RFC 2309
This document is obsolete. Please refer to RFC 7567.






Network Working Group                                 B. Braden, USC/ISI
Request for Comments: 2309                             D. Clark, MIT LCS
Category: Informational                                J. Crowcroft, UCL
                                                 B. Davie, Cisco Systems
                                               S. Deering, Cisco Systems
                                                          D. Estrin, USC
                                                          S. Floyd, LBNL
                                                       V. Jacobson, LBNL
                                                  G. Minshall, Fiberlane
                                                       C. Partridge, BBN
                                      L. Peterson, University of Arizona
                                      K. Ramakrishnan, ATT Labs Research
                                                  S. Shenker, Xerox PARC
                                                  J. Wroclawski, MIT LCS
                                                          L. Zhang, UCLA
                                                              April 1998



     Recommendations on Queue Management and Congestion Avoidance
                            in the Internet



Status of Memo

      This memo provides information for the Internet community.  It
      does not specify an Internet standard of any kind.  Distribution
      of this memo is unlimited.

Copyright Notice



      Copyright (C) The Internet Society (1998).  All Rights Reserved.

Abstract



      This memo presents two recommendations to the Internet community
      concerning measures to improve and preserve Internet performance.
      It presents a strong recommendation for testing, standardization,
      and widespread deployment of active queue management in routers,
      to improve the performance of today's Internet.  It also urges a
      concerted effort of research, measurement, and ultimate deployment
      of router mechanisms to protect the Internet from flows that are
      not sufficiently responsive to congestion notification.







Braden, et. al.              Informational                      [Page 1]

RFC 2309          Internet Performance Recommendations        April 1998


1. INTRODUCTION



   The Internet protocol architecture is based on a connectionless end-
   to-end packet service using the IP protocol.  The advantages of its
   connectionless design, flexibility and robustness, have been amply
   demonstrated.  However, these advantages are not without cost:
   careful design is required to provide good service under heavy load.
   In fact, lack of attention to the dynamics of packet forwarding can
   result in severe service degradation or "Internet meltdown".  This
   phenomenon was first observed during the early growth phase of the
   Internet of the mid 1980s [Nagle84], and is technically called
   "congestion collapse".

   The original fix for Internet meltdown was provided by Van Jacobson.
   Beginning in 1986, Jacobson developed the congestion avoidance
   mechanisms that are now required in TCP implementations [Jacobson88,
   HostReq89].  These mechanisms operate in the hosts to cause TCP
   connections to "back off" during congestion.  We say that TCP flows
   are "responsive" to congestion signals (i.e., dropped packets) from
   the network.  It is primarily these TCP congestion avoidance
   algorithms that prevent the congestion collapse of today's Internet.

   However, that is not the end of the story.  Considerable research has
   been done on Internet dynamics since 1988, and the Internet has
   grown.  It has become clear that the TCP congestion avoidance
   mechanisms [RFC2001], while necessary and powerful, are not
   sufficient to provide good service in all circumstances.  Basically,
   there is a limit to how much control can be accomplished from the
   edges of the network.  Some mechanisms are needed in the routers to
   complement the endpoint congestion avoidance mechanisms.

   It is useful to distinguish between two classes of router algorithms
   related to congestion control: "queue management" versus "scheduling"
   algorithms.  To a rough approximation, queue management algorithms
   manage the length of packet queues by dropping packets when necessary
   or appropriate, while scheduling algorithms determine which packet to
   send next and are used primarily to manage the allocation of
   bandwidth among flows.  While these two router mechanisms are closely
   related, they address rather different performance issues.

   This memo highlights two router performance issues.  The first issue
   is the need for an advanced form of router queue management that we
   call "active queue management."  Section 2 summarizes the benefits
   that active queue management can bring.  Section 3 describes a
   recommended active queue management mechanism, called Random Early
   Detection or "RED".  We expect that the RED algorithm can be used
   with a wide variety of scheduling algorithms, can be implemented
   relatively efficiently, and will provide significant Internet



Braden, et. al.              Informational                      [Page 2]

RFC 2309          Internet Performance Recommendations        April 1998


   performance improvement.

   The second issue, discussed in Section 4 of this memo, is the
   potential for future congestion collapse of the Internet due to flows
   that are unresponsive, or not sufficiently responsive, to congestion
   indications.  Unfortunately, there is no consensus solution to
   controlling congestion caused by such aggressive flows; significant
   research and engineering will be required before any solution will be
   available.  It is imperative that this work be energetically pursued,
   to ensure the future stability of the Internet.

   Section 5 concludes the memo with a set of recommendations to the
   IETF concerning these topics.

   The discussion in this memo applies to "best-effort" traffic.  The
   Internet integrated services architecture, which provides a mechanism
   for protecting individual flows from congestion, introduces its own
   queue management and scheduling algorithms [Shenker96, Wroclawski96].
   Similarly, the discussion of queue management and congestion control
   requirements for differential services is a separate issue.  However,
   we do not expect the deployment of integrated services and
   differential services to significantly diminish the importance of the
   best-effort traffic issues discussed in this memo.

   Preparation of this memo resulted from past discussions of end-to-end
   performance, Internet congestion, and RED in the End-to-End Research
   Group of the Internet Research Task Force (IRTF).

2. THE NEED FOR ACTIVE QUEUE MANAGEMENT



   The traditional technique for managing router queue lengths is to set
   a maximum length (in terms of packets) for each queue, accept packets
   for the queue until the maximum length is reached, then reject (drop)
   subsequent incoming packets until the queue decreases because a
   packet from the queue has been transmitted.  This technique is known
   as "tail drop", since the packet that arrived most recently (i.e.,
   the one on the tail of the queue) is dropped when the queue is full.
   This method has served the Internet well for years, but it has two
   important drawbacks.

   1.   Lock-Out

        In some situations tail drop allows a single connection or a few
        flows to monopolize queue space, preventing other connections
        from getting room in the queue.  This "lock-out" phenomenon is
        often the result of synchronization or other timing effects.





Braden, et. al.              Informational                      [Page 3]

RFC 2309          Internet Performance Recommendations        April 1998


   2.   Full Queues

        The tail drop discipline allows queues to maintain a full (or,
        almost full) status for long periods of time, since tail drop
        signals congestion (via a packet drop) only when the queue has
        become full.  It is important to reduce the steady-state queue
        size, and this is perhaps queue management's most important
        goal.

        The naive assumption might be that there is a simple tradeoff
        between delay and throughput, and that the recommendation that
        queues be maintained in a "non-full" state essentially
        translates to a recommendation that low end-to-end delay is more
        important than high throughput.  However, this does not take
        into account the critical role that packet bursts play in
        Internet performance.  Even though TCP constrains a flow's
        window size, packets often arrive at routers in bursts
        [Leland94].  If the queue is full or almost full, an arriving
        burst will cause multiple packets to be dropped.  This can
        result in a global synchronization of flows throttling back,
        followed by a sustained period of lowered link utilization,
        reducing overall throughput.

        The point of buffering in the network is to absorb data bursts
        and to transmit them during the (hopefully) ensuing bursts of
        silence.  This is essential to permit the transmission of bursty
        data.  It should be clear why we would like to have normally-
        small queues in routers: we want to have queue capacity to
        absorb the bursts.  The counter-intuitive result is that
        maintaining normally-small queues can result in higher
        throughput as well as lower end-to-end delay.  In short, queue
        limits should not reflect the steady state queues we want
        maintained in the network; instead, they should reflect the size
        of bursts we need to absorb.

   Besides tail drop, two alternative queue disciplines that can be
   applied when the queue becomes full are "random drop on full" or
   "drop front on full".  Under the random drop on full discipline, a
   router drops a randomly selected packet from the queue (which can be
   an expensive operation, since it naively requires an O(N) walk
   through the packet queue) when the queue is full and a new packet
   arrives.  Under the "drop front on full" discipline [Lakshman96], the
   router drops the packet at the front of the queue when the queue is
   full and a new packet arrives.  Both of these solve the lock-out
   problem, but neither solves the full-queues problem described above.






Braden, et. al.              Informational                      [Page 4]

RFC 2309          Internet Performance Recommendations        April 1998


   We know in general how to solve the full-queues problem for
   "responsive" flows, i.e., those flows that throttle back in response
   to congestion notification.  In the current Internet, dropped packets
   serve as a critical mechanism of congestion notification to end
   nodes.  The solution to the full-queues problem is for routers to
   drop packets before a queue becomes full, so that end nodes can
   respond to congestion before buffers overflow.  We call such a
   proactive approach "active queue management".  By dropping packets
   before buffers overflow, active queue management allows routers to
   control when and how many packets to drop.  The next section
   introduces RED, an active queue management mechanism that solves both
   problems listed above (given responsive flows).

   In summary, an active queue management mechanism can provide the
   following advantages for responsive flows.

   1.   Reduce number of packets dropped in routers

        Packet bursts are an unavoidable aspect of packet networks
        [Willinger95].  If all the queue space in a router is already
        committed to "steady state" traffic or if the buffer space is
        inadequate, then the router will have no ability to buffer
        bursts.  By keeping the average queue size small, active queue
        management will provide greater capacity to absorb naturally-
        occurring bursts without dropping packets.

        Furthermore, without active queue management, more packets will
        be dropped when a queue does overflow.  This is undesirable for
        several reasons.  First, with a shared queue and the tail drop
        discipline, an unnecessary global synchronization of flows
        cutting back can result in lowered average link utilization, and
        hence lowered network throughput.  Second, TCP recovers with
        more difficulty from a burst of packet drops than from a single
        packet drop.  Third, unnecessary packet drops represent a
        possible waste of bandwidth on the way to the drop point.

        We note that while RED can manage queue lengths and reduce end-
        to-end latency even in the absence of end-to-end congestion
        control, RED will be able to reduce packet dropping only in an
        environment that continues to be dominated by end-to-end
        congestion control.

   2.   Provide lower-delay interactive service

        By keeping the average queue size small, queue management will
        reduce the delays seen by flows.  This is particularly important
        for interactive applications such as short Web transfers, Telnet
        traffic, or interactive audio-video sessions, whose subjective



Braden, et. al.              Informational                      [Page 5]

RFC 2309          Internet Performance Recommendations        April 1998


        (and objective) performance is better when the end-to-end delay
        is low.

   3.   Avoid lock-out behavior



        Active queue management can prevent lock-out behavior by
        ensuring that there will almost always be a buffer available for
        an incoming packet.  For the same reason, active queue
        management can prevent a router bias against low bandwidth but
        highly bursty flows.

        It is clear that lock-out is undesirable because it constitutes
        a gross unfairness among groups of flows.  However, we stop
        short of calling this benefit "increased fairness", because
        general fairness among flows requires per-flow state, which is
        not provided by queue management.  For example, in a router
        using queue management but only FIFO scheduling, two TCP flows
        may receive very different bandwidths simply because they have
        different round-trip times [Floyd91], and a flow that does not
        use congestion control may receive more bandwidth than a flow
        that does.  Per-flow state to achieve general fairness might be
        maintained by a per-flow scheduling algorithm such as Fair
        Queueing (FQ) [Demers90], or a class-based scheduling algorithm
        such as CBQ [Floyd95], for example.

        On the other hand, active queue management is needed even for
        routers that use per-flow scheduling algorithms such as FQ or
        class-based scheduling algorithms such as CBQ.  This is because
        per-flow scheduling algorithms by themselves do nothing to
        control the overall queue size or the size of individual queues.
        Active queue management is needed to control the overall average
        queue sizes, so that arriving bursts can be accommodated without
        dropping packets.  In addition, active queue management should
        be used to control the queue size for each individual flow or
        class, so that they do not experience unnecessarily high delays.
        Therefore, active queue management should be applied across the
        classes or flows as well as within each class or flow.

        In short, scheduling algorithms and queue management should be
        seen as complementary, not as replacements for each other.  In
        particular, there have been implementations of queue management
        added to FQ, and work is in progress to add RED queue management
        to CBQ.








Braden, et. al.              Informational                      [Page 6]

RFC 2309          Internet Performance Recommendations        April 1998


3. THE QUEUE MANAGEMENT ALGORITHM "RED"

   Random Early Detection, or RED, is an active queue management
   algorithm for routers that will provide the Internet performance
   advantages cited in the previous section [RED93].  In contrast to
   traditional queue management algorithms, which drop packets only when
   the buffer is full, the RED algorithm drops arriving packets
   probabilistically.  The probability of drop increases as the
   estimated average queue size grows.  Note that RED responds to a
   time-averaged queue length, not an instantaneous one.  Thus, if the
   queue has been mostly empty in the "recent past", RED won't tend to
   drop packets (unless the queue overflows, of course!). On the other
   hand, if the queue has recently been relatively full, indicating
   persistent congestion, newly arriving packets are more likely to be
   dropped.

   The RED algorithm itself consists of two main parts: estimation of
   the average queue size and the decision of whether or not to drop an
   incoming packet.


   (a) Estimation of Average Queue Size

        RED estimates the average queue size, either in the forwarding
        path using a simple exponentially weighted moving average (such
        as presented in Appendix A of [Jacobson88]), or in the
        background (i.e., not in the forwarding path) using a similar
        mechanism.

           Note: The queue size can be measured either in units of
           packets or of bytes.  This issue is discussed briefly in
           [RED93] in the "Future Work" section.

           Note: when the average queue size is computed in the
           forwarding path, there is a special case when a packet
           arrives and the queue is empty.  In this case, the
           computation of the average queue size must take into account
           how much time has passed since the queue went empty.  This is
           discussed further in [RED93].


   (b) Packet Drop Decision

        In the second portion of the algorithm, RED decides whether or
        not to drop an incoming packet.  It is RED's particular
        algorithm for dropping that results in performance improvement
        for responsive flows.  Two RED parameters, minth (minimum
        threshold) and maxth (maximum threshold), figure prominently in



Braden, et. al.              Informational                      [Page 7]

RFC 2309          Internet Performance Recommendations        April 1998


        this decision process.  Minth specifies the average queue size
        *below which* no packets will be dropped, while maxth specifies
        the average queue size *above which* all packets will be
        dropped.  As the average queue size varies from minth to maxth,
        packets will be dropped with a probability that varies linearly
        from 0 to maxp.

           Note: a simplistic method of implementing this would be to
           calculate a new random number at each packet arrival, then
           compare that number with the above probability which varies
           from 0 to maxp.  A more efficient implementation, described
           in [RED93], computes a random number *once* for each dropped
           packet.

           Note: the decision whether or not to drop an incoming packet
           can be made in "packet mode", ignoring packet sizes, or in
           "byte mode", taking into account the size of the incoming
           packet.  The performance implications of the choice between
           packet mode or byte mode is discussed further in [Floyd97].

   RED effectively controls the average queue size while still
   accommodating bursts of packets without loss.  RED's use of
   randomness breaks up synchronized processes that lead to lock-out
   phenomena.

   There have been several implementations of RED in routers, and papers
   have been published reporting on experience with these
   implementations ([Villamizar94], [Gaynor96]).  Additional reports of
   implementation experience would be welcome, and will be posted on the
   RED web page [REDWWW].

   All available empirical evidence shows that the deployment of active
   queue management mechanisms in the Internet would have substantial
   performance benefits.  There are seemingly no disadvantages to using
   the RED algorithm, and numerous advantages.  Consequently, we believe
   that the RED active queue management algorithm should be widely
   deployed.

   We should note that there are some extreme scenarios for which RED
   will not be a cure, although it won't hurt and may still help.  An
   example of such a scenario would be a very large number of flows,
   each so tiny that its fair share would be less than a single packet
   per RTT.








Braden, et. al.              Informational                      [Page 8]

RFC 2309          Internet Performance Recommendations        April 1998


4. MANAGING AGGRESSIVE FLOWS



   One of the keys to the success of the Internet has been the
   congestion avoidance mechanisms of TCP.  Because TCP "backs off"
   during congestion, a large number of TCP connections can share a
   single, congested link in such a way that bandwidth is shared
   reasonably equitably among similarly situated flows.  The equitable
   sharing of bandwidth among flows depends on the fact that all flows
   are running basically the same congestion avoidance algorithms,
   conformant with the current TCP specification [HostReq89].

   We introduce the term "TCP-compatible" for a flow that behaves under
   congestion like a flow produced by a conformant TCP.  A TCP-
   compatible flow is responsive to congestion notification, and in
   steady-state it uses no more bandwidth than a conformant TCP running
   under comparable conditions (drop rate, RTT, MTU, etc.)

   It is convenient to divide flows into three classes: (1) TCP-
   compatible flows, (2) unresponsive flows, i.e., flows that do not
   slow down when congestion occurs, and (3) flows that are responsive
   but are not TCP-compatible.  The last two classes contain more
   aggressive flows that pose significant threats to Internet
   performance, as we will now discuss.

   o    Non-Responsive Flows

        There is a growing set of UDP-based applications whose
        congestion avoidance algorithms are inadequate or nonexistent
        (i.e, the flow does not throttle back upon receipt of congestion
        notification).  Such UDP applications include streaming
        applications like packet voice and video, and also multicast
        bulk data transport [SRM96].  If no action is taken, such
        unresponsive flows could lead to a new congestion collapse.

        In general, all UDP-based streaming applications should
        incorporate effective congestion avoidance mechanisms.  For
        example, recent research has shown the possibility of
        incorporating congestion avoidance mechanisms such as Receiver-
        driven Layered Multicast (RLM) within UDP-based streaming
        applications such as packet video [McCanne96; Bolot94]. Further
        research and development on ways to accomplish congestion
        avoidance for streaming applications will be very important.

        However, it will also be important for the network to be able to
        protect itself against unresponsive flows, and mechanisms to
        accomplish this must be developed and deployed.  Deployment of
        such mechanisms would provide incentive for every streaming
        application to become responsive by incorporating its own



Braden, et. al.              Informational                      [Page 9]

RFC 2309          Internet Performance Recommendations        April 1998


        congestion control.

   o    Non-TCP-Compatible Transport Protocols

        The second threat is posed by transport protocol implementations
        that are responsive to congestion notification but, either
        deliberately or through faulty implementations, are not TCP-
        compatible.  Such applications can grab an unfair share of the
        network bandwidth.

        For example, the popularity of the Internet has caused a
        proliferation in the number of TCP implementations.  Some of
        these may fail to implement the TCP congestion avoidance
        mechanisms correctly because of poor implementation.  Others may
        deliberately be implemented with congestion avoidance algorithms
        that are more aggressive in their use of bandwidth than other
        TCP implementations; this would allow a vendor to claim to have
        a "faster TCP".  The logical consequence of such implementations
        would be a spiral of increasingly aggressive TCP
        implementations, leading back to the point where there is
        effectively no congestion avoidance and the Internet is
        chronically congested.

        Note that there is a well-known way to achieve more aggressive
        TCP performance without even changing TCP: open multiple
        connections to the same place, as has been done in some Web
        browsers.

   The projected increase in more aggressive flows of both these
   classes, as a fraction of total Internet traffic, clearly poses a
   threat to the future Internet.  There is an urgent need for
   measurements of current conditions and for further research into the
   various ways of managing such flows.  There are many difficult issues
   in identifying and isolating unresponsive or non-TCP-compatible flows
   at an acceptable router overhead cost.  Finally, there is little
   measurement or simulation evidence available about the rate at which
   these threats are likely to be realized, or about the expected
   benefit of router algorithms for managing such flows.

   There is an issue about the appropriate granularity of a "flow".
   There are a few "natural" answers: 1) a TCP or UDP connection (source
   address/port, destination address/port); 2) a source/destination host
   pair; 3) a given source host or a given destination host.  We would
   guess that the source/destination host pair gives the most
   appropriate granularity in many circumstances.  However, it is
   possible that different vendors/providers could set different
   granularities for defining a flow (as a way of "distinguishing"
   themselves from one another), or that different granularities could



Braden, et. al.              Informational                     [Page 10]

RFC 2309          Internet Performance Recommendations        April 1998


   be chosen for different places in the network.  It may be the case
   that the granularity is less important than the fact that we are
   dealing with more unresponsive flows at *some* granularity.  The
   granularity of flows for congestion management is, at least in part,
   a policy question that needs to be addressed in the wider IETF
   community.

5. CONCLUSIONS AND RECOMMENDATIONS



   This discussion leads us to make the following recommendations to the
   IETF and to the Internet community as a whole.

   o    RECOMMENDATION 1:

        Internet routers should implement some active queue management
        mechanism to manage queue lengths, reduce end-to-end latency,
        reduce packet dropping, and avoid lock-out phenomena within the
        Internet.

        The default mechanism for managing queue lengths to meet these
        goals in FIFO queues is Random Early Detection (RED) [RED93].
        Unless a developer has reasons to provide another equivalent
        mechanism, we recommend that RED be used.

   o    RECOMMENDATION 2:

        It is urgent to begin or continue research, engineering, and
        measurement efforts contributing to the design of mechanisms to
        deal with flows that are unresponsive to congestion notification
        or are responsive but more aggressive than TCP.

   Although there has already been some limited deployment of RED in the
   Internet, we may expect that widespread implementation and deployment
   of RED in accordance with Recommendation 1 will expose a number of
   engineering issues.  For example, such issues may include:
   implementation questions for Gigabit routers, the use of RED in layer
   2 switches, and the possible use of additional considerations, such
   as priority, in deciding which packets to drop.

   We again emphasize that the widespread implementation and deployment
   of RED would not, in and of itself, achieve the goals of
   Recommendation 2.

   Widespread implementation and deployment of RED will also enable the
   introduction of other new functionality into the Internet.  One
   example of an enabled functionality would be the addition of explicit
   congestion notification [Ramakrishnan97] to the Internet
   architecture, as a mechanism for congestion notification in addition



Braden, et. al.              Informational                     [Page 11]

RFC 2309          Internet Performance Recommendations        April 1998


   to packet drops.  A second example of new functionality would be
   implementation of queues with packets of different drop priorities;
   packets would be transmitted in the order in which they arrived, but
   during times of congestion packets of the lower drop priority would
   be preferentially dropped.

6. References



   [Bolot94] Bolot, J.-C., Turletti, T., and Wakeman, I., Scalable
   Feedback Control for Multicast Video Distribution in the Internet,
   ACM SIGCOMM '94, Sept. 1994.

   [Demers90] Demers, A., Keshav, S., and Shenker, S., Analysis and
   Simulation of a Fair Queueing Algorithm, Internetworking: Research
   and Experience, Vol. 1, 1990, pp. 3-26.

   [Floyd91] Floyd, S., Connections with Multiple Congested Gateways in
   Packet-Switched Networks Part 1: One-way Traffic.  Computer
   Communications Review, Vol.21, No.5, October 1991, pp.  30-47.  URL
   http://ftp.ee.lbl.gov/floyd/.

   [Floyd95] Floyd, S., and Jacobson, V., Link-sharing and Resource
   Management Models for Packet Networks. IEEE/ACM Transactions on
   Networking, Vol. 3 No. 4, pp. 365-386, August 1995.

   [Floyd97] Floyd, S., RED: Discussions of Byte and Packet Modes, March
   1997 email, http://www-nrg.ee.lbl.gov/floyd/REDaveraging.txt.

   [Gaynor96] Gaynor, M., Proactive Packet Dropping Methods for TCP
   Gateways, October 1996, URL http://www.eecs.harvard.edu/~gaynor/
   final.ps.

   [HostReq89] Braden, R., Ed., "Requirements for Internet Hosts --
   Communication Layers", STD 3, RFC 1122, October 1989.

   [Jacobson88] V. Jacobson, Congestion Avoidance and Control, ACM
   SIGCOMM '88, August 1988.

   [Lakshman96] T. V. Lakshman, Arnie Neidhardt, Teunis Ott, The Drop
   From Front Strategy in TCP Over ATM and Its Interworking with Other
   Control Features, Infocom 96, MA28.1.

   [Leland94] W. Leland, M. Taqqu, W. Willinger, and D. Wilson, On the
   Self-Similar Nature of Ethernet Traffic (Extended Version), IEEE/ACM
   Transactions on Networking, 2(1), pp. 1-15, February 1994.






Braden, et. al.              Informational                     [Page 12]

RFC 2309          Internet Performance Recommendations        April 1998


   [McCanne96] McCanne, S., Jacobson, V., and M. Vetterli, Receiver-
   driven Layered Multicast, ACM SIGCOMM

   [Nagle84] Nagle, J., "Congestion Control in IP/TCP", RFC 896, January
   1984.

   [Ramakrishnan97] Ramakrishnan, K. K., and S. Floyd, "A Proposal to
   add Explicit Congestion Notification (ECN) to IPv6 and to TCP", Work
   in Progress.

   [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.

   [REDWWW] Floyd, S., The RED Web Page, 1997, URL
   http://ftp.ee.lbl.gov/floyd/red.html.

   [RFC 2001] Stevens, W., "TCP Slow Start, Congestion Avoidance, Fast
   Retransmit, and Fast Recovery Algorithms", RFC 2001, January 1997.

   [Shenker96] Shenker, S., Partridge, C., and R. Guerin, "Specification
   of Guaranteed Quality of Service", Work in Progress.

   [SRM96] Floyd. S., Jacobson, V., McCanne, S., Liu, C., and L. Zhang,
   A Reliable Multicast Framework for Light-weight Sessions and
   Application Level Framing.  ACM SIGCOMM '96, pp 342-355.

   [Villamizar94] Villamizar, C., and Song, C., High Performance TCP in
   ANSNET. Computer Communications Review, V. 24 N. 5, October 1994, pp.
   45-60.  URL http://ftp.ans.net/pub/papers/tcp-performance.ps.

   [Willinger95] W. Willinger, M. S. Taqqu, R. Sherman, D. V.  Wilson,
   Self-Similarity Through High-Variability:  Statistical Analysis of
   Ethernet LAN Traffic at the Source Level, ACM SIGCOMM '95, pp.  100-
   113, August 1995.

   [Wroclawski96] Wroclawski, J., "Specification of the Controlled-Load
   Network Element Service", Work in Progress.












Braden, et. al.              Informational                     [Page 13]

RFC 2309          Internet Performance Recommendations        April 1998


Security Considerations

   While security is a very important issue, it is largely orthogonal to
   the performance issues discussed in this memo.  We note, however,
   that denial-of-service attacks may create unresponsive traffic flows
   that are indistinguishable from flows from normal high-bandwidth
   isochronous applications, and the mechanism suggested in
   Recommendation 2 will be equally applicable to such attacks.

Authors' Addresses



   Bob Braden
   USC Information Sciences Institute
   4676 Admiralty Way
   Marina del Rey, CA 90292

   Phone: 310-822-1511
   EMail: Braden@ISI.EDU

   David D. Clark
   MIT Laboratory for Computer Science
   545 Technology Sq.
   Cambridge, MA  02139

   Phone: 617-253-6003
   EMail: DDC@lcs.mit.edu

   Jon Crowcroft
   University College London
   Department of Computer Science
   Gower Street
   London, WC1E 6BT
   ENGLAND

   Phone: +44 171 380 7296
   EMail: Jon.Crowcroft@cs.ucl.ac.uk

   Bruce Davie
   Cisco Systems, Inc.
   250 Apollo Drive
   Chelmsford, MA 01824

   Phone:
   EMail: bdavie@cisco.com







Braden, et. al.              Informational                     [Page 14]

RFC 2309          Internet Performance Recommendations        April 1998


   Steve Deering
   Cisco Systems, Inc.
   170 West Tasman Drive
   San Jose, CA 95134-1706

   Phone: 408-527-8213
   EMail: deering@cisco.com

   Deborah Estrin
   USC Information Sciences Institute
   4676 Admiralty Way
   Marina del Rey, CA 90292

   Phone: 310-822-1511
   EMail: Estrin@usc.edu

   Sally Floyd
   Lawrence Berkeley National Laboratory,
   MS 50B-2239,
   One Cyclotron Road,
   Berkeley CA 94720

   Phone:  510-486-7518
   EMail: Floyd@ee.lbl.gov

   Van Jacobson
   Lawrence Berkeley National Laboratory,
   MS 46A,
   One Cyclotron Road,
   Berkeley CA 94720

   Phone: 510-486-7519
   EMail: Van@ee.lbl.gov

   Greg Minshall
   Fiberlane Communications
   1399 Charleston Road
   Mountain View, CA  94043

   Phone:  +1 650 237 3164
   EMail:  Minshall@fiberlane.com










Braden, et. al.              Informational                     [Page 15]

RFC 2309          Internet Performance Recommendations        April 1998


   Craig Partridge
   BBN Technologies
   10 Moulton St.
   Cambridge MA 02138

   Phone: 510-558-8675
   EMail: craig@bbn.com

   Larry Peterson
   Department of Computer Science
   University of Arizona
   Tucson, AZ 85721

   Phone: 520-621-4231
   EMail: LLP@cs.arizona.edu

   K. K. Ramakrishnan
   AT&T Labs. Research
   Rm. A155
   180 Park Avenue
   Florham Park, N.J. 07932



   Phone: 973-360-8766
   EMail: KKRama@research.att.com

   Scott Shenker
   Xerox PARC
   3333 Coyote Hill Road
   Palo Alto, CA 94304

   Phone: 415-812-4840
   EMail: Shenker@parc.xerox.com

   John Wroclawski
   MIT Laboratory for Computer Science
   545 Technology Sq.
   Cambridge, MA  02139

   Phone: 617-253-7885
   EMail: JTW@lcs.mit.edu

   Lixia Zhang
   UCLA
   4531G Boelter Hall
   Los Angeles, CA 90024

   Phone: 310-825-2695
   EMail: Lixia@cs.ucla.edu



Braden, et. al.              Informational                     [Page 16]

RFC 2309          Internet Performance Recommendations        April 1998


Full Copyright Statement



   Copyright (C) The Internet Society (1998).  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.
























Braden, et. al.              Informational                     [Page 17]