RFC 3150






Network Working Group                                         S. Dawkins
Request for Comments: 3150                                 G. Montenegro
BCP: 48                                                         M . Kojo
Category: Best Current Practice                                V. Magret
                                                               July 2001


           End-to-end Performance Implications of Slow Links

Status of this Memo



   This document specifies an Internet Best Current Practices for the
   Internet Community, and requests discussion and suggestions for
   improvements.  Distribution of this memo is unlimited.

Copyright Notice



   Copyright (C) The Internet Society (2001).  All Rights Reserved.

Abstract



   This document makes performance-related recommendations for users of
   network paths that traverse "very low bit-rate" links.

   "Very low bit-rate" implies "slower than we would like".  This
   recommendation may be useful in any network where hosts can saturate
   available bandwidth, but the design space for this recommendation
   explicitly includes connections that traverse 56 Kb/second modem
   links or 4.8 Kb/second wireless access links - both of which are
   widely deployed.

   This document discusses general-purpose mechanisms.  Where
   application-specific mechanisms can outperform the relevant general-
   purpose mechanism, we point this out and explain why.

   This document has some recommendations in common with RFC 2689,
   "Providing integrated services over low-bitrate links", especially in
   areas like header compression.  This document focuses more on
   traditional data applications for which "best-effort delivery" is
   appropriate.











Dawkins, et al.          Best Current Practice                  [Page 1]

RFC 3150                   PILC - Slow Links                   July 2001


Table of Contents



   1.0 Introduction .................................................  2
   2.0 Description of Optimizations .................................  3
           2.1 Header Compression Alternatives ......................  3
           2.2 Payload Compression Alternatives .....................  5
           2.3 Choosing MTU sizes ...................................  5
           2.4 Interactions with TCP Congestion Control [RFC2581] ...  6
           2.5 TCP Buffer Auto-tuning ...............................  9
           2.6 Small Window Effects ................................. 10
   3.0 Summary of Recommended Optimizations ......................... 10
   4.0 Topics For Further Work ...................................... 12
   5.0 Security Considerations ...................................... 12
   6.0 IANA Considerations .......................................... 13
   7.0 Acknowledgements ............................................. 13
   8.0 References ................................................... 13
   Authors' Addresses ............................................... 16
   Full Copyright Statement ......................................... 17

1.0 Introduction



   The Internet protocol stack was designed to operate in a wide range
   of link speeds, and has met this design goal with only a limited
   number of enhancements (for example, the use of TCP window scaling as
   described in "TCP Extensions for High Performance" [RFC1323] for
   very-high-bandwidth connections).

   Pre-World Wide Web application protocols tended to be either
   interactive applications sending very little data (e.g., Telnet) or
   bulk transfer applications that did not require interactive response
   (e.g., File Transfer Protocol, Network News).  The World Wide Web has
   given us traffic that is both interactive and often "bulky",
   including images, sound, and video.

   The World Wide Web has also popularized the Internet, so that there
   is significant interest in accessing the Internet over link speeds
   that are much "slower" than typical office network speeds.  In fact,
   a significant proportion of the current Internet users is connected
   to the Internet over a relatively slow last-hop link.  In future, the
   number of such users is likely to increase rapidly as various mobile
   devices are foreseen to to be attached to the Internet over slow
   wireless links.

   In order to provide the best interactive response for these "bulky"
   transfers, implementors may wish to minimize the number of bits
   actually transmitted over these "slow" connections.  There are two





Dawkins, et al.          Best Current Practice                  [Page 2]

RFC 3150                   PILC - Slow Links                   July 2001


   areas that can be considered - compressing the bits that make up the
   overhead associated with the connection, and compressing the bits
   that make up the payload being transported over the connection.

   In addition, implementors may wish to consider TCP receive window
   settings and queuing mechanisms as techniques to improve performance
   over low-speed links.  While these techniques do not involve protocol
   changes, they are included in this document for completeness.

2.0 Description of Optimizations



   This section describes optimizations which have been suggested for
   use in situations where hosts can saturate their links.  The next
   section summarizes recommendations about the use of these
   optimizations.

2.1 Header Compression Alternatives



   Mechanisms for TCP and IP header compression defined in [RFC1144,
   RFC2507, RFC2508, RFC2509, RFC3095] provide the following benefits:

      -  Improve interactive response time

      -  Decrease header overhead (for a typical dialup MTU of 296
         bytes, the overhead of TCP/IP headers can decrease from about
         13 percent with typical 40-byte headers to 1-1.5 percent with
         with 3-5 byte compressed headers, for most packets).  This
         enables use of small packets for delay-sensitive low data-rate
         traffic and good line efficiency for bulk data even with small
         segment sizes (for reasons to use a small MTU on slow links,
         see section 2.3)

      -  Many slow links today are wireless and tend to be significantly
         lossy.  Header compression reduces packet loss rate over lossy
         links (simply because shorter transmission times expose packets
         to fewer events that cause loss).

   [RFC1144] header compression is a Proposed Standard for TCP Header
   compression that is widely deployed.  Unfortunately it is vulnerable
   on lossy links, because even a single bit error results in loss of
   synchronization between the compressor and decompressor.  It uses TCP
   timeouts to detect a loss of such synchronization, but these errors
   result in loss of data (up to a full TCP window), delay of a full
   RTO, and unnecessary slow-start.







Dawkins, et al.          Best Current Practice                  [Page 3]

RFC 3150                   PILC - Slow Links                   July 2001


   A more recent header compression proposal [RFC2507] includes an
   explicit request for retransmission of an uncompressed packet to
   allow resynchronization without waiting for a TCP timeout (and
   executing congestion avoidance procedures).  This works much better
   on links with lossy characteristics.

   The above scheme ceases to perform well under conditions as extreme
   as those of many cellular links (error conditions of 1e-3 or 1e-2 and
   round trip times over 100 ms.).  For these cases, the 'Robust Header
   Compression' working group has developed ROHC [RFC3095].  Extensions
   of ROHC to support compression of TCP headers are also under
   development.

   [RFC1323] defines a "TCP Timestamp" option, used to prevent
   "wrapping" of the TCP sequence number space on high-speed links, and
   to improve TCP RTT estimates by providing unambiguous TCP roundtrip
   timings.  Use of TCP timestamps prevents header compression, because
   the timestamps are sent as TCP options.  This means that each
   timestamped header has TCP options that differ from the previous
   header, and headers with changed TCP options are always sent
   uncompressed.  In addition, timestamps do not seem to have much of an
   impact on RTO estimation [AlPa99].

   Nevertheless, the ROHC working group is developing schemes to
   compress TCP headers, including options such as timestamps and
   selective acknowledgements.

   Recommendation: Implement [RFC2507], in particular as it relates to
   IPv4 tunnels and Minimal Encapsulation for Mobile IP, as well as TCP
   header compression for lossy links and links that reorder packets.
   PPP capable devices should implement "IP Header Compression over PPP"
   [RFC2509].  Robust Header Compression [RFC3095] is recommended for
   extremely slow links with very high error rates (see above), but
   implementors should judge if its complexity is justified (perhaps by
   the cost of the radio frequency resources).

   [RFC1144] header compression should only be enabled when operating
   over reliable "slow" links.

   Use of TCP Timestamps [RFC1323] is not recommended with these
   connections, because it complicates header compression.  Even though
   the Robust Header Compression (ROHC) working group is developing
   specifications to remedy this, those mechanisms are not yet fully
   developed nor deployed, and may not be generally justifiable.
   Furthermore, connections traversing "slow" links do not require
   protection against TCP sequence-number wrapping.





Dawkins, et al.          Best Current Practice                  [Page 4]

RFC 3150                   PILC - Slow Links                   July 2001


2.2 Payload Compression Alternatives



   Compression of IP payloads is also desirable on "slow" network links.
   "IP Payload Compression Protocol (IPComp)" [RFC2393] defines a
   framework where common compression algorithms can be applied to
   arbitrary IP segment payloads.

   IP payload compression is something of a niche optimization.  It is
   necessary because IP-level security converts IP payloads to random
   bitstreams, defeating commonly-deployed link-layer compression
   mechanisms which are faced with payloads that have no redundant
   "information" that can be more compactly represented.

   However, many IP payloads are already compressed (images, audio,
   video, "zipped" files being transferred), or are already encrypted
   above the IP layer (e.g., SSL [SSL]/TLS [RFC2246]).  These payloads
   will not "compress" further, limiting the benefit of this
   optimization.

   For uncompressed HTTP payload types, HTTP/1.1 [RFC2616] also includes
   Content-Encoding and Accept-Encoding headers, supporting a variety of
   compression algorithms for common compressible MIME types like
   text/plain.  This leaves only the HTTP headers themselves
   uncompressed.

   In general, application-level compression can often outperform
   IPComp, because of the opportunity to use compression dictionaries
   based on knowledge of the specific data being compressed.

   Extensive use of application-level compression techniques will reduce
   the need for IPComp, especially for WWW users.

   Recommendation: IPComp may optionally be implemented.

2.3 Choosing MTU Sizes



   There are several points to keep in mind when choosing an MTU for
   low-speed links.

   First, if a full-length MTU occupies a link for longer than the
   delayed ACK timeout (typically 200 milliseconds, but may be up to 500
   milliseconds), this timeout will cause an ACK to be generated for
   every segment, rather than every second segment, as occurs with most
   implementations of the TCP delayed ACK algorithm.







Dawkins, et al.          Best Current Practice                  [Page 5]

RFC 3150                   PILC - Slow Links                   July 2001


   Second, "relatively large" MTUs, which take human-perceptible amounts
   of time to be transmitted into the network, create human-perceptible
   delays in other flows using the same link.  [RFC1144] considers
   100-200 millisecond delays as human-perceptible.  The convention of
   choosing 296-byte MTUs (with header compression enabled) for dialup
   access is a compromise that limits the maximum link occupancy delay
   with full-length MTUs close to 200 milliseconds on 9.6 Kb/second
   links.

   Third, on last-hop links using a larger link MTU size, and therefore
   larger MSS, would allow a TCP sender to increase its congestion
   window faster in bytes than when using a smaller MTU size (and a
   smaller MSS).  However, with a smaller MTU size, and a smaller MSS
   size, the congestion window, when measured in segments, increases
   more quickly than it would with a larger MSS size.  Connections using
   smaller MSS sizes are more likely to be able to send enough segments
   to generate three duplicate acknowledgements, triggering fast
   retransmit/fast recovery when packet losses are encountered.  Hence,
   a smaller MTU size is useful for slow links with lossy
   characteristics.

   Fourth, using a smaller MTU size also decreases the queuing delay of
   a TCP flow (and thereby RTT) compared to use of larger MTU size with
   the same number of packets in a queue.  This means that a TCP flow
   using a smaller segment size and traversing a slow link is able to
   inflate the congestion window (in number of segments) to a larger
   value while experiencing the same queuing delay.

   Finally, some networks charge for traffic on a per-packet basis, not
   on a per-kilobyte basis.  In these cases, connections using a larger
   MTU may be charged less than connections transferring the same number
   of bytes using a smaller MTU.

   Recommendation: If it is possible to do so, MTUs should be chosen
   that do not monopolize network interfaces for human-perceptible
   amounts of time, and implementors should not chose MTUs that will
   occupy a network interface for significantly more than 100-200
   milliseconds.

2.4 Interactions with TCP Congestion Control [RFC2581]



   In many cases, TCP connections that traverse slow links have the slow
   link as an "access" link, with higher-speed links in use for most of
   the connection path.  One common configuration might be a laptop
   computer using dialup access to a terminal server (a last-hop
   router), with an HTTP server on a high-speed LAN "behind" the
   terminal server.




Dawkins, et al.          Best Current Practice                  [Page 6]

RFC 3150                   PILC - Slow Links                   July 2001


   In this case, the HTTP server may be able to place packets on its
   directly-attached high-speed LAN at a higher rate than the last-hop
   router can forward them on the low-speed link.  When the last-hop
   router falls behind, it will be unable to buffer the traffic intended
   for the low-speed link, and will become a point of congestion and
   begin to drop the excess packets.  In particular, several packets may
   be dropped in a single transmission window when initial slow start
   overshoots the last-hop router buffer.

   Although packet loss is occurring, it isn't detected at the TCP
   sender until one RTT time after the router buffer space is exhausted
   and the first packet is dropped.  This late congestion signal allows
   the congestion window to increase up to double the size it was at the
   time the first packet was dropped at the router.

   If the link MTU is large enough to take more than the delayed ACK
   timeout interval to transmit a packet, an ACK is sent for every
   segment and the congestion window is doubled in a single RTT.  If a
   smaller link MTU is in use and delayed ACKs can be utilized, the
   congestion window increases by a factor of 1.5 in one RTT.  In both
   cases the sender continues transmitting packets well beyond the
   congestion point of the last-hop router, resulting in multiple packet
   losses in a single window.

   The self-clocking nature of TCP's slow start and congestion avoidance
   algorithms prevent this buffer overrun from continuing.  In addition,
   these algorithms allow senders to "probe" for available bandwidth -
   cycling through an increasing rate of transmission until loss occurs,
   followed by a dramatic (50-percent) drop in transmission rate.  This
   happens when a host directly connected to a low-speed link offers an
   advertised window that is unrealistically large for the low-speed
   link.  During the congestion avoidance phase the peer host continues
   to probe for available bandwidth, trying to fill the advertised
   window, until packet loss occurs.

   The same problems may also exist when a sending host is directly
   connected to a slow link as most slow links have some local buffer in
   the link interface.  This link interface buffer is subject to
   overflow exactly in the same way as the last-hop router buffer.

   When a last-hop router with a small number of buffers per outbound
   link is used, the first buffer overflow occurs earlier than it would
   if the router had a larger number of buffers.  Subsequently with a
   smaller number of buffers the periodic packet losses occur more
   frequently during congestion avoidance, when the sender probes for
   available bandwidth.





Dawkins, et al.          Best Current Practice                  [Page 7]

RFC 3150                   PILC - Slow Links                   July 2001


   The most important responsibility of router buffers is to absorb
   bursts.  Too few buffers (for example, only three buffers per
   outbound link as described in [RFC2416]) means that routers will
   overflow their buffer pools very easily and are unlikely to absorb
   even a very small burst.  When a larger number of router buffers are
   allocated per outbound link, the buffer space does not overflow as
   quickly but the buffers are still likely to become full due to TCP's
   default behavior.  A larger number of router buffers leads to longer
   queuing delays and a longer RTT.

   If router queues become full before congestion is signaled or remain
   full for long periods of time, this is likely to result in "lock-
   out", where a single connection or a few connections occupy the
   router queue space, preventing other connections from using the link
   [RFC2309], especially when a tail drop queue management discipline is
   being used.

   Therefore, it is essential to have a large enough number of buffers
   in routers to be able to absorb data bursts, but keep the queues
   normally small.  In order to achieve this it has been recommended in
   [RFC2309] that an active queue management mechanism, like Random
   Early Detection (RED) [RED93], should be implemented in all Internet
   routers, including the last-hop routers in front of a slow link.  It
   should also be noted that RED requires a sufficiently large number of
   router buffers to work properly.  In addition, the appropriate
   parameters of RED on a last-hop router connected to a slow link will
   likely deviate from the defaults recommended.

   Active queue management mechanism do not eliminate packet drops but,
   instead, drop packets at earlier stage to solve the full-queue
   problem for flows that are responsive to packet drops as congestion
   signal.  Hosts that are directly connected to low-speed links may
   limit the receive windows they advertise in order to lower or
   eliminate the number of packet drops in a last-hop router.  When
   doing so one should, however, take care that the advertised window is
   large enough to allow full utilization of the last-hop link capacity
   and to allow triggering fast retransmit, when a packet loss is
   encountered.  This recommendation takes two forms:

   -  Modern operating systems use relatively large default TCP receive
      buffers compared to what is required to fully utilize the link
      capacity of low-speed links.  Users should be able to choose the
      default receive window size in use - typically a system-wide
      parameter.  (This "choice" may be as simple as "dial-up access/LAN
      access" on a dialog box - this would accommodate many environments
      without requiring hand-tuning by experienced network engineers.)





Dawkins, et al.          Best Current Practice                  [Page 8]

RFC 3150                   PILC - Slow Links                   July 2001


   -  Application developers should not attempt to manually manage
      network bandwidth using socket buffer sizes.  Only in very rare
      circumstances will an application actually know both the bandwidth
      and delay of a path and be able to choose a suitably low (or high)
      value for the socket buffer size to obtain good network
      performance.

   This recommendation is not a general solution for any network path
   that might involve a slow link.  Instead, this recommendation is
   applicable in environments where the host "knows" it is always
   connected to other hosts via "slow links".  For hosts that may
   connect to other host over a variety of links (e.g., dial-up laptop
   computers with LAN-connected docking stations), buffer auto-tuning
   for the receive buffer is a more reasonable recommendation, and is
   discussed below.

2.5 TCP Buffer Auto-tuning



   [SMM98] recognizes a tension between the desire to allocate "large"
   TCP buffers, so that network paths are fully utilized, and a desire
   to limit the amount of memory dedicated to TCP buffers, in order to
   efficiently support large numbers of connections to hosts over
   network paths that may vary by six orders of magnitude.

   The technique proposed is to dynamically allocate TCP buffers, based
   on the current congestion window, rather than attempting to
   preallocate TCP buffers without any knowledge of the network path.

   This proposal results in receive buffers that are appropriate for the
   window sizes in use, and send buffers large enough to contain two
   windows of segments, so that SACK and fast recovery can recover
   losses without forcing the connection to use lengthy retransmission
   timeouts.

   While most of the motivation for this proposal is given from a
   server's perspective, hosts that connect using multiple interfaces
   with markedly-different link speeds may also find this kind of
   technique useful.  This is true in particular with slow links, which
   are likely to dominate the end-to-end RTT.  If the host is connected
   only via a single slow link interface at a time, it is fairly easy to
   (dynamically) adjust the receive window (and thus the advertised
   window) to a value appropriate for the slow last-hop link with known
   bandwidth and delay characteristics.

   Recommendation: If a host is sometimes connected via a slow link but
   the host is also connected using other interfaces with markedly-
   different link speeds, it may use receive buffer auto-tuning to
   adjust the advertised window to an appropriate value.



Dawkins, et al.          Best Current Practice                  [Page 9]

RFC 3150                   PILC - Slow Links                   July 2001


2.6 Small Window Effects



   If a TCP connection stabilizes with a congestion window of only a few
   segments (as could be expected on a "slow" link), the sender isn't
   sending enough segments to generate three duplicate acknowledgements,
   triggering fast retransmit and fast recovery.  This means that a
   retransmission timeout is required to repair the loss - dropping the
   TCP connection to a congestion window with only one segment.

   [TCPB98] and [TCPF98] observe that (in studies of network trace
   datasets) it is relatively common for TCP retransmission timeouts to
   occur even when some duplicate acknowledgements are being sent.  The
   challenge is to use these duplicate acknowledgements to trigger fast
   retransmit/fast recovery without injecting traffic into the network
   unnecessarily - and especially not injecting traffic in ways that
   will result in instability.

   The "Limited Transmit" algorithm [RFC3042] suggests sending a new
   segment when the first and second duplicate acknowledgements are
   received, so that the receiver is more likely to be able to continue
   to generate duplicate acknowledgements until the TCP retransmit
   threshold is reached, triggering fast retransmit and fast recovery.
   When the congestion window is small, this is very useful in assisting
   fast retransmit and fast recovery to recover from a packet loss
   without using a retransmission timeout.  We note that a maximum of
   two additional new segments will be sent before the receiver sends
   either a new acknowledgement advancing the window or two additional
   duplicate acknowledgements, triggering fast retransmit/fast recovery,
   and that these new segments will be acknowledgement-clocked, not
   back-to-back.

   Recommendation: Limited Transmit should be implemented in all hosts.

3.0 Summary of Recommended Optimizations



   This section summarizes our recommendations regarding the previous
   standards-track mechanisms, for end nodes that are connected via a
   slow link.

   Header compression should be implemented.  [RFC1144] header
   compression can be enabled over robust network links.  [RFC2507]
   should be used over network connections that are expected to
   experience loss due to corruption as well as loss due to congestion.
   For extremely lossy and slow links, implementors should evaluate ROHC
   [RFC3095] as a potential solution.  [RFC1323] TCP timestamps must be
   turned off because (1) their protection against TCP sequence number
   wrapping is unjustified for slow links, and (2) they complicate TCP
   header compression.



Dawkins, et al.          Best Current Practice                 [Page 10]

RFC 3150                   PILC - Slow Links                   July 2001


   IP Payload Compression [RFC2393] should be implemented, although
   compression at higher layers of the protocol stack (for example [RFC
   2616]) may make this mechanism less useful.

   For HTTP/1.1 environments, [RFC2616] payload compression should be
   implemented and should be used for payloads that are not already
   compressed.

   Implementors should choose MTUs that don't monopolize network
   interfaces for more than 100-200 milliseconds, in order to limit the
   impact of a single connection on all other connections sharing the
   network interface.

   Use of active queue management is recommended on last-hop routers
   that provide Internet access to host behind a slow link.  In
   addition, number of router buffers per slow link should be large
   enough to absorb concurrent data bursts from more than a single flow.
   To absorb concurrent data bursts from two or three TCP senders with a
   typical data burst of three back-to-back segments per sender, at
   least six (6) or nine (9) buffers are needed.  Effective use of
   active queue management is likely to require even larger number of
   buffers.

   Implementors should consider the possibility that a host will be
   directly connected to a low-speed link when choosing default TCP
   receive window sizes.

   Application developers should not attempt to manually manage network
   bandwidth using socket buffer sizes as only in very rare
   circumstances an application will be able to choose a suitable value
   for the socket buffer size to obtain good network performance.

   Limited Transmit [RFC3042] should be implemented in all end hosts as
   it assists in triggering fast retransmit when congestion window is
   small.

   All of the mechanisms described above are stable standards-track RFCs
   (at Proposed Standard status, as of this writing).

   In addition, implementors may wish to consider TCP buffer auto-
   tuning, especially when the host system is likely to be used with a
   wide variety of access link speeds.  This is not a standards-track
   TCP mechanism but, as it is an operating system implementation issue,
   it does not need to be standardized.

   Of the above mechanisms, only Header Compression (for IP and TCP) may
   cease to work in the presence of end-to-end IPSEC.  However,
   [RFC3095] does allow compressing the ESP header.



Dawkins, et al.          Best Current Practice                 [Page 11]

RFC 3150                   PILC - Slow Links                   July 2001


4.0 Topics For Further Work



   In addition to the standards-track mechanisms discussed above, there
   are still opportunities to improve performance over low-speed links.

   "Sending fewer bits" is an obvious response to slow link speeds.  The
   now-defunct HTTP-NG proposal [HTTP-NG] replaced the text-based HTTP
   header representation with a binary representation for compactness.
   However, HTTP-NG is not moving forward and HTTP/1.1 is not being
   enhanced to include a more compact HTTP header representation.
   Instead, the Wireless Application Protocol (WAP) Forum has opted for
   the XML-based Wireless Session Protocol [WSP], which includes a
   compact header encoding mechanism.

   It would be nice to agree on a more compact header representation
   that will be used by all WWW communities, not only the wireless WAN
   community.  Indeed, general XML content encodings have been proposed
   [Millau], although they are not yet widely adopted.

   We note that TCP options which change from segment to segment
   effectively disable header compression schemes deployed today,
   because there's no way to indicate that some fields in the header are
   unchanged from the previous segment, while other fields are not.  The
   Robust Header Compression working group is developing such schemes
   for TCP options such as timestamps and selective acknowledgements.
   Hopefully, documents subsequent to [RFC3095] will define such
   specifications.

   Another effort worth following is that of 'Delta Encoding'.  Here,
   clients that request a slightly modified version of some previously
   cached resource would receive a succinct description of the
   differences, rather than the entire resource [HTTP-DELTA].

5.0 Security Considerations



   All recommendations included in this document are stable standards-
   track RFCs (at Proposed Standard status, as of this writing) or
   otherwise do not suggest any changes to any protocol.  With the
   exception of Van Jacobson compression [RFC1144] and [RFC2507,
   RFC2508, RFC2509], all other mechanisms are applicable to TCP
   connections protected by end-to-end IPSec.  This includes ROHC
   [RFC3095], albeit partially, because even though it can compress the
   outermost ESP header to some extent, encryption still renders any
   payload data uncompressible (including any subsequent protocol
   headers).






Dawkins, et al.          Best Current Practice                 [Page 12]

RFC 3150                   PILC - Slow Links                   July 2001


6.0 IANA Considerations



   This document is a pointer to other, existing IETF standards.  There
   are no new IANA considerations.

7.0 Acknowledgements



   This recommendation has grown out of "Long Thin Networks" [RFC2757],
   which in turn benefited from work done in the IETF TCPSAT working
   group.

8.0 References



   [AlPa99]     Mark Allman and Vern Paxson, "On Estimating End-to-End
                Network Path Properties", in ACM SIGCOMM 99 Proceedings,
                1999.

   [HTTP-DELTA] J. Mogul, et al., "Delta encoding in HTTP", Work in
                Progress.

   [HTTP-NG]    Mike Spreitzer, Bill Janssen, "HTTP 'Next Generation'",
                9th International WWW Conference, May, 2000.  Also
                available as:  http://www.www9.org/w9cdrom/60/60.html

   [Millau]     Marc Girardot, Neel Sundaresan, "Millau: an encoding
                format for efficient representation and exchange of XML
                over the Web", 9th International WWW Conference, May,
                2000.  Also available as:
                http://www.www9.org/w9cdrom/154/154.html

   [PAX97]      Paxson, V., "End-to-End Internet Packet Dynamics", 1997,
                in SIGCOMM 97 Proceedings, available as:
                http://www.acm.org/sigcomm/ccr/archive/ccr-toc/ccr-toc-
                97.html

   [RED93]      Floyd, S., and Jacobson, V., Random Early Detection
                gateways for Congestion Avoidance, IEEE/ACM Transactions
                on Networking, V.1 N.4, August 1993, pp. 397-413.  Also
                available from http://ftp.ee.lbl.gov/floyd/red.html.

   [RFC1144]    Jacobson, V., "Compressing TCP/IP Headers for Low-Speed
                Serial Links", RFC 1144, February 1990.









Dawkins, et al.          Best Current Practice                 [Page 13]

RFC 3150                   PILC - Slow Links                   July 2001


   [RFC1323]    Jacobson, V., Braden, R. and D. Borman, "TCP Extensions
                for High Performance", RFC 1323, May 1992.

   [RFC2246]    Dierks, T. and C. Allen, "The TLS Protocol: Version
                1.0", RFC 2246, January 1999.

   [RFC2309]    Braden, R., Clark, D., Crowcroft, J., Davie, B.,
                Deering, S., Estrin, D., Floyd, S., Jacobson, V.,
                Minshall, G., Partridge, C., Peterson, L., Ramakrishnan,
                K., Shenker, S., Wroclawski, J. and L. Zhang,
                "Recommendations on Queue Management and Congestion
                Avoidance in the Internet", RFC 2309, April 1998.

   [RFC2393]    Shacham, A., Monsour, R., Pereira, R. and M. Thomas, "IP
                Payload Compression Protocol (IPComp)", RFC 2393,
                December 1998.

   [RFC2401]    Kent, S. and R. Atkinson, "Security Architecture for the
                Internet Protocol", RFC 2401, November 1998.

   [RFC2416]    Shepard, T. and C. Partridge, "When TCP Starts Up With
                Four Packets Into Only Three Buffers", RFC 2416,
                September 1998.

   [RFC2507]    Degermark, M., Nordgren, B. and S. Pink, "IP Header
                Compression", RFC 2507, February 1999.

   [RFC2508]    Casner, S. and V. Jacobson. "Compressing IP/UDP/RTP
                Headers for Low-Speed Serial Links", RFC 2508, February
                1999.

   [RFC2509]    Engan, M., Casner, S. and C. Bormann, "IP Header
                Compression over PPP", RFC 2509, February 1999.

   [RFC2581]    Allman, M., Paxson, V. and W. Stevens, "TCP Congestion
                Control", RFC 2581, April 1999.

   [RFC2616]    Fielding, R., Gettys, J., Mogul, J., Frystyk, H.,
                Masinter, L., Leach, P. and T. Berners-Lee, "Hypertext
                Transfer Protocol -- HTTP/1.1", RFC 2616, June 1999.

   [RFC2757]    Montenegro, G., Dawkins, S., Kojo, M., Magret, V., and
                N. Vaidya, "Long Thin Networks", RFC 2757, January 2000.

   [RFC3042]    Allman, M., Balakrishnan, H. and S. Floyd, "Enhancing
                TCP's Loss Recovery Using Limited Transmit", RFC 3042,
                January 2001.




Dawkins, et al.          Best Current Practice                 [Page 14]

RFC 3150                   PILC - Slow Links                   July 2001


   [RFC3095]    Bormann, C., Burmeister, C., Degermark, M., Fukushima,
                H., Hannu, H., Jonsson, L-E., Hakenberg, R., Koren, T.,
                Le, K., Liu, Z., Martensson, A., Miyazaki, A., Svanbro,
                K., Wiebke, T., Yoshimura, T. and H. Zheng, "RObust
                Header Compression (ROHC): Framework and four Profiles:
                RTP, UDP ESP and uncompressed", RFC 3095, July 2001.

   [SMM98]      Jeffrey Semke, Matthew Mathis, and Jamshid Mahdavi,
                "Automatic TCP Buffer Tuning", in ACM SIGCOMM 98
                Proceedings 1998. Available from
                http://www.acm.org/sigcomm/sigcomm98/tp/abs_26.html.

   [SSL]        Alan O. Freier, Philip Karlton, Paul C. Kocher, The SSL
                Protocol: Version 3.0, March 1996.  (Expired Internet-
                Draft, available from
                http://home.netscape.com/eng/ssl3/ssl-toc.html)

   [TCPB98]     Hari Balakrishnan, Venkata N. Padmanabhan, Srinivasan
                Seshan, Mark Stemm, Randy H. Katz, "TCP Behavior of a
                Busy Internet Server: Analysis and Improvements", IEEE
                Infocom, March 1998. Available from:
                http://www.cs.berkeley.edu/~hari/papers/infocom98.ps.gz

   [TCPF98]     Dong Lin and H.T. Kung, "TCP Fast Recovery Strategies:
                Analysis and Improvements", IEEE Infocom, March 1998.
                Available from:
                http://www.eecs.harvard.edu/networking/papers/ infocom-
                tcp-final-198.pdf

   [WSP]        Wireless Application Protocol Forum, "WAP Wireless
                Session Protocol Specification", approved 4 May, 2000,
                available from
                http://www1.wapforum.org/tech/documents/WAP-203-WSP-
                20000504-a.pdf.  (informative reference).

















Dawkins, et al.          Best Current Practice                 [Page 15]

RFC 3150                   PILC - Slow Links                   July 2001


Authors' Addresses



   Questions about this document may be directed to:

   Spencer Dawkins
   Fujitsu Network Communications
   2801 Telecom Parkway
   Richardson, Texas 75082

   Phone:  +1-972-479-3782
   EMail: spencer.dawkins@fnc.fujitsu.com


   Gabriel Montenegro
   Sun Microsystems Laboratories, Europe
   29, chemin du Vieux Chene
   38240 Meylan, FRANCE

   Phone:  +33 476 18 80 45
   EMail: gab@sun.com


   Markku Kojo
   Department of Computer Science
   University of Helsinki
   P.O. Box 26 (Teollisuuskatu 23)
   FIN-00014 HELSINKI
   Finland

   Phone: +358-9-1914-4179
   Fax:   +358-9-1914-4441
   EMail: kojo@cs.helsinki.fi


   Vincent Magret
   Alcatel Internetworking, Inc.
   26801 W. Agoura road
   Calabasas, CA, 91301

   Phone: +1 818 878 4485
   EMail: vincent.magret@alcatel.com










Dawkins, et al.          Best Current Practice                 [Page 16]

RFC 3150                   PILC - Slow Links                   July 2001


Full Copyright Statement



   Copyright (C) The Internet Society (2001).  All Rights Reserved.

   This document and translations of it may be copied and furnished to
   others, and derivative works that comment on or otherwise explain it
   or assist in its implementation may be prepared, copied, published
   and distributed, in whole or in part, without restriction of any
   kind, provided that the above copyright notice and this paragraph are
   included on all such copies and derivative works.  However, this
   document itself may not be modified in any way, such as by removing
   the copyright notice or references to the Internet Society or other
   Internet organizations, except as needed for the purpose of
   developing Internet standards in which case the procedures for
   copyrights defined in the Internet Standards process must be
   followed, or as required to translate it into languages other than
   English.

   The limited permissions granted above are perpetual and will not be
   revoked by the Internet Society or its successors or assigns.

   This document and the information contained herein is provided on an
   "AS IS" basis and THE INTERNET SOCIETY AND THE INTERNET ENGINEERING
   TASK FORCE DISCLAIMS ALL WARRANTIES, EXPRESS OR IMPLIED, INCLUDING
   BUT NOT LIMITED TO ANY WARRANTY THAT THE USE OF THE INFORMATION
   HEREIN WILL NOT INFRINGE ANY RIGHTS OR ANY IMPLIED WARRANTIES OF
   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE.

Acknowledgement



   Funding for the RFC Editor function is currently provided by the
   Internet Society.



















Dawkins, et al.          Best Current Practice                 [Page 17]