified in an Experimental RFC (RFC 4138) that has
   been implemented in a number of operating systems since it was
   published.  Gained experience has been documented in a separate
   document [KYHS07], and can be summarized as follows.

   If the TCP sender employs F-RTO, it is able to detect spurious RTOs
   and avoid the unnecessary retransmission of the whole window of data.
   Because F-RTO avoids the unnecessary retransmissions after a spurious
   RTO, it is able to adhere to the packet conservation principle,
   unlike a regular TCP that enters the slow-start recovery
   unnecessarily and inappropriately restarts the ACK clock while there
   are segments outstanding in the network.  When a spurious RTO has
   been detected, a sender can select an appropriate congestion control
   response instead of setting the congestion window to one segment.
   Because F-RTO avoids unnecessary retransmissions, it is able to take
   the round-trip time of the delayed segments into account when
   calculating the RTO estimate, which may help in avoiding further
   spurious retransmission timeouts.

   Experimental results with the basic F-RTO have been reported in an
   emulated network using a Linux implementation [SKR03].  Also,
   different congestion control responses along with the SACK-enhanced
   version of F-RTO were tested in a similar environment [Sar03].  There
   are publications analyzing F-RTO performance over commercial Wideband
   Code Division Multiple Access (W-CDMA) networks, and in an emulated
   High-Speed Downlink Packet Access (HSDPA) network [Yam05, Hok05].
   Also, Microsoft reported positive experiences with their
   implementation of F-RTO at the IETF-68 meeting.

   It is known that some spurious RTOs may remain undetected by F-RTO if
   duplicate acknowledgments arrive at the sender immediately after the
   spurious RTO, for example due to packet reordering or packet loss.
   There are rare corner cases where F-RTO could "hide" a packet loss
   and therefore lead to inappropriate behavior with non-conservative
   congestion control response: first, if a massive packet reordering
   occurred so that the acknowledgment of RTO retransmission arrived at
   the sender before the acknowledgments of original transmissions, the
   sender might not detect the loss of the segment that triggered the
   RTO.  Second, a malicious receiver could lead F-RTO to make a wrong
   conclusion after an RTO by acknowledging segments it has not
   received.  Such a receiver would, however, risk breaking the
   consistency of the TCP state between the sender and receiver, causing
   the connection to become unusable, which cannot be of any benefit to
   the receiver.  Therefore, we believe it is not likely that receivers
   would start employing such tricks on a significant scale.  Finally,
   loss of the unnecessary RTO retransmission cannot be detected without
   using some explicit acknowledgment scheme such as DSACK.  This is
   common to the other mechanisms for detecting spurious RTO, as well as
   to regular TCP that does not use DSACK.  We note that if the
   congestion control response to spurious RTO is conservative enough,
   the above corner cases do not cause problems due to increased
   congestion.

6.  Security Considerations

   The main security threat regarding F-RTO is the possibility that a
   receiver could mislead the sender into setting too large a congestion
   window after an RTO.  There are two possible ways a malicious
   receiver could trigger a wrong output from the F-RTO algorithm.
   First, the receiver can acknowledge data that it has not received.
   Second, it can delay acknowledgment of a segment it has received
   earlier, and acknowledge the segment after the TCP sender has been
   deluded to enter algorithm step 3.

   If the receiver acknowledges a segment it has not really received,
   the sender can be led to declare spurious timeout in the F-RTO
   algorithm, step 3.  However, because the sender will have an
   incorrect state, it cannot retransmit the segment that has never
   reached the receiver.  Therefore, this attack is unlikely to be
   useful for the receiver to maliciously gain a larger congestion
   window.

   A common case for a retransmission timeout is that a fast
   retransmission of a segment is lost.  If all other segments have been
   received, the RTO retransmission causes the whole window to be
   acknowledged at once.  This case is recognized in F-RTO algorithm
   branch (2a).  However, if the receiver only acknowledges one segment
   after receiving the RTO retransmission, and then the rest of the
   segments, it could cause the timeout to be declared spurious when it
   is not.  Therefore, it is suggested that, when an RTO occurs during
   the fast recovery phase, the sender would not fully revert the
   congestion window even if the timeout was declared spurious.
   Instead, the sender would reduce the congestion window to 1.

   If there is more than one segment missing at the time of a
   retransmission timeout, the receiver does not benefit from misleading
   the sender to declare a spurious timeout because the sender would
   have to go through another recovery period to retransmit the missing
   segments, usually after an RTO has elapsed.

7.  Acknowledgments

   The authors would like to thank Alfred Hoenes, Ilpo Jarvinen, and
   Murari Sridharan for the comments on this document.

   We are also thankful to Reiner Ludwig, Andrei Gurtov, Josh Blanton,
   Mark Allman, Sally Floyd, Yogesh Swami, Mika Liljeberg, Ivan Arias
   Rodriguez, Sourabh Ladha, Martin Duke, Motoharu Miyake, Ted Faber,
   Samu Kontinen, and Kostas Pentikousis who gave valuable feedback
   during the preparation of RFC 4138, the precursor of this document.

Appendix A.  Discussion of Window-Limited Cases

   When the advertised window limits the transmission of two new
   previously unsent segments, or there are no new data to send, it is
   recommended in F-RTO algorithm step (2b) that the TCP sender continue
   with the conventional RTO recovery algorithm.  The disadvantage is
   that the sender may continue unnecessary retransmissions due to
   possible spurious timeout.  This section briefly discusses the
   options that can potentially improve performance when transmitting
   previously unsent data is not possible.

   - The TCP sender could reserve an unused space of a size of one or
     two segments in the advertised window to ensure the use of
     algorithms such as F-RTO or Limited Transmit [ABF01] in receiver
     window-limited situations.  On the other hand, while doing this,
     the TCP sender should ensure that the window of outstanding
     segments is large enough for proper utilization of the available
     pipe.

   - Use additional information if available, e.g., TCP timestamps with
     the Eifel Detection algorithm, for detecting a spurious timeout.
     However, Eifel detection may yield different results from F-RTO
     when ACK losses and an RTO occur within the same round-trip time
     [SKR03].

   - Retransmit data from the tail of the retransmission queue and
     continue with step 3 of the F-RTO algorithm.  It is possible that
     the retransmission will be made unnecessarily.  Furthermore, the
     operation of the SACK-based F-RTO algorithm would need to consider
     this case separately, to not use the retransmitted segment to
     indicate spurious timeout.  Given these considerations, this option
     is not recommended.

   - Send a zero-sized segment below SND.UNA, similar to a TCP Keep-
     Alive probe, and continue with step 3 of the F-RTO algorithm.
     Because the receiver replies with a duplicate ACK, the sender is
     able to detect whether the timeout was spurious from the incoming
     acknowledgment.  This method does not send data unnecessarily, but
     it delays the recovery by one round-trip time in cases where the
     timeout was not spurious.  Therefore, this method is not
     encouraged.

   - In receiver-limited cases, send one octet of new data, regardless
     of the advertised window limit, and continue with step 3 of the
     F-RTO algorithm.  It is possible that the receiver will have free
     buffer space to receive the data by the time the segment has
     propagated through the network, in which case no harm is done.  If
     the receiver is not capable of receiving the segment, it rejects
     the segment and sends a duplicate ACK.

Appendix B.  Changes since RFC 4138

     Changes from RFC 4138 are summarized below, apart from minor
     editing and language improvements.

   * Modified the basic F-RTO algorithm and the SACK-enhanced F-RTO
     algorithm to prevent the TCP sender from applying the F-RTO
     algorithm if the retransmission timer expires when an earlier RTO
     recovery is underway, except when the retransmission timer expires
     multiple times for the same segment.

   * Clarified behavior on multiple timeouts.

   * Added a paragraph on acknowledgments that do not acknowledge new
     data but are not duplicate acknowledgments.

   * Clarified the SACK-algorithm a bit, and added one paragraph of
     description of the basic idea of the algorithm.

   * Removed SCTP considerations.

   * Removed earlier Appendix sections, except Appendix C from RFC 4138,
     which is now Appendix A.

   * Clarified text about the possible response algorithms.

   * Added section that summarizes the evaluation of RFC 4138.

References

Normative References

   [APB09]   Allman, M., Paxson, V., and E. Blanton, "TCP Congestion
             Control", RFC 5681, September 2009.

   [BAFW03]  Blanton, E., Allman, M., Fall, K., and L. Wang, "A
             Conservative Selective Acknowledgment (SACK)-based Loss
             Recovery Algorithm for TCP", RFC 3517, April 2003.

   [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate
             Requirement Levels", BCP 14, RFC 2119, March 1997.

   [FHG04]   Floyd, S., Henderson, T., and A. Gurtov, "The NewReno
             Modification to TCP's Fast Recovery Algorithm", RFC 3782,
             April 2004.

   [MMFR96]  Mathis, M., Mahdavi, J., Floyd, S., and A. Romanow, "TCP
             Selective Acknowledgment Options", RFC 2018, October 1996.

   [PA00]    Paxson, V. and M. Allman, "Computing TCP's Retransmission
             Timer", RFC 2988, November 2000.

   [Pos81]   Postel, J., "Transmission Control Protocol", STD 7, RFC
             793, September 1981.

Informative References

   [ABF01]   Allman, M., Balakrishnan, H., and S. Floyd, "Enhancing
             TCP's Loss Recovery Using Limited Transmit", RFC 3042,
             January 2001.

   [BA04]    Blanton, E. and M. Allman, "Using TCP Duplicate Selective
             Acknowledgement (DSACKs) and Stream Control Transmission
             Protocol (SCTP) Duplicate Transmission Sequence Numbers
             (TSNs) to Detect Spurious Retransmissions", RFC 3708,
             February 2004.

   [BBA06]   Blanton, J., Blanton, E., and M. Allman, "Using Spurious
             Retransmissions to Adapt the Retransmission Timeout", Work
             in Progress, December 2006.

   [BBJ92]   Jacobson, V., Braden, R., and D. Borman, "TCP Extensions
             for High Performance", RFC 1323, May 1992.

   [FMMP00]  Floyd, S., Mahdavi, J., Mathis, M., and M. Podolsky, "An
             Extension to the Selective Acknowledgement (SACK) Option
             for TCP", RFC 2883, July 2000.

   [GL02]    Gurtov A. and R. Ludwig, "Evaluating the Eifel Algorithm
             for TCP in a GPRS Network", In Proc. European Wireless,
             Florence, Italy, February 2002.

   [GL03]    Gurtov A. and R. Ludwig, "Responding to Spurious Timeouts
             in TCP", In Proc. IEEE INFOCOM 03, San Francisco, CA, USA,
             March 2003.

   [Jac88]   Jacobson, V., "Congestion Avoidance and Control", In Proc.
             ACM SIGCOMM 88.

   [Hok05]   Hokamura, A., et al., "Performance Evaluation of F-RTO and
             Eifel Response Algorithms over W-CDMA packet network", In
             Proc. Wireless Personal Multimedia Communications
             (WPMC'05), Sept. 2005.

   [KYHS07]  Kojo, M., Yamamoto, K., Hata, M., and P. Sarolahti,
             "Evaluation of RFC 4138", Work in Progress, November 2007.

   [LG05]    Ludwig, R. and A. Gurtov, "The Eifel Response Algorithm for
             TCP", RFC 4015, February 2005.

   [LK00]    Ludwig R. and R.H. Katz, "The Eifel Algorithm: Making TCP
             Robust Against Spurious Retransmissions", ACM SIGCOMM
             Computer Communication Review, 30(1), January 2000.

   [LM03]    Ludwig, R. and M. Meyer, "The Eifel Detection Algorithm for
             TCP", RFC 3522, April 2003.

   [Nag84]   Nagle, J., "Congestion control in IP/TCP internetworks",
             RFC 896, January 1984.

   [SK05]    Sarolahti, P. and M. Kojo, "Forward RTO-Recovery (F-RTO):
             An Algorithm for Detecting Spurious Retransmission Timeouts
             with TCP and the Stream Control Transmission Protocol
             (SCTP)", RFC 4138, August 2005.

   [SKR03]   Sarolahti, P., Kojo, M., and K. Raatikainen, "F-RTO: An
             Enhanced Recovery Algorithm for TCP Retransmission
             Timeouts", ACM SIGCOMM Computer Communication Review,
             33(2), April 2003.

   [Sar03]   Sarolahti, P., "Congestion Control on Spurious TCP
             Retransmission Timeouts", In Proc. of IEEE Globecom 2003,
             San Francisco, CA, USA. December 2003.

   [SL03]    Swami Y. and K. Le, "DCLOR: De-correlated Loss Recovery
             using SACK Option for spurious timeouts", Work in Progress,
             September 2003.

   [Ste07]   Stewart, R., Ed., "Stream Control Transmission Protocol",
             RFC 4960, September 2007.

   [Yam05]   Yamamoto, K., et al., "Effects of F-RTO and Eifel Response
             Algorithms for W-CDMA and HSDPA networks", In Proc.
             Wireless Personal Multimedia Communications (WPMC'05),
             September 2005.

Authors' Addresses

   Pasi Sarolahti
   Nokia Research Center
   P.O. Box 407
   FI-00045 NOKIA GROUP
   Finland
   Phone: +358 50 4876607
   EMail: pasi.sarolahti@iki.fi

   Markku Kojo
   University of Helsinki
   P.O. Box 68
   FI-00014 UNIVERSITY OF HELSINKI
   Finland
   Phone: +358 9 19151305
   EMail: kojo@cs.helsinki.fi

   Kazunori Yamamoto
   NTT Docomo, Inc.
   3-5 Hikarinooka, Yokosuka, Kanagawa, 239-8536, Japan
   Phone: +81-46-840-3812
   EMail: yamamotokaz@nttdocomo.co.jp

   Max Hata
   NTT Docomo, Inc.
   3-5 Hikarinooka, Yokosuka, Kanagawa, 239-8536, Japan
   Phone: +81-46-840-3812
   EMail: hatama@s1.nttdocomo.co.jp