rfc1254.txt

RFC 的详细文档！

Network Working Group                                          A. Mankin
Request for Comments: 1254                                         MITRE
                                                         K. Ramakrishnan
                                           Digital Equipment Corporation
                                                                 Editors
                                                             August 1991


                   Gateway Congestion Control Survey

Status of this Memo

   This memo provides information for the Internet community.  It is a
   survey of some of the major directions and issues.  It does not
   specify an Internet standard.  Distribution of this memo is
   unlimited.

Abstract

   The growth of network intensive Internet applications has made
   gateway congestion control a high priority.  The IETF Performance and
   Congestion Control Working Group surveyed and reviewed gateway
   congestion control and avoidance approaches.  The purpose of this
   paper is to present our review of the congestion control approaches,
   as a way of encouraging new discussion and experimentation.  Included
   in the survey are Source Quench, Random Drop, Congestion Indication
   (DEC Bit), and Fair Queueing.  The task remains for Internet
   implementors to determine and agree on the most effective mechanisms
   for controlling gateway congestion.

1.  Introduction

   Internet users regularly encounter congestion, often in mild forms.
   However, severe congestion episodes have been reported also; and
   gateway congestion remains an obstacle for Internet applications such
   as scientific supercomputing data transfer.  The need for Internet
   congestion control originally became apparent during several periods
   of 1986 and 1987, when the Internet experienced the "congestion
   collapse" condition predicted by Nagle [Nag84].  A large number of
   widely dispersed Internet sites experienced simultaneous slowdown or
   cessation of networking services for prolonged periods.  BBN, the
   firm responsible for maintaining the then backbone of the Internet,
   the ARPANET, responded to the collapse by adding link capacity
   [Gar87].

   Much of the Internet now uses as a transmission backbone the National
   Science Foundation Network (NSFNET). Extensive monitoring and
   capacity planning are being done for the NSFNET backbone; still, as



Performance and Congestion Control Working Group                [Page 1]

RFC 1254           Gateway Congestion Control Survey         August 1991


   the demand for this capacity grows, and as resource-intensive
   applications such as wide-area file system management [Sp89]
   increasingly use the backbone, effective congestion control policies
   will be a critical requirement.

   Only a few mechanisms currently exist in Internet hosts and gateways
   to avoid or control congestion.  The mechanisms for handling
   congestion set forth in the specifications for the DoD Internet
   protocols are limited to:

      Window flow control in TCP [Pos81b], intended primarily for
      controlling the demand on the receiver's capacity, both in terms
      of processing and buffers.

      Source quench in ICMP, the message sent by IP to request that a
      sender throttle back [Pos81a].

   One approach to enhancing Internet congestion control has been to
   overlay the simple existing mechanisms in TCP and ICMP with more
   powerful ones.  Since 1987, the TCP congestion control policy, Slow-
   start, a collection of several algorithms developed by Van Jacobson
   and Mike Karels [Jac88], has been widely adopted. Successful Internet
   experiences with Slow-start led to the Host Requirements RFC [HREQ89]
   classifying the algorithms as mandatory for TCP.  Slow-start modifies
   the user's demand when congestion reaches such a point that packets
   are dropped at the gateway.  By the time such overflows occur, the
   gateway is congested.  Jacobson writes that the Slow-start policy is
   intended to function best with a complementary gateway policy
   [Jac88].

1.1  Definitions

   The characteristics of the Internet that we are interested in include
   that it is, in general, an arbitrary mesh-connected network.  The
   internetwork protocol is connectionless.  The number of users that
   place demands on the network is not limited by any explicit
   mechanism; no reservation of resources occurs and transport layer
   set-ups are not disallowed due to lack of resources.  A path from a
   source to destination host may have multiple hops, through several
   gateways and links.  Paths through the Internet may be heterogeneous
   (though homogeneous paths also exist and experience congestion).
   That is, links may be of different speeds.  Also, the gateways and
   hosts may be of different speeds or may be providing only a part of
   their processing power to communication-related activity.  The
   buffers for storing information flowing through Internet gateways are
   finite.  The nature of the internet protocol is to drop packets when
   these buffers overflow.




Performance and Congestion Control Working Group                [Page 2]

RFC 1254           Gateway Congestion Control Survey         August 1991


   Gateway congestion arises when the demand for one or more of the
   resources of the gateway exceeds the capacity of that resource.  The
   resources include transmission links, processing, and space used for
   buffering.  Operationally, uncongested gateways operate with little
   queueing on average, where the queue is the waiting line for a
   particular resource of the gateway.  One commonly used quantitative
   definition [Kle79] for when a resource is congested is when the
   operating point is greater than the point at which resource power is
   maximum, where resource power is defined as the ratio of throughput
   to delay (See Section 2.2).  At this operating point, the average
   queue size is close to one, including the packet in service.  Note
   that this is a long-term average queue size.  Several definitions
   exist for the timescale of averaging for congestion detection and
   control, such as dominant round-trip time and queue regeneration
   cycle (see Section 2.1).

   Mechanisms for handling congestion may be divided into two
   categories, congestion recovery and congestion avoidance.  Congestion
   recovery tries to restore an operating state, when demand has already
   exceeded capacity.  Congestion avoidance is preventive in nature.  It
   tries to keep the demand on the network at or near the point of
   maximum power, so that congestion never occurs.  Without congestion
   recovery, the network may cease to operate entirely (zero
   throughput), whereas the Internet has been operating without
   congestion avoidance for a long time.  Overall performance may
   improve with an effective congestion avoidance mechanism.  Even if
   effective congestion avoidance was in place, congestion recovery
   schemes would still be required, to retain throughput in the face of
   sudden changes (increase of demand, loss of resources) that can lead
   to congestion.

   In this paper, the term "user" refers to each individual transport
   (TCP or another transport protocol) entity.  For example, a TCP
   connection is a "user" in this terminology.  The terms "flow" and
   "stream" are used by some authors in the same sense.  Some of the
   congestion control policies discussed in this paper, such as
   Selective Feedback Congestion Indication and Fair Queueing aggregate
   multiple TCP connections from a single host (or between a source
   host-destination host pair) as a virtual user.

   The term "cooperating transport entities" will be defined as a set of
   TCP connections (for example) which follow an effective method of
   adjusting their demand on the Internet in response to congestion.
   The most restrictive interpretation of this term is that the
   transport entities follow identical algorithms for congestion control
   and avoidance.  However, there may be some variation in these
   algorithms.  The extent to which heterogeneous end-system congestion
   control and avoidance may be accommodated by gateway policies should



Performance and Congestion Control Working Group                [Page 3]

RFC 1254           Gateway Congestion Control Survey         August 1991


   be a subject of future research. The role played in Internet
   performance of non-cooperating transport entities is discussed in
   Section 5.

1.2  Goals and Scope of This Paper

   The task remains for Internet implementors to determine effective
   mechanisms for controlling gateway congestion.  There has been
   minimal common practice on which to base recommendations for Internet
   gateway congestion control.  In this survey, we describe the
   characteristics of one experimental gateway congestion management
   policy, Random Drop, and several that are better-known:  Source
   Quench, Congestion Indication, Selective Feedback Congestion
   Indication, and Fair Queueing, both Bit-Round and Stochastic.  A
   motivation for documenting Random Drop is that it has as primary
   goals low overhead and suitability for scaling up for Internets with
   higher speed links.  Both of these are important goals for future
   gateway implementations that will have fast links, fast processors,
   and will have to serve large numbers of interconnected hosts.

   The structure of this paper is as follows.  First, we discuss
   performance goals, including timescale and fairness considerations.
   Second, we discuss the gateway congestion control policies.  Random
   Drop is sketched out, with a recommendation for using it for
   congestion recovery and a separate section on its use as congestion
   avoidance.  Third, since gateway congestion control in itself does
   not change the end-systems' demand, we briefly present the effective
   responses to these policies by two end-system congestion control
   schemes, Slow-start and End-System Congestion Indication.  Among our
   conclusions, we address the issues of transport entities that do not
   cooperate with gateway congestion control.  As an appendix, because
   of the potential interactions with gateway congestion policies, we
   report on a scheme to help in controlling the performance of Internet
   gateways to connection-oriented subnets (in particular, X.25).

   Resources in the current Internet are not charged to users of them.
   Congestion avoidance techniques cannot be expected to help when users
   increase beyond the capacity of the underlying facilities.  There are
   two possible solutions for this, increase the facilities and
   available bandwidth, or forcibly reduce the demand.  When congestion
   is persistent despite implemented congestion control mechanisms,
   administrative responses are needed.  These are naturally not within
   the scope of this paper.  Also outside the scope of this paper are
   routing techniques that may be used to relocate demand away from
   congested individual resources (e.g., path-splitting and load-
   balancing).





Performance and Congestion Control Working Group                [Page 4]

RFC 1254           Gateway Congestion Control Survey         August 1991


2.  Performance Goals

   To be able to discuss design and use of various mechanisms for
   improving Internetwork performance, we need to have clear performance
   goals for the operation of gateways and sets of end-systems.
   Internet experience shows that congestion control should be based on
   adaptive principles; this requires efficient computation of metrics
   by algorithms for congestion control.  The first issue is that of the
   interval over which these metrics are estimated and/or measured.

2.1  Interval for Measurement/Estimation of Performance Metrics

   Network performance metrics may be distorted if they are computed
   over intervals that are too short or too long relative to the dynamic
   characteristics of the network.  For instance, within a small
   interval, two FTP users with equal paths may appear to have sharply
   different demands, as an effect of brief, transient fluctuations in
   their respective processing.  An overly long averaging interval
   results in distortions because of the changing number of users
   sharing the resource measured during the time.  It is similarly
   important for congestion control mechanisms exerted at end systems to
   find an appropriate interval for control.

   The first approach to the monitoring, or averaging, interval for