TCPM WG                                                       J. Touch
Internet Draft                                              Independent
Intended status: Informational                                 M. Welzl
Obsoletes: 2140                                                S. Islam
Expires: October 2019 May 2020                                    University of Oslo
                                                         April 15,
                                                      November 19, 2019

                      TCP Control Block Interdependence

Status of this Memo

   This Internet-Draft is submitted in full conformance with the
   provisions of BCP 78 and BCP 79.

   This document may contain material from IETF Documents or IETF
   Contributions published or made publicly available before November
   10, 2008. The person(s) controlling the copyright in some of this
   material may not have granted the IETF Trust the right to allow
   modifications of such material outside the IETF Standards Process.
   Without obtaining an adequate license from the person(s) controlling
   the copyright in such materials, this document may not be modified
   outside the IETF Standards Process, and derivative works of it may
   not be created outside the IETF Standards Process, except to format
   it for publication as an RFC or to translate it into languages other
   than English.

   Internet-Drafts are working documents of the Internet Engineering
   Task Force (IETF), its areas, and its working groups.  Note that
   other groups may also distribute working documents as Internet-

   Internet-Drafts are draft documents valid for a maximum of six
   months and may be updated, replaced, or obsoleted by other documents
   at any time.  It is inappropriate to use Internet-Drafts as
   reference material or to cite them other than as "work in progress."

   The list of current Internet-Drafts can be accessed at

   The list of Internet-Draft Shadow Directories can be accessed at

   This Internet-Draft will expire on October 15, 2019. May 19, 2020.

Copyright Notice

   Copyright (c) 2019 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
   ( 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.


   This memo provides guidance to TCP implementers that are intended to
   help improve convergence to steady-state operation without affecting
   interoperability. It updates and replaces RFC 2140's description of
   interdependent TCP control blocks and the ways that part of TCP
   state can be shared among similar concurrent or consecutive
   connections. TCP state includes a combination of parameters, such as
   connection state, current round-trip time estimates, congestion
   control information, and process information. Most of this state is
   maintained on a per-connection basis in the TCP Control Block (TCB),
   but implementations can (and do) share certain TCB information
   across connections to the same host. Such sharing is intended to
   improve overall transient transport performance, while maintaining
   backward-compatibility with existing implementations. The sharing
   described herein is limited to only the TCB initialization and so
   has no effect on the long-term behavior of TCP after a connection
   has been established.

Table of Contents

   1. Introduction...................................................3
   2. Conventions used in this document..............................3 document..............................4
   3. Terminology....................................................4
   4. The TCP Control Block (TCB)....................................4
   5. TCB Interdependence............................................5
   6. An Example of Temporal Sharing.................................5 Sharing.................................6
   7. An Example of Ensemble Sharing.................................9
   8. Compatibility Issues..........................................11
   9. Implications..................................................13
   10. Implementation Observations..................................14
   11. Updates to RFC 2140..........................................15
   12. Security Considerations......................................16
   13. IANA Considerations..........................................16
   14. References...................................................16
      14.1. Normative References....................................16
      14.2. Informative References..................................17
   15. Acknowledgments..............................................19
   16. Change log...................................................19
   17. log...................................................20
   Appendix A: A : TCB sharing history..............................21
   18. history.................................22
   Appendix B: Options..........................................22 B : TCP Option Sharing and Caching......................22
   Appendix C : Automating the Initial Window in TCP over Long
      C.1. Introduction.............................................25
      C.2. Design Considerations....................................25
      C.3. Proposed IW Algorithm....................................26
      C.4. Discussion...............................................29
      C.5. Observations.............................................30

1. Introduction

   TCP is a connection-oriented reliable transport protocol layered
   over IP [RFC793]. Each TCP connection maintains state, usually in a
   data structure called the TCP Control Block (TCB). The TCB contains
   information about the connection state, its associated local
   process, and feedback parameters about the connection's transmission
   properties. As originally specified and usually implemented, most
   TCB information is maintained on a per-connection basis. Some
   implementations can (and now do) share certain TCB information
   across connections to the same host [RFC2140]. Such sharing is
   intended to lead to better overall transient performance, especially
   for numerous short-lived and simultaneous connections, as often used
   in the World-Wide Web [Be94],[Br02]. This sharing of state is
   intended to help TCP connections converge to steady-state behavior
   more quickly without affecting TCP interoperability.

   This document updates RFC 2140's discussion of TCB state sharing and
   provides a complete replacement for that document. This state
   sharing affects only TCB initialization [RFC2140] and thus has no
   effect on the long-term behavior of TCP after a connection has been
   established nor on interoperability. Path information shared across
   SYN destination port numbers assumes that TCP segments having the
   same host-pair experience the same path properties, irrespective of
   TCP port numbers. The observations about TCB sharing in this
   document apply similarly to any protocol with congestion state,
   including SCTP [RFC4960] and DCCP [RFC4340], as well as for
   individual subflows in Multipath TCP [RFC6824].

2. Conventions used in this document

   The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
   "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.

   However, this document is intended to describe behavior that is
   already permitted by TCP implementers. As a result, it provides
   informative guidance but does not use such normative language,
   except when quoting other documents.

3. Terminology

   Host - a source or sink of TCP segments associated with a single IP

   Host-pair - a pair of hosts and their corresponding IP addresses

   Path - an Internet path between the IP addresses of two hosts

4. The TCP Control Block (TCB)

   A TCB describes the data associated with each connection, i.e., with
   each association of a pair of applications across the network. The
   TCB contains at least the following information [RFC793]:

        Local process state
            pointers to send and receive buffers
            pointers to retransmission queue and current segment
            pointers to Internet Protocol (IP) PCB
        Per-connection shared state
                connection state
                local and remote host numbers and ports
                TCP option state
                send and receive window state (size*, current number)
                round-trip time and variance
                cong. window size (snd_cwnd)*
                cong. window size threshold (ssthresh)*
                max window size seen*
                round-trip time and variance#

   The per-connection information is shown as split into macro-state
   and micro-state, terminology borrowed from [Co91]. Macro-state
   describes the protocol for establishing the initial shared state
   about the connection; we include the endpoint numbers and components
   (timers, flags) required upon commencement that are later used to
   help maintain that state. Micro-state describes the protocol after a
   connection has been established, to maintain the reliability and
   congestion control of the data transferred in the connection.

   We further distinguish two other classes of shared micro-state that
   are associated more with host-pairs than with application pairs. One
   class is clearly host-pair dependent (#, e.g., MSS, MMS, PMTU, RTT),
   and the other is host-pair dependent in its aggregate (*, e.g.,
   congestion window information, current window sizes, etc.).

5. TCB Interdependence

   There are two cases of TCB interdependence. Temporal sharing occurs
   when the TCB of an earlier (now CLOSED) connection to a host is used
   to initialize some parameters of a new connection to that same host,
   i.e., in sequence. Ensemble sharing occurs when a currently active
   connection to a host is used to initialize another (concurrent)
   connection to that host.

6. An Example of Temporal Sharing

   The TCB data cache is accessed in two ways: it is read to initialize
   new TCBs and written when more current per-host state is available.
   New TCBs can be initialized using context from past connections as

             TEMPORAL SHARING - TCB Initialization

                     Cached TCB     New TCB
                     old_MMS_S      old_MMS_S or not cached

                     old_MMS_R      old_MMS_R or not cached

                     old_sendMSS    old_sendMSS

                     old_PMTU       old_PMTU

                     old_RTT        old_RTT

                     old_RTTvar     old_RTTvar

                     old_option     (option specific)

                     old_ssthresh   old_ssthresh

                     old_snd_cwnd   old_snd_cwnd

   Sections 8 and 9 discuss compatibility issues and implications of
   sharing the specific information listed above. Section 10 gives an
   overview of known implementations.

   Most cached TCB values are updated when a connection closes. The
   exceptions are MMS_R and MMS_S, which are reported by IP [RFC1122],
   PMTU which is updated after Path MTU Discovery
   [RFC1191][RFC4821][RFC8201], and sendMSS, which is updated if the
   MSS option is received in the TCP SYN header.

   Sharing sendMSS information affects only data in the SYN of the next
   connection, because sendMSS information is typically included in
   most TCP SYN segments. Caching PMTU can accelerate the efficiency of
   PMTUD, but can also result in black-holing until corrected if in
   error. Caching MMS_R and MMS_S may be of little direct value as they
   are reported by the local IP stack anyway.

   The way in which other TCP option state can be shared depends on the
   details of that option. E.g., TFO state includes the TCP Fast Open
   Cookie [RFC7413] or, in case TFO fails, a negative TCP Fast Open
   response. RFC 7413 states, "The client MUST cache negative responses
   from the server in order to avoid potential connection failures.
   Negative responses include the server not acknowledging the data in
   the SYN, ICMP error messages, and (most importantly) no response
   (SYN-ACK) from the server at all, i.e., connection timeout." [RFC
   7413]. TFOinfo is cached when a connection is established.

   Other TCP option state might not be as readily cached. E.g., TCP-AO
   [RFC5925] success or failure between a host pair for a single SYN
   destination port might be usefully cached. TCP-AO success or failure
   to other SYN destination ports on that host pair is never useful to
   cache because TCP-AO security parameters can vary per service.

   The table below gives an overview of option-specific information
   that can be shared. Additional information on TCP options and
   sharing is provided in Appendix B.

             TEMPORAL SHARING - Option info

             Cached               New
             old_TFO_Cookie       old_TFO_Cookie

             old_TFO_Failure      old_TFO_Failure
                TEMPORAL SHARING - Cache Updates

      Cached TCB     Current TCB     when?   New Cached TCB
      old_MMS_S      curr_ MMS_S     OPEN    curr MMS_S

      old_MMS_R      curr_ MMS_R     OPEN    curr_MMS_R

      old_sendMSS    curr_sendMSS    MSSopt  curr_sendMSS

      old_PMTU       curr_PMTU       PMTUD   curr_PMTU

      old_RTT        curr_RTT        CLOSE   merge(curr,old)

      old_RTTvar     curr_RTTvar     CLOSE   merge(curr,old)

      old_option     curr option     ESTAB   (depends on option)

      old_ssthresh   curr_ssthresh   CLOSE   merge(curr,old)

      old_snd_cwnd   curr_snd_cwnd   CLOSE   merge(curr,old)

   Caching PMTU and sendMSS is trivial; reported values are cached, and
   the most recent values are used. The cache is updated when the MSS
   option is received in a SYN or after PMTUD (i.e., when an ICMPv4
   Fraqmentation Needed [RFC1191] or ICMPv6 Packet Too Big message is
   received [RFC8201] or the equivalent is inferred, e.g. as from
   PLPMTUD [RFC4821]), respectively, so the cache always has the most
   recent values from any connection. For sendMSS, the cache is
   consulted only at connection establishment and not otherwise
   updated, which means that MSS options do not affect current
   connections. The default sendMSS is never saved; only reported MSS
   values update the cache, so an explicit override is required to
   reduce the sendMSS. There is no particular benefit to caching MMS_S
   and MMS R as these are reported by the local IP stack.

   TCP options are copied or merged depending on the details of each
   option, where "merge" is some function that combines the values of
   "curr" and "old". E.g., TFO state is updated when a connection is
   established and read before establishing a new connection.

   RTT values are updated by formulae that merge the old and new
   values. Dynamic RTT estimation requires a sequence of RTT
   measurements. As a result, the cached RTT (and its variance) is an
   average of its previous value with the contents of the currently
   active TCB for that host, when a TCB is closed. RTT values are
   updated only when a connection is closed. The method for merging old
   and current values needs to attempt to reduce the transient for new

   The updates for RTT, RTTvar and ssthresh rely on existing
   information, i.e., old values. Should no such values exist, the
   current values are cached instead.

                TEMPORAL SHARING - Option info Updates

   Cached          Current          when?   New Cached
   old_TFO_Cookie  old_TFO_Cookie   ESTAB   old_TFO_Cookie

   old_TFO_Failure old_TFO_Failure  ESTAB   old_TFO_Failure

7. An Example of Ensemble Sharing

   Sharing cached TCB data across concurrent connections requires
   attention to the aggregate nature of some of the shared state. For
   example, although MSS and RTT values can be shared by copying, it
   may not be appropriate to simply copy congestion window or ssthresh
   information; instead, the new values can be a function (f) of the
   cumulative values and the number of connections (N).

               ENSEMBLE SHARING - TCB Initialization

                  Cached TCB          New TCB
                  old_MMS_S           old_MMS_S

                  old_MMS_R           old_MMS_R

                  old_sendMSS         old_sendMSS

                  old_PMTU            old_PMTU

                  old_RTT             old_RTT

                  old_RTTvar          old_RTTvar

                  old ssthresh sum    f(old ssthresh sum, N)

                  old snd_cwnd sum    f(old snd cwnd sum, N)

                  old_option          (option-specific)

   Sections 8 and 9 discuss compatibility issues and implications of
   sharing the specific information listed above.

   The table below gives an overview of option-specific information
   that can be shared.

             ENSEMBLE SHARING  Option info

             Cached               New
             old_TFO_Cookie       old_TFO_Cookie

             old_TFO_Failure      old_TFO_Failure

             ENSEMBLE SHARING - Cache Updates

         Cached TCB   Current TCB   when?      New Cached TCB
         old_MMS_S    curr_MMS_S    OPEN       curr_MMS_S

         old_MMS_R    curr_MMS_R    OPEN       curr_MMS_R

         old_sendMSS  curr_sendMSS  MSSopt     curr_sendMSS

         old_PMTU     curr_PMTU     PMTUD      curr_PMTU

         old_RTT      curr_RTT      update     rtt_update(old,curr)

         old_RTTvar   curr_RTTvar   update     rtt_update(old,curr)

         old ssthresh curr ssthresh update     adjust sum as appopriate

         old snd_cwnd curr snd_cwnd update     adjust sum as appopriate

         old_option   curr option   (depends)  (option specific)

   For ensemble sharing, TCB information should be cached as early as
   possible, sometimes before a connection is closed. Otherwise,
   opening multiple concurrent connections may not result in TCB data
   sharing if no connection closes before others open. The amount of
   work involved in updating the aggregate average should be minimized,
   but the resulting value should be equivalent to having all values
   measured within a single connection. The function "rtt_update" in
   the ensemble sharing table indicates this operation, which occurs
   whenever the RTT would have been updated in the individual TCP
   connection. As a result, the cache contains the shared RTT
   variables, which no longer need to reside in the TCB.

   Congestion window size and ssthresh aggregation are more complicated
   in the concurrent case. When there is an ensemble of connections, we
   need to decide how that ensemble would have shared these variables,
   in order to derive initial values for new TCBs.

                ENSEMBLE SHARING - Option info Updates

   Cached          Current          when?   New Cached
   old_TFO_Cookie  old_TFO_Cookie   ESTAB   old_TFO_Cookie

   old_TFO_Failure old_TFO_Failure  ESTAB   old_TFO_Failure

   Any assumption of this sharing can be incorrect because identical
   endpoint address pairs may not share network paths. In current
   implementations, new congestion windows are set at an initial value
   of 4-10 segments [RFC3390][RFC6928], so that the sum of the current
   windows is increased for any new connection. This can have
   detrimental consequences where several connections share a highly
   congested link.

   There are several ways to initialize the congestion window in a new
   TCB among an ensemble of current connections to a host. Current TCP
   implementations initialize it to four segments as standard [rfc3390]
   and 10 segments experimentally [RFC6928]. These approaches assume
   that new connections should behave as conservatively as possible.
   The algorithm described in [Ba12] adjusts the initial cwnd depending
   on the cwnd values of ongoing connections. There have also been
   suggestions to use the kind of sharing mechanisms described in this
   document over long timescales to adapt TCP's initial window
   automatically [To13].
   automatically, as described further in Appendix A [To12].

8. Compatibility Issues

   For the congestion and current window information, the initial
   values computed by TCB interdependence may not be consistent with
   the long-term aggregate behavior of a set of concurrent connections
   between the same endpoints. Under conventional TCP congestion
   control, if a single existing connection has converged to a
   congestion window of 40 segments, two newly joining concurrent
   connections assume initial windows of 10 segments [RFC6928], and the
   current connection's window doesn't decrease to accommodate this
   additional load and connections can mutually interfere. One example
   of this is seen on low-bandwidth, high-delay links, where concurrent
   connections supporting Web traffic can collide because their initial
   windows were too large, even when set at one segment.

   The authors of [Hu12] recommend caching ssthresh for temporal
   sharing only when flows are long. Some studies suggest that sharing
   ssthresh between short flows can deteriorate the performance of
   individual connections [Hu12, Du16], although this may benefit
   aggregate network performance.

   Due to mechanisms like ECMP and LAG [RFC7424], TCP connections
   sharing the same host-pair may not always share the same path. This
   does not matter for host-specific information such as RWIN and TCP
   option state, such as TFOinfo. When TCB information is shared across
   different SYN destination ports, path-related information can be
   incorrect; however, the impact of this error is potentially
   diminished if (as discussed here) TCB sharing affects only the
   transient event of a connection start or if TCB information is
   shared only within connections to the same SYN destination port. In
   case of Temporal Sharing, TCB information could also become invalid
   over time. Because this is similar to the case when a connection
   becomes idle, mechanisms that address idle TCP connections (e.g.,
   [RFC7661]) could also be applied to TCB cache management, especially
   when TCP Fast Open is used [RFC7413].

   There may be additional considerations to the way in which TCB
   interdependence rebalances congestion feedback among the current
   connections, e.g., it may be appropriate to consider the impact of a
   connection being in Fast Recovery [RFC5861] [RFC5681] or some other similar
   unusual feedback state, e.g., as inhibiting or affecting the
   calculations described herein.

   TCP is sometimes used in situations where packets of the same host-
   pair do not always take the same path. Multipath routing that relies
   on examining transport headers, such as ECMP and LAG, may not result
   in repeatable path selection when TCP segments are encapsulated,
   encrypted, or altered - for example, in some Virtual Private Network
   (VPN) tunnels that rely on proprietary encapsulation. Similarly,
   such approaches cannot operate deterministically when the TCP header
   is encrypted, e.g., when using IPsec ESP. TCB interdependence among
   the entire set sharing the same endpoint IP addresses should work
   without problems under these circumstances. Moreover, measures to
   increase the probability that connections use the same path could be
   applied: e.g., the connections could be given the same IPv6 flow
   label. TCB interdependence can also be extended to sets of host IP
   address pairs that share the same network path conditions, such as
   when a group of addresses is on the same LAN (see Section 9).

   It can be wrong to share TCB information between TCP connections on
   the same host as identified by the IP address if an IP address is
   assigned to a new host (e.g., IP address spinning, as is used by
   ISPs to inhibit running servers). It can be wrong if Network Address
   (and Port) Translation (NA(P)T) [RFC2663] or any other IP sharing
   mechanism is used. Such mechanisms are less likely to be used with
   IPv6. Other methods to identify a host could also be considered to
   make correct TCB sharing more likely. Moreover, some TCB information
   is about dominant path properties rather than the specific host. IP
   addresses may differ, yet the relevant part of the path may be the

9. Implications

   There are several implications to incorporating TCB interdependence
   in TCP implementations. First, it may reduce the need for
   application-layer multiplexing for performance enhancement
   [RFC7231]. Protocols like HTTP/2 [RFC7540] avoid connection
   reestablishment costs by serializing or multiplexing a set of per-
   host connections across a single TCP connection. This avoids TCP's
   per-connection OPEN handshake and also avoids recomputing the MSS,
   RTT, and congestion window values. By avoiding the so-called, "slow-
   start restart," performance can be optimized [Hu01]. TCB
   interdependence can provide the "slow-start restart avoidance" of
   multiplexing, without requiring a multiplexing mechanism at the
   application layer.

   TCB interdependence pushes some of the TCP implementation from the
   traditional transport layer (in the ISO model), to the network
   layer. This acknowledges that some state is in fact per-host-pair or
   can be per-path as indicated solely by that host-pair. Transport
   protocols typically manage per-application-pair associations (per
   stream), and network protocols manage per-host-pair and path
   associations (routing). Round-trip time, MSS, and congestion
   information could be more appropriately handled in a network-layer
   fashion, aggregated among concurrent connections, and shared across
   connection instances [RFC3124].

   An earlier version of RTT sharing suggested implementing RTT state
   at the IP layer, rather than at the TCP layer. Our observations
   describe sharing state among TCP connections, which avoids some of
   the difficulties in an IP-layer solution. One such problem of an IP
   layer solution is determining the correspondence between packet
   exchanges using IP header information alone, where such
   correspondence is needed to compute RTT. Because TCB sharing
   computes RTTs inside the TCP layer using TCP header information, it
   can be implemented more directly and simply than at the IP layer.
   This is a case where information should be computed at the transport
   layer, but could be shared at the network layer.

   Per-host-pair associations are not the limit of these techniques. It
   is possible that TCBs could be similarly shared between hosts on a
   subnet or within a cluster, because the predominant path can be
   subnet-subnet, rather than host-host. Additionally, TCB
   interdependence can be applied to any protocol with congestion
   state, including SCTP [RFC4960] and DCCP [RFC4340], as well as for
   individual subflows in Multipath TCP [RFC6824].

   There may be other information that can be shared between concurrent
   connections. For example, knowing that another connection has just
   tried to expand its window size and failed, a connection may not
   attempt to do the same for some period. The idea is that existing
   TCP implementations infer the behavior of all competing connections,
   including those within the same host or subnet. One possible
   optimization is to make that implicit feedback explicit, via
   extended information associated with the endpoint IP address and its
   TCP implementation, rather than per-connection state in the TCB.

   Like the initial version of this document [RFC2140], this update's
   approach to TCB interdependence focuses on sharing a set of TCBs by
   updating the TCB state to reduce the impact of transients when
   connections begin or end. Other mechanisms have since been proposed
   to continuously share information between all ongoing communication
   (including connectionless protocols), updating the congestion state
   during any congestion-related event (e.g., timeout, loss
   confirmation, etc.) [RFC3124]. By dealing exclusively with
   transients, TCB interdependence is more likely to exhibit the same
   behavior as unmodified, independent TCP connections.

10. Implementation Observations

   The observation that some TCB state is host-pair specific rather
   than application-pair dependent is not new and is a common
   engineering decision in layered protocol implementations. Although
   now deprecated, T/TCP [RFC1644] was the first to propose using
   caches in order to maintain TCB states (see Appendix A for more

   The table below describes the current implementation status for some
   TCB information in Linux kernel version 4.6, FreeBSD 10 and Windows
   (as of October 2016). In the table, "shared" only refers to temporal

      TCB data     Status
      old MMS_S    Not shared

      old MMS_R    Not shared

      old_sendMSS  Cached and shared in Linux (MSS)

      old PMTU     Cached and shared in FreeBSD and Windows (PMTU)

      old_RTT      Cached and shared in FreeBSD and Linux

      old_RTTvar   Cached and shared in FreeBSD

      old TFOinfo  Cached and shared in Linux and Windows

      old_snd_cwnd Not shared

      old_ssthresh Cached and shared in FreeBSD and Linux:
                   FreeBSD: arithmetic
                   mean of ssthresh and previous value if
                   a previous value exists;
                   Linux: depending on state,
                   max(cwnd/2, ssthresh) in most cases

11. Updates to RFC 2140

   This document updates the description of TCB sharing in RFC 2140 and
   its associated impact on existing and new connection state,
   providing a complete replacement for that document [RFC2140]. It
   clarifies the previous description and terminology and extends the
   mechanism to its impact on new protocols and mechanisms, including
   multipath TCP, fast open, PLPMTUD, NAT, and the TCP Authentication

   The detailed impact on TCB state addresses TCB parameters in greater
   detail, addressing RSS in both the send and receive direction, MSS
   and send-MSS separately, adds path MTU and ssthresh, and addresses
   the impact on TCP option state.

   New sections have been added to address compatibility issues and
   implementation observations. The relation of this work to T/TCP has
   been moved to an appendix discussion Appendix A on history, partly to reflect the
   deprecation of that protocol.

   Appendix C has been added to discuss the potential to use temporal
   sharing over long timescales to adapt TCP's initial window
   automatically, largely imported from [To12].

   Finally, this document updates and significantly expands the
   referenced literature.

12. Security Considerations

   These presented implementation methods do not have additional
   ramifications for explicit attacks. They may be susceptible to
   denial-of-service attacks if not otherwise secured. For example, an
   application can open a connection and set its window size to zero,
   denying service to any other subsequent connection between those

   TCB sharing may be susceptible to denial-of-service attacks,
   wherever the TCB is shared, between connections in a single host, or
   between hosts if TCB sharing is implemented within a subnet (see
   Implications section). Some shared TCB parameters are used only to
   create new TCBs, others are shared among the TCBs of ongoing
   connections. New connections can join the ongoing set, e.g., to
   optimize send window size among a set of connections to the same

   Attacks on parameters used only for initialization affect only the
   transient performance of a TCP connection. For short connections,
   the performance ramification can approach that of a denial-of-
   service attack. E.g., if an application changes its TCB to have a
   false and small window size, subsequent connections would experience
   performance degradation until their window grew appropriately.

13. IANA Considerations

   There are no IANA implications or requests in this document.

   This section should be removed upon final publication as an RFC.

14. References

14.1. Normative References

   This document has no normative references.

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

   [RFC8174] Leiba., B., "Ambiguity of Uppercase vs Lowercase in RFC
             2119 Key Words", RFC 8174, May 2017.

14.2. Informative References

   [Br02]    Brownlee, N. and K. Claffy, "Understanding Internet
             Traffic Streams: Dragonflies and Tortoises",

   [Al10]    Allman, M., "Initial Congestion Window Specification",
             (work in progress), draft-allman-tcpm-bump-initcwnd-00,
             Nov. 2010.

   [Ba12]    Barik, R., Welzl, M., Ferlin, S., Alay, O., " LISA: A
             Linked Slow-Start Algorithm for MPTCP", IEEE
             Communications Magazine p110-117, 2002. ICC, Kuala
             Lumpur, Malaysia, May 23-27 2016.

   [Be94]    Berners-Lee, T., et al., "The World-Wide Web,"
             Communications of the ACM, V37, Aug. 1994, pp. 76-82.

   [Br94]    Braden, B., "T/TCP -- Transaction TCP: Source Changes for
             Sun OS 4.1.3,", Release 1.0, USC/ISI, September 14, 1994.

   [Br02]    Brownlee, N. and K. Claffy, "Understanding Internet
             Traffic Streams: Dragonflies and Tortoises", IEEE
             Communications Magazine p110-117, 2002.

   [Co91]    Comer, D., Stevens, D., Internetworking with TCP/IP, V2,
             Prentice-Hall, NJ, 1991.

   [FreeBSD] FreeBSD source code, Release 2.10,

   [Du16]    Dukkipati, N., Yuchung C., and Amin V., "Research
             Impacting the Practice of Congestion Control." ACM SIGCOMM
             CCR (editorial), on-line post, July 2016.

   [FreeBSD] FreeBSD source code, Release 2.10,

   [Hu01]    Hugues, A., Touch, J., Heidemann, J., "Issues in Slow-
             Start Restart After Idle", draft-hughes-restart-00
             (expired), Dec. 2001.

   [Hu12]    Hurtig, P., Brunstrom, A., "Enhanced metric caching for
             short TCP flows," 2012 IEEE International Conference on
             Communications (ICC), Ottawa, ON, 2012, pp. 1209-1213.

   [Ba12]    Barik, R., Welzl, M., Ferlin, S., Alay, O., " LISA: A
             Linked Slow-Start Algorithm for MPTCP", IEEE ICC, Kuala
             Lumpur, Malaysia, May 23-27 2016.

   [Ja88]    Jacobson, V., M. Karels, "Congestion Avoidance and
             Control", Proc. Sigcomm 1988.

   [RFC793]  Postel, Jon, "Transmission Control Protocol," Network
             Working Group RFC-793/STD-7, ISI, Sept. 1981.

   [RFC1122] Braden, R. (ed), "Requirements for Internet Hosts --
             Communication Layers", RFC-1122, Oct. 1989.

   [RFC1191] Mogul, J., Deering, S., "Path MTU Discovery," RFC 1191,
             Nov. 1990.

   [RFC1644] Braden, R., "T/TCP -- TCP Extensions for Transactions
             Functional Specification," RFC-1644, July 1994.

   [RFC1379] Braden, R., "Transaction TCP -- Concepts," RFC-1379,
             September 1992.

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

   [RFC2001] Stevens, W., "TCP Slow Start, Congestion Avoidance, Fast
             Retransmit, and Fast Recovery Algorithms", RFC2001
             (Standards Track), Jan. 1997.

   [RFC2140] Touch, J., "TCP Control Block Interdependence", RFC 2140,
             April 1997.

   [RFC2663] Srisuresh, P., Holdrege, M., "IP Network Address
             Translator (NAT) Terminology and Considerations", RFC-
             2663, August 1999.


   [RFC2414] Allman, M., Floyd, S., Partridge, C., "Increasing TCP's
             Initial Window," Window", RFC 2414 (Experimental), Sept. 1998.

   [RFC2581] Allman, M., Paxson, V., Stevens, W., "TCP Congestion
             Control," RFC2581 (Standards Track), Apr. 1999.

   [RFC2663] Srisuresh, P., Holdrege, M., "IP Network Address
             Translator (NAT) Terminology and Considerations", RFC-
             2663, August 1999.

   [RFC2861] Handley, M., Padhye, J., Floyd, S., "TCP Congestion Window
             Validation", RFC2861 (Experimental), June 2000.

   [RFC3390] Allman, M., Floyd, S., Partridge, C., "Increasing TCP's
             Initial Window," RFC 3390, Oct. 2002.

   [RFC7231] Fielding, R., J. Reshke, Eds., "HTTP/1.1 Semantics and
             Content," RFC-7231, June 2014.

   [RFC3124] Balakrishnan, H., Seshan, S., "The Congestion Manager,"
             RFC 3124, June 2001.

   [RFC4340] Kohler, E., Handley, M., Floyd, S., "Datagram Congestion
             Control Protocol (DCCP)," RFC 4340, Mar. 2006.

   [RFC4821] Mathis, M., Heffner, J., "Packetization Layer Path MTU
             Discovery," RFC 4821, Mar. 2007.

   [RFC4960] Stewart, R., (Ed.), "Stream Control Transmission
             Protocol," RFC4960, Sept. 2007.


   [RFC5681] Allman, M., Paxson, V., Blanton, E., "TCP Congestion
             Control," RFC 5861, Sept. 5681 (Standards Track), Sep. 2009.

   [RFC5925] Touch, J., Mankin, A., Bonica, R., "The TCP Authentication
             Option," RFC 5925, June 2010.

   [RFC6824] Ford, A., Raiciu, C., Handley, M., Bonaventure, O., "TCP
             Extensions for Multipath Operation with Multiple
             Addresses," RFC 6824, Jan. 2013.

   [RFC6928] Chu, J., Dukkipati, N., Cheng, Y., Mathis, M., "Increasing
             TCP's Initial Window," RFC 6928, Apr. 2013.

   [RFC7231] Fielding, R., J. Reshke, Eds., "HTTP/1.1 Semantics and
             Content," RFC-7231, June 2014.

   [RFC7323] Borman, D., B. Braden, V. Jacobson, R. Scheffenegger
             (Ed.), "TCP Extensions for High Performance," RFC 7323,
             Sept. 2014.

   [RFC7413] Cheng, Y., Chu, J., Radhakrishnan, S., Jain, A., "TCP Fast
             Open", RFC 7413, Dec. 2014.

   [RFC7424] Krishnan, R., Yong, L., Ghanwani, A., So, N., Khasnabish,
             B., "Mechanisms for Optimizing Link Aggregation Group
             (LAG) and Equal-Cost Multipath (ECMP) Component Link
             Utilization in Networks", RFC 7424, Jan. 2015

   [RFC7540] Belshe, M., Peon, R., Thomson, M., "Hypertext Transfer
             Protocol Version 2 (HTTP/2)", RFC 7540, May 2015.

   [RFC7661] Fairhurst, G., Sathiaseelan, A., Secchi, R., "Updating TCP
             to Support Rate-Limited Traffic", RFC 7661, Oct. 2015.

   [RFC8174] Leiba., B., "Ambiguity of Uppercase vs Lowercase in RFC
             2119 Key Words", RFC 8174, May 2017.

   [RFC8201] McCann, J., Deering. S., Mogul, J., Hinden, R. (Ed.),
             "Path MTU Discovery for IP version 6," RFC 8201, Jul.


   [To12]    Touch, J., "Automating the Initial Window in TCP," draft-
             touch-tcpm-automatic-iw-03 (expired), Jan. 2013. July 2012.

15. Acknowledgments

   The authors would like to thank for Praveen Balasubramanian for
   information regarding TCB sharing in Windows, and Yuchung Cheng,
   Lars Eggert, Ilpo Jarvinen and Michael Scharf for comments on
   earlier versions of the draft. Earlier revisions of this work
   received funding from a collaborative research project between the
   University of Oslo and Huawei Technologies Co., Ltd. and were partly
   supported by USC/ISI's Postel Center.

   This document was prepared using

16. Change log

   This section should be removed upon final publication as an RFC.


      - Added Appendix C to address long-timescale temporal adaptation.


      - Re-issued as draft-ietf-tcpm-2140bis due to WG adoption.
      - Cleaned orphan references to T/TCP, removed incomplete refs
      - Moved references to informative section and updated Sec 2
      - Updated to clarify no impact to interoperability
      - Updated appendix B to avoid 2119 language


     - Changed to update 2140, cite it normatively, and summarize the
        updates in a separate section


     - Fixed some TBDs.


     - Removed BCP-style recommendations and fixed some TBDs.


      - Updated Touch's affiliation and address information


      - Stated that our OS implementation overview table only covers
      temporal sharing.

      - Correctly reflected sharing of old_RTT in Linux in the
   implementation overview table.

      - Marked entries that are considered safe to share with an
   asterisk (suggestion was to split the table)

      - Discussed correct host identification: NATs may make IP
   addresses the wrong input, could e.g. use HTTP cookie.

      - Included MMS_S and MMS_R from RFC1122; fixed the use of MSS and

      - Added information about option sharing, listed options in the
   Appendix B

Authors' Addresses

   Joe Touch
   Manhattan Beach, CA 90266

   Phone: +1 (310) 560-0334

   Michael Welzl
   University of Oslo
   PO Box 1080 Blindern
   Oslo  N-0316

   Phone: +47 22 85 24 20

   Safiqul Islam
   University of Oslo
   PO Box 1080 Blindern
   Oslo  N-0316

   Phone: +47 22 84 08 37


Appendix A: TCB sharing history

   T/TCP proposed using caches to maintain TCB information across
   instances (temporal sharing), e.g., smoothed RTT, RTT variance,
   congestion avoidance threshold, and MSS [RFC1644]. These values were
   in addition to connection counts used by T/TCP to accelerate data
   delivery prior to the full three-way handshake during an OPEN. The
   goal was to aggregate TCB components where they reflect one
   association - that of the host-pair, rather than artificially
   separating those components by connection.

   At least one T/TCP implementation saved the MSS and aggregated the
   RTT parameters across multiple connections, but omitted caching the
   congestion window information [Br94], as originally specified in
   [RFC1379]. Some T/TCP implementations immediately updated MSS when
   the TCP MSS header option was received [Br94], although this was not
   addressed specifically in the concepts or functional specification
   [RFC1379][RFC1644]. In later T/TCP implementations, RTT values were
   updated only after a CLOSE, which does not benefit concurrent

   Temporal sharing of cached TCB data was originally implemented in
   the SunOS 4.1.3 T/TCP extensions [Br94] and the FreeBSD port of same
   [FreeBSD]. As mentioned before, only the MSS and RTT parameters were
   cached, as originally specified in [RFC1379]. Later discussion of
   T/TCP suggested including congestion control parameters in this
   cache; for example, [RFC1644] (Section 3.1) hints at initializing
   the congestion window to the old window size.


Appendix B: Options TCP Option Sharing and Caching

   In addition to the options that can be cached and shared, this memo
   also lists known options for which state is unsafe to be kept. This
   list is meant to avoid work duplication and should be removed upon

   Obsolete (unsafe to keep state):



      PO Conn permitted

      PO service profile



      Alt CS req

      Alt CS data

   No state to keep:










   Unsafe to keep state:

      Skeeter (DH exchange - might be obsolete, though)

      Bubba (DH exchange - might really be obsolete, though)

      Trailer CS

      SCPS capabilities


      Records boundaries

      Corruption experienced

      TCP Compression

      Quickstart response


      MPTCP (can we cache when this fails?)

      TFO success

   Safe but optional to keep state:


      TFO failure (so we don't try again, since it's optional)

   Safe and necessary to keep state:

      TFP cookie (if TFO succeeded in the past)

Appendix C: Automating the Initial Window in TCP over Long Timescales

   Note: this section is taken verbatim from [To12], updated to refer
   to itself as an appendix.

C.1. Introduction

   TCP's congestion control algorithm uses an initial window value
   (IW), both as a starting point for new connections and after one RTO
   or more [RFC2581][RFC2861]. This value has evolved over time,
   originally one maximum segment size (MSS), and increased to the
   lesser of four MSS or 4,380 bytes [RFC3390][RFC5681]. For typical
   Internet connections with an maximum transmission units (MTUs) of
   1500 bytes, this permits three segments of 1,460 bytes each.

   The IW value was originally implied in the original TCP congestion
   control description, and documented as a standard in 1997
   [RFC2001][Ja88]. The value was last updated in 1998 experimentally,
   and moved to the standards track in 2002 [RFC2414][RFC3390]. There
   have been recent proposals to update the IW based on further
   increases in host and router capabilities and network capacity, some
   focusing on specific values (e.g., IW=10), and others prescribing a
   schedule for increases over time (e.g., IW=6 for 2011, increasing by
   1-2 MSS per year).

   This appendix discusses how TCP can objectively measure when an IW
   is too large, and that such feedback should be used over long
   timescales to adjust the IW automatically. The result should be
   safer to deploy and might avoid the need to repeatedly revisit IW
   size over time.

   Note that this mechanism attempts to make the IW more adaptive over
   time. It can increase the IW beyond that which is currently
   recommended for widescale deployment, and so its use should be
   carefully monitored.

C.2. Design Considerations

   TCP's IW value has existed statically for over two decades, so any
   solution to adjusting the IW dynamically should have similarly
   stable, non-invasive effects on the performance and complexity of
   TCP. In order to be fair, the IW should be similar for most machines
   on the public Internet. Finally, a desirable goal is to develop a
   self-correcting algorithm, so that IW values that cause network
   problems can be avoided. To that end, we propose the following list
   of design goals:

   o  Impart little to no impact to TCP in the absence of loss, i.e.,
      it should not increase the complexity of default packet
      processing in the normal case.

   o  Adapt to network feedback over long timescales, avoiding values
      that persistently cause network problems.

   o  Decrease the IW in the presence of sustained loss of IW segments,
      as determined over a number of different connections.

   o  Increase the IW in the absence of sustained loss of IW segments,
      as determined over a number of different connections.

   o  Operate conservatively, i.e., tend towards leaving the IW the
      same in the absence of sufficient information, and give greater
      consideration to IW segment loss than IW segment success.

   We expect that, without other context, a good IW algorithm will
   converge to a single value, but this is not required. An endpoint
   with additional context or information, or deployed in a constrained
   environment, can always use a different value. In specific,
   information from previous connections, or sets of connections with a
   similar path, can already be used as context for such decisions (as
   noted in the core of this document).

   However, if a given IW value persistently causes packet loss during
   the initial burst of packets, it is clearly inappropriate and could
   be inducing unnecessary loss in other competing connections. This
   might happen for sites behind very slow boxes with small buffers,
   which may or may not be the first hop.

C.3. Proposed IW Algorithm

   Below is a simple description of the proposed IW algorithm. It
   relies on the following parameters:

   o  MinIW = 3 MSS or 4,380 bytes (as per RFC3390]

   o  MaxIW = 10

   o  MulDecr = 0.5

   o  AddIncr = 2 MSS

   o  Threshold = 0.05
   We assume that the minimum IW (MinIW) should be as currently
   specified [RFC3390]. The maximum IW can be set to a fixed value
   [RFC6928], or set based on a schedule if trusted time references are
   available [Al10]; here we prefer a fixed value. We also propose to
   use an AIMD algorithm, with increase and decreases as noted.

   Although these parameters are somewhat arbitrary, their initial
   values are not important except that the algorithm is AIMD and the
   MaxIW should not exceed that recommended for other systems on the
   Internet. Current proposals, including default current operation,
   are degenerate cases of the algorithm below for given parameters -
   notably MulDec = 1.0 and AddIncr = 0 MSS, thus disabling the
   automatic part of the algorithm.

   The proposed algorithm is as follows:

   1. On boot:

      IW = MaxIW; # assume this is in bytes, and an even number of MSS

   2. Upon starting a new connection

      CWND = IW;
      IWnotchecked = 1; # true

   3. During a connection's SYN-ACK processing, if SYN-ACK includes
      ECN, treat as if the IW is too large

      if (IWnotchecked && (synackecn == 1)) {
         IWnotchecked = 0; # never check again

   4. During a connection, if retransmission occurs, check the seqno of
      the outgoing packet (in bytes) to see if the resent segment fixes
      an IW loss:

      if (Retransmitting && IWnotchecked && ((ISN - seqno) < IW))) {
         IWnotchecked = 0; # never do this entire "if" again
      } else {
         IWnotchecked = 0; # you're beyond the IW so stop checking

   5. Once every 1000 conections, as a separate process (i.e., not as
      part of processing a given connection):

      if (conncount > 1000) {
         if (losscount/conncount > threshold) {
            # the number of connections with errors is too high
            IW = IW * MulDecr;
         } else {
            IW = IW + AddIncr;

   We recognize that this algorithm can yield a false positive when the
   sequence number wraps around. This can be avoided using either PAWS
   [RFC7323] context or 64-bit internal sequence numbers (as in TCP-AO
   [RFC5925]). Alternately, false positives can be allowed since they
   are expected to be infrequent and thus will not affect the overall
   statistics of the algorithm.

   The following additional constraints are imposed:

   >> The automatic IW algorithm MUST initialize to MaxIW, in the
   absence of other context information.

   If there are too few connections to make a decision or if there is
   otherwise insufficient information to increase the IW, then the
   MaxIW defaults to the current recommended value.

   >> An implementation may allow the MaxIW to grow beyond the
   currently recommended Internet default, but not more than 2 segments
   per calendar year.

   If an endpoint has a persistent history of successfully transmitting
   IW segments without loss, then it is allowed to probe the Internet
   to determine if larger IW values have similar success. This probing
   is limited and requires a trusted time source, otherwise the MaxIW
   remains constant.

   >> An implementation MUST adjust the IW based on loss statistics at
   least once every 1000 connections.

   An endpoint needs to be sufficiently reactive to IW loss.

   >> An implementation MUST decrease the IW by at least one MSS when
   indicated during an evaluation interval.

   An endpoint that detects loss needs to decrease its IW by at least
   one MSS, otherwise it is not participating in an automatic reactive

   >> An implementation MUST increase by no more than 2 MSS per
   evaluation interval.

   An endpoint that does not experience IW loss needs to probe the
   network incrementally.

   >> An implementation SHOULD use an IW that is an integer multiple of
   2 MSS.

   The IW should remain a multiple of 2 MSS segments, to enable
   efficient ACK compression without incurring unnecessary timeouts.

   >> An implementation MUST decrease the IW if more than 95% of
   connections have IW losses.

   Again, this is to ensure an implementation is sufficiently reactive.

   >> An implementation MAY group IW values and statistics within
   subsets of connections. Such grouping MAY use any information about
   connections to form groups except loss statistics.

   There are some TCP connections which might not be counted at all,
   such as those to/from loopback addresses, or those within the same
   subnet as that of a local interface (for which congestion control is
   sometimes disabled anyway). This may also include connections that
   terminate before the IW is full, i.e., as a separate check at the
   time of the connection closing.

   The period over which the IW is updated is intended to be a long
   timescale, e.g., a month or so, or 1,000 connections, whichever is
   longer. An implementation might check the IW once a month, and
   simply not update the IW or clear the connection counts in months
   where the number of connections is too small.

C.4. Discussion

   There are numerous parameters to the above algorithm that are
   compliant with the given requirements; this is intended to allow
   variation in configuration and implementation while ensuring that
   all such algorithms are reactive and safe.

   This algorithm continues to assume segments because that is the
   basis of most TCP implementations. It might be useful to consider
   revising the specifications to allow byte-based congestion given
   sufficient experience.

   The algorithm checks for IW losses only during the first IW after a
   connection start; it does not check for IW losses elsewhere the IW
   is used, e.g., during slow-start restarts.

   >> An implementation MAY detect IW losses during slow-start restarts
   in addition to losses during the first IW of a connection. In this
   case, the implementation MUST count each restart as a "connection"
   for the purposes of connection counts and periodic rechecking of the
   IW value.

   False positives can occur during some kinds of segment reordering,
   e.g., that might trigger spurious retransmissions even without a
   true segment loss. These are not expected to be sufficiently common
   to dominate the algorithm and its conclusions.

   This mechanism does require additional per-connection state which is
   currently common in some implementations, and is useful for other
   reasons (e.g., the ISN is used in TCP-AO [RFC5925]). The mechanism
   also benefits from persistent state kept across reboots, as would be
   other state sharing mechanisms (e.g., TCP Control Block Sharing
   [RFC2140]). The mechanism is inspired by RFC 2140's use of
   information across connections.

   The receive window (RWIN) is not involved in this calculation. The
   size of RWIN is determined by receiver resources, and provides space
   to accommodate segment reordering. It is not involved with
   congestion control, which is the focus of this document and its
   management of the IW.

C.5. Observations

   The IW may not converge to a single, global value. It also may not
   converge at all, but rather may oscillate by a few MSS as it
   repeatedly probes the Internet for larger IWs and fails. Both
   properties are consistent with TCP behavior during each individual

   This mechanism assumes that losses during the IW are due to IW size.
   Persistent errors that drop packets for other reasons - e.g., OS
   bugs, can cause false positives. Again, this is consistent with
   TCP's basic assumption that loss is caused by congestion and
   requires backoff. This algorithm treats the IW of new connections as
   a long-timescale backoff system.