rfc2861.txt

RFC 的详细文档！

Network Working Group                                         M. Handley
Request for Comments: 2861                                     J. Padhye
Category: Experimental                                          S. Floyd
                                                                   ACIRI
                                                               June 2000


                    TCP Congestion Window Validation

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 (2000).  All Rights Reserved.

Abstract

   TCP's congestion window controls the number of packets a TCP flow may
   have in the network at any time.  However, long periods when the
   sender is idle or application-limited can lead to the invalidation of
   the congestion window, in that the congestion window no longer
   reflects current information about the state of the network.  This
   document describes a simple modification to TCP's congestion control
   algorithms to decay the congestion window cwnd after the transition
   from a sufficiently-long application-limited period, while using the
   slow-start threshold ssthresh to save information about the previous
   value of the congestion window.

   An invalid congestion window also results when the congestion window
   is increased (i.e., in TCP's slow-start or congestion avoidance
   phases) during application-limited periods, when the previous value
   of the congestion window might never have been fully utilized.  We
   propose that the TCP sender should not increase the congestion window
   when the TCP sender has been application-limited (and therefore has
   not fully used the current congestion window).  We have explored
   these algorithms both with simulations and with experiments from an
   implementation in FreeBSD.

1.  Conventions and Acronyms

   The keywords MUST, MUST NOT, REQUIRED, SHALL, SHALL NOT, SHOULD,
   SHOULD NOT, RECOMMENDED, MAY, and OPTIONAL, when they appear in this
   document, are to be interpreted as described in [B97].



Handley, et al.               Experimental                      [Page 1]

RFC 2861            TCP Congestion Window Validation           June 2000


2. Introduction

   TCP's congestion window controls the number of packets a TCP flow may
   have in the network at any time.  The congestion window is set using
   an Additive-Increase, Multiplicative-Decrease (AIMD) mechanism that
   probes for available bandwidth, dynamically adapting to changing
   network conditions.  This AIMD mechanism works well when the sender
   continually has data to send, as is typically the case for TCP used
   for bulk-data transfer.  In contrast, for TCP used with telnet
   applications, the data sender often has little or no data to send,
   and the sending rate is often determined by the rate at which data is
   generated by the user.  With the advent of the web, including
   developments such as TCP senders with dynamically-created data and
   HTTP 1.1 with persistent-connection TCP, the interaction between
   application-limited periods (when the sender sends less than is
   allowed by the congestion or receiver windows) and network-limited
   periods (when the sender is limited by the TCP window) becomes
   increasingly important.  More precisely, we define a network-limited
   period as any period when the sender is sending a full window of
   data.

   Long periods when the sender is application-limited can lead to the
   invalidation of the congestion window.  During periods when the TCP
   sender is network-limited, the value of the congestion window is
   repeatedly "revalidated" by the successful transmission of a window
   of data without loss.  When the TCP sender is network-limited, there
   is an incoming stream of acknowledgements that "clocks out" new data,
   giving concrete evidence of recent available bandwidth in the
   network.  In contrast, during periods when the TCP sender is
   application-limited, the estimate of available capacity represented
   by the congestion window may become steadily less accurate over time.
   In particular, capacity that had once been used by the network-
   limited connection might now be used by other traffic.

   Current TCP implementations have a range of behaviors for starting up
   after an idle period.  Some current TCP implementations slow-start
   after an idle period longer than the RTO estimate, as suggested in
   [RFC2581] and in the appendix of [VJ88], while other implementations
   don't reduce their congestion window after an idle period.  RFC 2581
   [RFC2581] recommends the following: "a TCP SHOULD set cwnd to no more
   than RW [the initial window] before beginning transmission if the TCP
   has not sent data in an interval exceeding the retransmission
   timeout."  A proposal for TCP's slow-start after idle has also been
   discussed in [HTH98].  The issue of validation of congestion
   information during idle periods has also been addressed in contexts
   other than TCP and IP, for example in "Use-it or Lose-it" mechanisms
   for ATM networks [J96,J95].




Handley, et al.               Experimental                      [Page 2]

RFC 2861            TCP Congestion Window Validation           June 2000


   To address the revalidation of the congestion window after a
   application-limited period, we propose a simple modification to TCP's
   congestion control algorithms to decay the congestion window cwnd
   after the transition from a sufficiently-long application-limited
   period (i.e., at least one roundtrip time) to a network-limited
   period.  In particular, we propose that after an idle period, the TCP
   sender should reduce its congestion window by half for every RTT that
   the flow has remained idle.

   When the congestion window is reduced, the slow-start threshold
   ssthresh remains as "memory" of the recent congestion window.
   Specifically, ssthresh is never decreased when cwnd is reduced after
   an application-limited period; before cwnd is reduced, ssthresh is
   set to the maximum of its current value, and half-way between the old
   and the new values of cwnd.  This use of ssthresh allows a TCP sender
   increasing its sending rate after an application-limited period to
   quickly slow-start to recover most of the previous value of the
   congestion window.  To be more precise, if ssthresh is less than 3/4
   cwnd when the congestion window is reduced after an application-
   limited period, then ssthresh is increased to 3/4 cwnd before the
   reduction of the congestion window.

   An invalid congestion window also results when the congestion window
   is increased (i.e., in TCP's slow-start or congestion avoidance
   phases) during application-limited periods, when the previous value
   of the congestion window might never have been fully utilized.  As
   far as we know, all current TCP implementations increase the
   congestion window when an acknowledgement arrives, if allowed by the
   receiver's advertised window and the slow-start or congestion
   avoidance window increase algorithm, without checking to see if the
   previous value of the congestion window has in fact been used.  This
   document proposes that the window increase algorithm not be invoked
   during application-limited periods [MSML99].  In particular, the TCP
   sender should not increase the congestion window when the TCP sender
   has been application-limited (and therefore has not fully used the
   current congestion window).  This restriction prevents the congestion
   window from growing arbitrarily large, in the absence of evidence
   that the congestion window can be supported by the network.  From
   [MSML99, Section 5.2]: "This restriction assures that [cwnd] only
   grows as long as TCP actually succeeds in injecting enough data into
   the network to test the path."

   A somewhat-orthogonal problem associated with maintaining a large
   congestion window after an application-limited period is that the
   sender, with a sudden large amount of data to send after a quiescent
   period, might immediately send a full congestion window of back-to-
   back packets.  This problem of sending large bursts of packets back-
   to-back can be effectively handled using rate-based pacing (RBP,



Handley, et al.               Experimental                      [Page 3]

RFC 2861            TCP Congestion Window Validation           June 2000


   [VH97]), or using a maximum burst size control [FF96].  We would
   contend that, even with mechanisms for limiting the sending of back-
   to-back packets or pacing packets out over the period of a roundtrip
   time, an old congestion window that has not been fully used for some
   time can not be trusted as an indication of the bandwidth currently
   available for that flow.  We would contend that the mechanisms to
   pace out packets allowed by the congestion window are largely
   orthogonal to the algorithms used to determine the appropriate size
   of the congestion window.

3. Description

   When a TCP sender has sufficient data available to fill the available
   network capacity for that flow, cwnd and ssthresh get set to
   appropriate values for the network conditions.  When a TCP sender
   stops sending, the flow stops sampling the network conditions, and so
   the value of the congestion window may become inaccurate.  We believe
   the correct conservative behavior under these circumstances is to
   decay the congestion window by half for every RTT that the flow
   remains inactive.  The value of half is a very conservative figure
   based on how quickly multiplicative decrease would have decayed the
   window in the presence of loss.

   Another possibility is that the sender may not stop sending, but may
   become application-limited rather than network-limited, and offer
   less data to the network than the congestion window allows to be
   sent.  In this case the TCP flow is still sampling network
   conditions, but is not offering sufficient traffic to be sure that
   there is still sufficient capacity in the network for that flow to
   send a full congestion window.  Under these circumstances we believe
   the correct conservative behavior is for the sender to keep track of
   the maximum amount of the congestion window used during each RTT, and
   to decay the congestion window each RTT to midway between the current
   cwnd value and the maximum value used.

   Before the congestion window is reduced, ssthresh is set to the
   maximum of its current value and 3/4 cwnd.  If the sender then has
   more data to send than the decayed cwnd allows, the TCP will slow-
   start (perform exponential increase) at least half-way back up to the
   old value of cwnd.

   The justification for this value of "3/4 cwnd" is that 3/4 cwnd is a
   conservative estimate of the recent average value of the congestion
   window, and the TCP should safely be able to slow-start at least up
   to this point.  For a TCP in steady-state that has been reducing its
   congestion window each time the congestion window reached some
   maximum value `maxwin', the average congestion window has been 3/4
   maxwin.  On average, when the connection becomes application-limited,



Handley, et al.               Experimental                      [Page 4]

RFC 2861            TCP Congestion Window Validation           June 2000


   cwnd will be 3/4 maxwin, and in this case cwnd itself represents the
   average value of the congestion window.  However, if the connection
   happens to become application-limited when cwnd equals maxwin, then
   the average value of the congestion window is given by 3/4 cwnd.

   An alternate possibility would be to set ssthresh to the maximum of
   the current value of ssthresh, and the old value of cwnd, allowing
   TCP to slow-start all of the way back up to the old value of cwnd.
   Further experimentation can be used to evaluate these two options for
   setting ssthresh.

   For the separate issue of the increase of the congestion window in
   response to an acknowledgement, we believe the correct behavior is
   for the sender to increase the congestion window only if the window
   was full when the acknowledgment arrived.

   We term this set of modifications to TCP Congestion Window Validation
   (CWV) because they are related to ensuring the congestion window is
   always a valid reflection of the current network state as probed by
   the connection.

3.1. The basic algorithm for reducing the congestion window

   A key issue in the CWV algorithm is to determine how to apply the
   guideline of reducing the congestion window once for every roundtrip
   time that the flow is application-limited.  We use TCP's
   retransmission timer (RTO) as a reasonable upper bound on the
   roundtrip time, and reduce the congestion window roughly once per
   RTO.

   This basic algorithm could be implemented in TCP as follows: When TCP
   sends a new packet it checks to see if more than RTO seconds have
   elapsed since the previous packet was sent.  If RTO has elapsed,
   ssthresh is set to the maximum of 3/4 cwnd and the current value of
   ssthresh, and then the congestion window is halved for every RTO that
   elapsed since the previous packet was sent.  In addition, T_prev is
   set to the current time, and W_used is reset to zero.  T_prev will be
   used to determine the elapsed time since the sender last was network-
   limited or had reduced cwnd after an idle period.  When the sender is
   application-limited, W_used holds the maximum congestion window
   actually used since the sender was last network-limited.

   The mechanism for determining the number of RTOs in the most recent
   idle period could also be implemented by using a timer that expires
   every RTO after the last packet was sent instead of a check per
   packet - efficiency constraints on different operating systems may
   dictate which is more efficient to implement.




Handley, et al.               Experimental                      [Page 5]

RFC 2861            TCP Congestion Window Validation           June 2000


   After TCP sends a packet, it also checks to see if that packet filled
   the congestion window.  If so, the sender is network-limited, and
   sets the variable T_prev to the current TCP clock time, and the
   variable W_used to zero.

   When TCP sends a packet that does not fill the congestion window, and
   the TCP send queue is empty, then the sender is application-limited.
   The sender checks to see if the amount of unacknowledged data is
   greater than W_used; if so, W_used is set to the amount of
   unacknowledged data.  In addition TCP checks to see if the elapsed
   time since T_prev is greater than RTO.  If so, then the TCP has not
   just reduced its congestion window following an idle period.  The TCP
   has been application-limited rather than network-limited for at least
   an entire RTO interval, but for less than two RTO intervals.  In this
   case, TCP sets ssthresh to the maximum of 3/4 cwnd and the current
   value of ssthresh, and reduces its congestion window to
   (cwnd+W_used)/2.  W_used is then set to zero, and T_prev is set to
   the current time, so a further reduction will not take place until at
   least another RTO period has elapsed.  Thus, during an application-
   limited period the CWV algorithm reduces the congestion window once
   per RTO.

3.2.  Pseudo-code for reducing the congestion window