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
congestion control is one based on round-trip times. The rationale
for it is as follows: control mechanisms must adapt to changes in
Internet congestion as quickly as possible. Even on an uncongested
path, changed conditions will not be detected by the sender faster
than a round-trip time. The effect of a sending end-system's control
will also not be seen in less than a round-trip time in the entire
path as well as at the end systems. For the control mechanism to be
adaptive, new information on the path is needed before making a
modification to the control exerted. The statistics and metrics used
in congestion control must be able to provide information to the
control mechanism so that it can make an informed decision.
Transient information which may be obsolete before a change is made
by the end-system should be avoided. This implies the
monitoring/estimating interval is one lasting one or more round
trips. The requirements described here give bounds on:
How short an interval: not small enough that obsolete information
is used for control;
How long: not more than the period at which the end-system makes
changes.
But, from the point of view of the gateway congestion control policy,
what is a round-trip time? If all the users of a given gateway have
Performance and Congestion Control Working Group [Page 5]
RFC 1254 Gateway Congestion Control Survey August 1991
the same path through the Internet, they also have the same round-
trip time through the gateway. But this is rarely the case.
A meaningful interval must be found for users with both short and
long paths. Two approaches have been suggested for estimating this
dynamically, queue regeneration cycle and frequency analysis.
Use of the queue regeneration cycle has been described as part of the
Congestion Indication policy. The time period used for averaging
here begins when a resource goes from the idle to busy state. The
basic interval for averaging is a "regeneration cycle" which is in
the form of busy and idle intervals. Because an average based on a
single previous regeneration may become old information, the
recommendation in [JRC87] is to average over the sum of two
intervals, that is, the previous (busy and idle) period, and the time
since the beginning of the current busy period.
If the gateway users are window-based transport entities, it is
possible to see how the regeneration interval responds to their
round-trip times. If a user with a long round-trip time has the
dominant traffic, the queue length may be zero only when that user is
awaiting acknowledgements. Then the users with short paths will
receive gateway congestion information that is averaged over several
of their round-trip times. If the short path traffic dominates the
activity in the gateway, i.e., the connections with shorter round-
trip times are the dominant users of the gateway resources, then the
regeneration interval is shorter and the information communicated to
them can be more timely. In this case, users with longer paths
receive, in one of their round-trip times, multiple samples of the
dominant traffic; the end system averaging is based on individual
user's intervals, so that these multiple samples are integrated
appropriately for these connections with longer paths.
The use of frequency analysis has been described by [Jac89]. In this
approach, the gateway congestion control is done at intervals based
on spectral analysis of the traffic arrivals. It is possible for
users to have round-trip times close to each other, but be out of
phase from each other. A spectral analysis algorithm detects this.
Otherwise, if multiple round-trip times are significant, multiple
intervals will be identified. Either one of these will be
predominant, or several will be comparable. An as yet difficult
problem for the design of algorithms accomplishing this technique is
the likelihood of "locking" to the frequency of periodic traffic of
low intensity, such as routing updates.
Performance and Congestion Control Working Group [Page 6]
RFC 1254 Gateway Congestion Control Survey August 1991
2.2 Power and its Relationship to the Operating Point
Performance goals for a congestion control/avoidance strategy embody
a conflict in that they call for as high a throughput as possible,
with as little delay as possible. A measure that is often used to
reflect the tradeoff between these goals is power, the ratio of
throughput to delay. We would like to maximize the value of power
for a given resource. In the standard expression for power,
Power = (Throughput^alpha)/Delay
the exponent alpha is chosen for throughput, based on the relative
emphasis placed on throughput versus delay: if throughput is more
important, then a value of A alpha greater than one is chosen. If
throughput and delay are equally important (e.g., both bulk transfer
traffic and interactive traffic are equally important), then alpha
equal to one is chosen. The operating point where power is maximized
is the "knee" in the throughput and delay curves. It is desirable
that the operating point of the resource be driven towards the knee,
where power is maximized. A useful property of power is that it is
decreasing whether the resource is under- or over-utilized relative
to the knee.
In an internetwork comprising nodes and links of diverse speeds and
utilization, bottlenecks or concentrations of demand may form. Any
particular user can see a single bottleneck, which is the slowest or
busiest link or gateway in the path (or possibly identical "balanced"
bottlenecks). The throughput that the path can sustain is limited by
the bottleneck. The delay for packets through a particular path is
determined by the service times and queueing at each individual hop.
The queueing delay is dominated by the queueing at the bottleneck
resource(s). The contribution to the delay over other hops is
primarily the service time, although the propagation delay over
certain hops, such as a satellite link, can be significant. We would
like to operate all shared resources at their knee and maximize the
power of every user's bottleneck.
The above goal underscores the significance of gateway congestion
control. If techniques can be found to operate gateways at their
resource knee, it can improve Internet performance broadly.
2.3 Fairness
We would like gateways to allocate resources fairly to users. A
concept of fairness is only relevant when multiple users share a
gateway and their total demand is greater than its capacity. If
demands were equal, a fair allocation of the resource would be to
provide an equal share to each user. But even over short intervals,
Performance and Congestion Control Working Group [Page 7]
RFC 1254 Gateway Congestion Control Survey August 1991
demands are not equal. Identifying the fair share of the resource
for the user becomes hard. Having identified it, it is desirable to
allocate at least this fair share to each user. However, not all
users may take advantage of this allocation. The unused capacity can
be given to other users. The resulting final allocation is termed a
maximally fair allocation. [RJC87] gives a quantitative method for
comparing the allocation by a given policy to the maximally fair
allocation.
It is known that the Internet environment has heterogeneous transport
entities, which do not follow the same congestion control policies
(our definition of cooperating transports). Then, the controls given
by a gateway may not affect all users and the congestion control
policy may have unequal effects. Is "fairness" obtainable in such a
heterogeneous community? In Fair Queueing, transport entities with
differing congestion control policies can be insulated from each
other and each given a set share of gateway bandwidth.
It is important to realize that since Internet gateways cannot refuse
new users, fairness in gateway congestion control can lead to all
users receiving small (sub-divided) amounts of the gateway resources
inadequate to meet their performance requirements. None of the
policies described in this paper currently addresses this. Then,
there may be policy reasons for unequal allocation of the gateway
resources. This has been addressed by Bit-Round Fair Queueing.
2.4 Network Management
Network performance goals may be assessed by measurements in either
the end-system or gateway frame of reference. Performance goals are
often resource-centered and the measurement of the global performance
of "the network," is not only difficult to measure but is also
difficult to define. Resource-centered metrics are more easily
obtained, and do not require synchronization. That resource-centered
metrics are appropriate ones for use in optimization of power is
shown by [Jaf81].
It would be valuable for the goal of developing effective gateway
congestion handling if Management Information Base (MIB) objects
useful for evaluating gateway congestion were developed. The
reflections on the control interval described above should be applied
when network management applications are designed for this purpose.
In particular, obtaining an instantaneous queue length from the
managed gateway is not meaningful for the purposes of congestion
management.
Performance and Congestion Control Working Group [Page 8]
RFC 1254 Gateway Congestion Control Survey August 1991
3. Gateway Congestion Control Policies
There have been proposed a handful of approaches to dealing with
congestion in the gateway. Some of these are Source Quench, Random
Drop, Congestion Indication, Selective Feedback Congestion
Indication, Fair Queueing, and Bit-Round Fair Queueing. They differ
in whether they use a control message, and indeed, whether they view
control of the end-systems as necessary, but none of them in itself
lowers the demand of users and consequent load on the network. End-
system policies that reduce demand in conjunction with gateway
congestion control are described in Section 4.
3.1 Source Quench
The method of gateway congestion control currently used in the
Internet is the Source Quench message of the RFC-792 [Pos81a]
Internet Control Message Protocol (ICMP). When a gateway responds to
congestion by dropping datagrams, it may send an ICMP Source Quench
message to the source of the dropped datagram. This is a congestion
recovery policy.
The Gateway Requirements RFC, RFC-1009 [GREQ87], specifies that
gateways should only send Source Quench messages with a limited
frequency, to conserve CPU resources during the time of heavy load.
We note that operating the gateway for long periods under such loaded
conditions should be averted by a gateway congestion control policy.
A revised Gateway Requirements RFC is being prepared by the IETF.
Another significant drawback of the Source Quench policy is that its
details are discretionary, or, alternatively, that the policy is
really a family of varied policies. Major Internet gateway
manufacturers have implemented a variety of source quench
frequencies. It is impossible for the end-system user on receiving a
Source Quench to be certain of the circumstances in which it was
issued. This makes the needed end-system response problematic: is
the Source Quench an indication of heavy congestion, approaching
congestion, a burst causing massive overload, or a burst slightly
exceeding reasonable load?
To the extent that gateways drop the last arrived datagram on
overload, Source Quench messages may be distributed unfairly. This
is because the position at the end of the queue may be unfairly often
occupied by the packets of low demand, intermittent users, since
these do not send regular bursts of packets that can preempt most of
the queue space.
[Fin89] developed algorithms for when to issue Source Quench and for
responding to it with a rate-reduction in the IP layer on the sending
Performance and Congestion Control Working Group [Page 9]
RFC 1254 Gateway Congestion Control Survey August 1991
host. The system controls end-to-end performance of connections in a
manner similar to the congestion avoidance portion of Slow-start TCP
[Jac88].
3.2 Random Drop
Random Drop is a gateway congestion control policy intended to give
feedback to users whose traffic congests the gateway by dropping
packets on a statistical basis. The key to this policy is the
hypothesis that a packet randomly selected from all incoming traffic
will belong to a particular user with a probability proportional to
the average rate of transmission of that user. Dropping a randomly
selected packet results in users which generate much traffic having
a greater number of packets dropped compared with those generating
little traffic. The selection of packets to be dropped is completely
uniform. Therefore, a user who generates traffic of an amount below
the "fair share" (as defined in Section 2.3) may also experience a
small amount of packet loss at a congested gateway. This character of
uniformity is in fact a primary goal that Random Drop attempts to
achieve.
The other primary goal that Random Drop attempts to achieve is a
theoretical overhead which is scaled to the number of shared
resources in the gateway rather than the number of its users. If a
gateway congestion algorithm has more computation the more users
there are, this can lead to processing demands that are higher as
congestion increases. Also the low-overhead goal of Random Drop
addresses concerns about the scale of gateway processing that will be
required in the mid-term Internet as gateways with fast processors
and links are shared by very large active sets of users.
3.2.1 For Congestion Recovery
Random Drop has been proposed as an improvement to packet dropping at
the operating point where the gateway's packet buffers overflow.
This is using Random Drop strictly as a congestion recovery
mechanism.
In Random Drop congestion recovery, instead of dropping the last
packet to arrive at the queue, a packet is selected randomly from the
queue. Measurements of an implementation of Random Drop Congestion
Recovery [Man90] showed that a user with a low demand, due to a
longer round-trip time path than other users of the gateway, had a
higher drop rate with RDCR than without. The throughput accorded to
users with the same round-trip time paths was nearly equal with RDCR
as compared to without it. These results suggest that RDCR should be
avoided unless it is used within a scheme that groups traffic more or
less by round-trip time.
Performance and Congestion Control Working Group [Page 10]
RFC 1254 Gateway Congestion Control Survey August 1991
3.2.2 For Congestion Avoidance
Random Drop is also proposed as a congestion avoidance policy
[Jac89]. The intent is to initiate dropping packets when the gateway
is anticipated to become congested and remain so unless there is some
control exercised. This implies selection of incoming packets to
be randomly dropped at a rate derived from identifying the level of
congestion at the gateway. The rate is the number of arrivals
allowed between drops. It depends on the current operating point and
the prediction of congestion.
A part of the policy is to determine that congestion will soon occur
and that the gateway is beginning to operate beyond the knee of the
power curve. With a suitably chosen interval (Section 2.1), the
number of packets from each individual user in a sample over that
interval is proportional to each user's demand on the gateway. Then,
dropping one or more random packets indicates to some user(s) the
need to reduce the level of demand that is driving the gateway beyond
the desired operating point. This is the goal that a policy of
Random Drop for congestion avoidance attempts to achieve.
There are several parameters to be determined for a Random Drop
congestion avoidance policy. The first is an interval, in terms of
number of packet arrivals, over which packets are dropped with
uniform probability. For instance, in a sample implementation, if
this interval spanned 2000 packet arrivals, and a suitable
probability of drop was 0.001, then two random variables would be
drawn in a uniform distribution in the range of 1 to 2,000. The
values drawn would be used by counting to select the packets dropped
at arrival. The second parameter is the value for the probability of
drop. This parameter would be a function of an estimate of the
number of users, their appropriate control intervals, and possibly
the length of time that congestion has persisted. [Jac89] has
suggested successively increasing the probability of drop when
congestion persists over multiple control intervals. The motivation
for increasing the packet drop probability is that the implicit
estimate of the number of users and random selection of their packets
to drop does not guarantee that all users have received enough
signals to decrease demand. Increasing the probability of drop
increases the probability that enough feedback is provided.
Congestion detection is also needed in Random Drop congestion
avoidance, and could be implemented in a variety of ways. The
simplest is a static threshold, but dynamically averaged measures of
demand or utilization are suggested.
The packets dropped in Random Drop congestion avoidance would not be
from the initial inputs to the gateway. We suggest that they would
be selected only from packets destined for the resource which is
Performance and Congestion Control Working Group [Page 11]
RFC 1254 Gateway Congestion Control Survey August 1991
predicted to be approaching congestion. For example, in the case of
a gateway with multiple outbound links, access to each individual
link is treated as a separate resource, the Random Drop policy is
applied at each link independently. Random Drop congestion avoidance
would provide uniform treatment of all cooperating transport users,
even over individual patterns of traffic multiplexed within one
user's stream. There is no aggregation of users.
Simulation studies [Zha89], [Has90] have presented evidence that
Random Drop is not fair across cooperating and non-cooperating
transport users. A transport user whose sending policies included
Go-Back-N retransmissions and did not include Slow-start received an
excessive share of bandwidth from a simple implementation of Random
Drop. The simultaneously active Slow-start users received unfairly
low shares. Considering this, it can be seen that when users do not
respond to control, over a prolonged period, the Random Drop
congestion avoidance mechanism would have an increased probability of
penalizing users with lower demand. Their packets dropped, these
users exert the controls leading to their giving up bandwidth.
Another problem can be seen to arise in Random Drop [She89] across
users whose communication paths are of different lengths. If the
path spans congested resources at multiple gateways, then the user's
probability of receiving an unfair drop and subsequent control (if
cooperating) is exponentially increased. This is a significant
scaling problem.
Unequal paths cause problems for other congestion avoidance policies
as well. Selective Feedback Congestion Indication was devised to
enhance Congestion Indication specifically because of the problem of
unequal paths. In Fair Queueing by source-destination pairs, each
path gets its own queue in all the gateways.
3.3 Congestion Indication
The Congestion Indication policy is often referred to as the DEC Bit
policy. It was developed at DEC [JRC87], originally for the Digital
Network Architecture (DNA). It has also been specified for the
congestion avoidance of the ISO protocols TP4 and CLNP [NIST88].
Like Source Quench, it uses explicit communications from the
congested gateway to the user. However, to use the lowest possible
network resources for indicating congestion, the information is
communicated in a single bit, the Congestion Experienced Bit, set in
the network header of the packets already being forwarded by a
gateway. Based on the condition of this bit, the end-system user
makes an adjustment to the sending window. In the NSP transport
protocol of DECNET, the source makes an adjustment to its window; in
the ISO transport protocol, TP4, the destination makes this
Performance and Congestion Control Working Group [Page 12]
RFC 1254 Gateway Congestion Control Survey August 1991
adjustment in the window offered to the sender.
This policy attempts to avoid congestion by setting the bit whenever
the average queue length over the previous queue regeneration cycle
plus part of the current cycle is one or more. The feedback is
determined independently at each resource.
3.4 Selective Feedback Congestion Indication
The simple Congestion Indication policy works based upon the total
demand on the gateway. The total number of users or the fact that
only a few of the users might be causing congestion is not
considered. For fairness, only those users who are sending more than
their fair share should be asked to reduce their load, while others
could attempt to increase where possible. In Selective Feedback
Congestion Indication, the Congestion Experienced Bit is used to
carry out this goal.
Selective Feedback works by keeping a count of the number of packets
sent by different users since the beginning of the queue averaging
interval. This is equivalent to monitoring their throughputs. Based
on the total throughput, a fair share for each user is determined and
the congestion bit is set, when congestion approaches, for the users
whose demand is higher than their fair share. If the gateway is
operating below the throughput-delay knee, congestion indications are
not set.
A min-max algorithm used to determine the fair share of capacity and
other details of this policy are described in [RJC87]. One parameter
to be computed is the capacity of each resource to be divided among
the users. This metric depends on the distribution of inter-arrival
times and packet sizes of the users. Attempting to determine these
in real time in the gateway is unacceptable. The capacity is instead
estimated from on the throughput seen when the gateway is operating
in congestion, as indicated by the average queue length. In
congestion (above the knee), the service rate of the gateway limits
its throughput. Multiplying the throughput obtained at this
operating point by a capacity factor (between 0.5 and 0.9) to adjust
for the distributions yields an acceptable capacity estimate in
simulations.
Selective Feedback Congestion Indication takes as input a vector of
the number of packets sent by each source-destination pair of end-
systems. Other alternatives include 1) destination address, 2)
input/output link, and 3) transport connection (source/destination
addresses and ports).
These alternatives give different granularities for fairness. In the
Performance and Congestion Control Working Group [Page 13]
RFC 1254 Gateway Congestion Control Survey August 1991
case where paths are the same or round-trip times of users are close
together, throughputs are equalized similarly by basing the selective
feedback on source or destination address. In fact, if the RTTs are
close enough, the simple congestion indication policy would result in
a fair allocation. Counts based on source/destination pairs ensure
that paths with different lengths and network conditions get a fair
throughput at the individual gateways. Counting packets based on
link pairs has a low overhead, but may result in unfairness to users
whose demand is below the fair share and are using a long path.
Counts based on transport layer connection identifiers, if this
information was available to Internet gateways, would make good
distinctions, since the differences of demand of different
applications and instances of applications would be separately
monitored.
Problems with Selective Feedback Congestion Indication include that
the gateway has significant processing to do. With the feasible
choice of aggregation at the gateway, unfairness across users within
the group is likely. For example, an interactive connection
aggregated with one or more bulk transfer connections will receive
congestion indications though its own use of the gateway resources is
very low.
3.5 Fair Queueing
Fair Queueing is the policy of maintaining separate gateway output
queues for individual end-systems by source-destination pair. In the
policy as proposed by [Nag85], the gateway's processing and link
resources are distributed to the end-systems on a round-robin basis.
On congestion, packets are dropped from the longest queue. This
policy leads to equal allocations of resources to each source-
destination pair. A source-destination pair that demands more than a
fair share simply increases its own queueing delay and congestion
drops.
3.5.1 Bit-Round Fair Queueing
An enhancement of Nagle Fair Queueing, the Bit-Round Fair Queuing
algorithm described and simulated by [DKS89] addresses several
shortcomings of Nagle's scheme. It computes the order of service to
packets using their lengths, with a technique that emulates a bit-
by-bit round-robin discipline, so that long packets do not get an
advantage over short ones. Otherwise the round-robin would be
unfair, for example, giving more bandwidth to hosts whose traffic is
mainly long packets than to hosts sourcing short packets.
The aggregation of users of a source-destination pair by Fair
Queueing has the property of grouping the users whose round-trips are
Performance and Congestion Control Working Group [Page 14]
RFC 1254 Gateway Congestion Control Survey August 1991
similar. This may be one reason that the combination of Bit-Round
Fair Queueing with Congestion Indication had particularly good
simulated performance in [DKS89].
The aggregation of users has the expected drawbacks, as well. A
TELNET user in a queue with an FTP user does not get delay benefits;
and host pairs with high volume of connections get treated the same
as a host pair with small number of connections and as a result gets
unfair services.
A problem can be mentioned about Fair Queueing, though it is related
to implementation, and perhaps not properly part of a policy
discussion. This is a concern that the resources (processing) used
for determining where to queue incoming packets would themselves be
subject to congestion, but not to the benefits of the Fair Queuing
discipline. In a situation where the gateway processor was not
adequate to the demands on it, the gateway would need an alternative
policy for congestion control of the queue awaiting processing.
Clever implementation can probably find an efficient way to route
packets to the queues they belong in before other input processing is
done, so that processing resources can be controlled, too. There is
in addition, the concern that bit-by-bit round FQ requires non-FCFS
queueing even within the same source destination pairs to allow for
fairness to different connections between these end systems.
Another potential concern about Fair Queueing is whether it can scale
up to gateways with very large source-destination populations. For
example, the state in a Fair Queueing implementation scales with the
number of active end-to-end paths, which will be high in backbone
gateways.
3.5.2 Stochastic Fairness Queuing
Stochastic Fairness Queueing (SFQ) has been suggested as a technique
[McK90] to address the implementation issues relating to Fair
Queueing. The first overhead that is reduced is that of looking up
the source-destination address pair in an incoming packet and
determining which queue that packet will have to be placed in. SFQ
does not require as many memory accesses as Fair Queueing to place
the packet in the appropriate queue. SFQ is thus claimed to be more
amenable to implementation for high-speed networks [McK90].
SFQ uses a simple hash function to map from the source-destination
address pair to a fixed set of queues. Since the assignment of an
address pair to a queue is probabilistic, there is the likelihood of
multiple address pairs colliding and mapping to the same queue. This
would potentially degrade the additional fairness that is gained with
Fairness Queueing. If two or more address pairs collide, they would
Performance and Congestion Control Working Group [Page 15]
RFC 1254 Gateway Congestion Control Survey August 1991
continue to do so. To deal with the situation when such a collision
occurs, SFQ periodically perturbs the hash function so that these
address pairs will be unlikely to collide subsequently.
The price paid for achieving this implementation efficiency is that
SFQ requires a potentially large number of queues (we must note
however, that these queues may be organized orthogonally from the
buffers in which packets are stored. The buffers themselves may be
drawn from a common pool, and buffers in each queue organized as a
linked list pointed to from each queue header). For example, [McK90]
indicates that to get good, consistent performance, we may need to
have up to 5 to 10 times the number of active source-destination
pairs. In a typical gateway, this may require around 1000 to 2000
queues.
[McK90] reports simulation results with SFQ. The particular hash
function that is studied is using the HDLC's cyclic redundancy check
(CRC). The hash function is perturbed by multiplying each byte by a
sequence number in the range 1 to 255 before applying the CRC. The
metric considered is the standard deviation of the number of packets
output per source-destination pair. A perfectly fair policy would
have a standard deviation of zero and an unfair policy would have a
large standard deviation. In the example studied (which has up to 20
source-destination (s-d) pairs going through a single overloaded
gateway), SFQ with 1280 queues (i.e., 64 times the number of source-
destination pairs), approaches about 3 times the standard deviation
of Fairness Queueing. This must be compared to a FCFS queueing
discipline having a standard deviation which is almost 175 times the
standard deviation of Fairness Queueing.
It is conjectured in [McK90] that a good value for the interval in
between perturbations of the hash function appears to be in the area
between twice the queue flush time of the stochastic fairness queue
and one-tenth the average conversation time between a source-
destination pair.
SFQ also may alleviate the anticipated scaling problems that may be
an issue with Fair Queueing by not strictly requiring the number of
queues equal to the possible source-destination pairs that may be
presenting a load on a particular gateway. However, SFQ achieves this
property by trading off some of the fairness for implementation
efficiency.
[McK90] examines alternative strategies for implementation of Fair
Queueing and SFQ and estimates their complexity on common hardware
platforms (e.g., a Motorola 68020). It is suggested that mapping an
IP address pair may require around 24 instructions per packet for
Fair Queueing in the best case; in contrast SFQ requires 10
Performance and Congestion Control Working Group [Page 16]
RFC 1254 Gateway Congestion Control Survey August 1991
instructions in the worst case. The primary source of this gain is
that SFQ avoids a comparison of the s-d address pair on the packet to
the identity of the queue header. The relative benefit of SFQ over
Fair Queueing is anticipated to be greater when the addresses are
longer.
SFQ offers promising implemenatation benefits. However, the price to
be paid in terms of having a significantly larger number of queues
(and the consequent data structures and their management) than the
number of s-d pairs placing a load on the gateway is a concern. SFQ
is also attractive in that it may be used in concert with the DEC-bit
scheme for Selective Feedback Congestion Indication to provide
fairness as well as congestion avoidance.
4. END-SYSTEM CONGESTION CONTROL POLICIES
Ideally in gateway congestion control, the end-system transport
entities adjust (decrease) their demand in response to control
exerted by the gateway. Schemes have been put in practice for
transport entities to adjust their demand dynamically in response to
congestion feedback. To accomplish this, in general, they call for
the user to gradually increase demand until information is received
that the load on the gateway is too high. In response to this
information, the user decreases load, then begins an exploratory
increases again. This cycle is repeated continuously, with the goal
that the total demand will oscillate around the optimal level.
We have already noted that a Slow-start connection may give up
considerable bandwidth when sharing a gateway with aggressive
transport entities. There is currently no way to enforce that end-
systems use a congestion avoidance policy, though the Host
Requirements RFC [HR89] has defined Slow-start as mandatory for TCP.
This adverse effect on Slow-start connections is mitigated by the
Fair Queueing policy. Our conclusions discuss further the
coexistence of different end-system strategies.
This section briefly presents two fielded transport congestion
control and avoidance schemes, Slow-start and End-System Congestion
Indication, and the responses by means of which they cooperate with
gateway policies. They both use the control paradigm described
above. Slow-start, as mentioned in Section 1, was developed by
[Jac88], and widely fielded in the BSD TCP implementation. End-
system Congestion Indication was developed by [JRC87]. It is fielded
in DEC's Digital Network Architecture, and has been specified as well
for ISO TP4 [NIST88].
Both Slow-start and End-system Congestion Indication view the
relationship between users and gateways as a control system. They
Performance and Congestion Control Working Group [Page 17]
RFC 1254 Gateway Congestion Control Survey August 1991
have feedback and control components, the latter further broken down
into a procedure bringing a new connection to equilibrium, and a
procedure to maintain demand at the proper level. They make use of
policies for increasing and decreasing the transport sender's window
size. These require the sender to follow a set of self-restraining
rules which dynamically adjust the send window whenever congestion is
sensed.
A predecessor of these, CUTE, developed by [Jai86], introduced the
use of retransmission timeouts as congestion feedback. The Slow-
start scheme was also designed to use timeouts in the same way. The
End-System policies for Congestion Indication use the Congestion
Experienced Bit delivered in the network header as the primary
feedback, but retransmission timeouts also provoke an end-system
congestion response.
In reliable transport protocols like TCP and TP4, the retransmission
timer must do its best to satisfy two conflicting goals. On one hand,
the timer must rapidly detect lost packets. And, on the other hand,
the timer must minimize false alarms. Since the retransmit timer is
used as a congestion signal in these end-system policies, it is all
the more important that timeouts reliably correspond to packet drops.
One important rule for retransmission is to avoid bad sampling in the
measurements that will be used in estimating the round-trip delay.
[KP87] describes techniques to ensure accurate sampling. The Host
Requirements RFC [HR89] makes these techniques mandatory for TCP.
The utilization of a resource can be defined as the ratio of its
average arrival rate to its mean service rate. This metric varies
between 0 and 1.0. In a state of congestion, one or more resources
(link, gateway buffer, gateway CPU) has a utilization approaching
1.0. The average delay (round trip time) and its variance approach
infinity. Therefore, as the utilization of a network increases, it
becomes increasingly important to take into account the variance of
the round trip time in estimating it for the retransmission timeout.
The TCP retransmission timer is based on an estimate of the round
trip time. [Jac88] calls the round trip time estimator the single
most important feature of any protocol implementation that expects to
survive a heavy load. The retransmit timeout procedure in RFC-793
[Pos81b] includes a fixed parameter, beta, to account for the
variance in the delay. An estimate of round trip time using the
suggested values for beta is valid for a network which is at most 30%
utilized. Thus, the RFC-793 retransmission timeout estimator will
fail under heavy congestion, causing unnecessary retransmissions that
add to the load, and making congestive loss detection impossible.
Slow-start TCP uses the mean deviation as an estimate of the variance
Performance and Congestion Control Working Group [Page 18]
RFC 1254 Gateway Congestion Control Survey August 1991
to improve the estimate. As a rough view of what happens with the
Slow-start retransmission calculation, delays can change by
approximately two standard deviations without the timer going off in
a false alarm. The same method of estimation may be applicable to
TP4. The procedure is:
Error = Measured - Estimated
Estimated = Estimated + Gain_1 * Error
Deviation = Deviation + Gain_2 * (|Error| - Deviation)
Timeout = Estimated + 2 * Deviation
Where: Gain_1, Gain_2 are gain factors.
4.1 Response to No Policy in Gateway
Since packets must be dropped during congestion because of the finite
buffer space, feedback of congestion to the users exists even when
there is no gateway congestion policy. Dropping the packets is an
attempt to recover from congestion, though it needs to be noted that
congestion collapse is not prevented by packet drops if end-systems
accelerate their sending rate in these conditions. The accurate
detection of congestive loss by all retransmission timers in the
end-systems is extremely important for gateway congestion recovery.
4.2 Response to Source Quench
Given that a Source Quench message has ambiguous meaning, but usually
is issued for congestion recovery, the response of Slow-start to a
Source Quench is to return to the beginning of the cycle of increase.
This is an early response, since the Source Quench on overflow will
also lead to a retransmission timeout later.
4.3 Response to Random Drop
The end-system's view of Random Drop is the same as its view of loss
caused by overflow at the gateway. This is a drawback of the use of
packet drops as congestion feedback for congestion avoidance: the
decrease policy on congestion feedback cannot be made more drastic
for overflows than for the drops the gateway does for congestion
avoidance. Slow-start responds rapidly to gateway feedback. In a
case where the users are cooperating (all Slow-start), a desired
outcome would be that this responsiveness would lead quickly to a
decreased probability of drop. There would be, as an ideal, a self-
adjusting overall amount of control.
Performance and Congestion Control Working Group [Page 19]
RFC 1254 Gateway Congestion Control Survey August 1991
4.4 Response to Congestion Indication
Since the Congestion Indication mechanism attempts to keep gateways
uncongested, cooperating end-system congestion control policies need
not reduce demand as much as with other gateway policies. The
difference between the Slow-start response to packet drops and the
End-System Congestion Indication response to Congestion Experienced
Bits is primarily in the decrease policy. Slow-start decreases the
window to one packet on a retransmission timeout. End-System
Congestion Indication decreases the window by a fraction (for
instance, to 7/8 of the original value), when the Congestion
Experienced Bit is set in half of the packets in a sample spanning
two round-trip times (two windows full).
4.5 Response to Fair Queuing
A Fair Queueing policy may issue control indications, as in the
simulated Bit-Round Fair Queueing with DEC Bit, or it may depend only
on the passive effects of the queueing. When the passive control is
the main effect (perhaps because most users are not responsive to
control indications), the behavior of retransmission timers will be
very important. If retransmitting users send more packets in
response to increases in their delay and drops, Fair Queueing will be
prone to degraded performance, though collapse (zero throughput for
all users) may be prevented for a longer period of time.
5. Conclusions
The impact of users with excessive demand is a driving force as
proposed gateway policies come closer to implementation. Random Drop
and Congestion Indication can be fair only if the transport entities
sharing the gateway are all cooperative and reduce demand as needed.
Within some portions of the Internet, good behavior of end-systems
eventually may be enforced through administration. But for various
reasons, we can expect non-cooperating transports to be a persistent
population in the Internet. Congestion avoidance mechanisms will not
be deployed all at once, even if they are adopted as standards.
Without enforcement, or even with penalties for excessive demand,
some end-systems will never implement congestion avoidance
mechanisms.
Since it is outside the context of any of the gateway policies, we
will mention here a suggestion for a relatively small-scale response
to users which implement especially aggressive policies. This has
been made experimentally by [Jac89]. It would implement a low
priority queue, to which the majority of traffic is not routed. The
candidate traffic to be queued there would be identified by a cache
of recent recipients of whatever control indications the gateway
Performance and Congestion Control Working Group [Page 20]
RFC 1254 Gateway Congestion Control Survey August 1991
policy makes. Remaining in the cache over multiple control intervals
is the criterion for being routed to the low priority queue. In
approaching or established congestion, the bandwidth given to the
low-service queue is decreased.
The goal of end-system cooperation itself has been questioned. As
[She89] points out, it is difficult to define. A TCP implementation
that retransmits before it determines that has been loss indicated
and in a Go-Back-N manner is clearly non-cooperating. On the other
hand, a transport entity with selective acknowledgement may demand
more from the gateways than TCP, even while responding to congestion
in a cooperative way.
Fair Queueing maintains its control of allocations regardless of the
end-system congestion avoidance policies. [Nag85] and [DKS89] argue
that the extra delays and congestion drops that result from
misbehavior could work to enforce good end-system policies. Are the
rewards and penalties less sharply defined when one or more
misbehaving systems cause the whole gateway's performance to be less?
While the tax on all users imposed by the "over-users" is much less
than in gateways with other policies, how can it be made sufficiently
low?
In the sections on Selective Feedback Congestion Indication and Bit-
Round Fair Queueing we have pointed out that more needs to be done on
two particular fronts:
How can increased computational overhead be avoided?
The allocation of resources to source-destination pairs is, in
many scenarios, unfair to individual users. Bit-Round Fair
Queueing offers a potential administrative remedy, but even if it
is applied, how should the unequal allocations be propagated
through multiple gateways?
The first question has been taken up by [McK90].
Since Selective Feedback Congestion Indication (or Congestion
Indication used with Fair Queueing) uses a network bit, its use in
the Internet requires protocol support; the transition of some
portions of the Internet to OSI protocols may make such a change
attractive in the future. The interactions between heterogeneous
congestion control policies in the Internet will need to be explored.
The goals of Random Drop Congestion Avoidance are presented in this
survey, but without any claim that the problems of this policy can be
solved. These goals themselves, of uniform, dynamic treatment of
users (streams/flows), of low overhead, and of good scaling
Performance and Congestion Control Working Group [Page 21]
RFC 1254 Gateway Congestion Control Survey August 1991
characteristics in large and loaded networks, are significant.
Appendix: Congestion and Connection-oriented Subnets
This section presents a recommendation for gateway implementation in
an areas that unavoidably interacts with gateway congestion control,
the impact of connection-oriented subnets, such as those based on
X.25.
The need to manage a connection oriented service (X.25) in order to
transport datagram traffic (IP) can cause problems for gateway
congestion control. Being a pure datagram protocol, IP provides no
information delimiting when a pair of IP entities need to establish a
session between themselves. The solution involves compromise among
delay, cost, and resources. Delay is introduced by call
establishment when a new X.25 SVC (Switched Virtual Circuit) needs to
be established, and also by queueing delays for the physical line.
Cost includes any charges by the X.25 network service provider.
Besides the resources most gateways have (CPU, memory, links), a
maximum supported or permitted number of virtual circuits may be in
contest.
SVCs are established on demand when an IP packet needs to be sent and
there is no SVC established or pending establishment to the
destination IP entity. Optionally, when cost considerations, and
sufficient numbers of unused virtual circuits allow, redundant SVCs
may be established between the same pair of IP entities. Redundant
SVCs can have the effect of improving performance when coping with
large end-to-end delay, small maximum packet sizes and small maximum
packet windows. It is generally preferred though, to negotiate large
packet sizes and packet windows on a single SVC. Redundant SVCs must
especially be discouraged when virtual circuit resources are small
compared with the number of destination IP entities among the active
users of the gateway.
SVCs are closed after some period of inactivity indicates that
communication may have been suspended between the IP entities. This
timeout is significant in the operation of the interface. Setting
the value too low can result in closing of the SVC even though the
traffic has not stopped. This results in potentially large delays
for the packets which reopen the SVC and may also incur charges
associated with SVC calls. Also, clearing of SVCs is, by definition,
nongraceful. If an SVC is closed before the last packets are
acknowledged, there is no guarantee of delivery. Packet losses are
introduced by this destructive close independent of gateway traffic
and congestion.
When a new circuit is needed and all available circuits are currently
Performance and Congestion Control Working Group [Page 22]
RFC 1254 Gateway Congestion Control Survey August 1991
in use, there is a temptation to pick one to close (possibly using
some Least Recently Used criterion). But if connectivity demands are
larger than available circuit resources, this strategy can lead to a
type of thrashing where circuits are constantly being closed and
reopened. In the worst case, a circuit is opened, a single packet
sent and the circuit closed (without a guarantee of packet delivery).
To counteract this, each circuit should be allowed to remain open a
minimum amount of time.
One possible SVC strategy is to dynamically change the time a circuit
will be allowed to remain open based on the number of circuits in
use. Three administratively controlled variables are used, a minimum
time, a maximum time and an adaptation factor in seconds per
available circuit. In this scheme, a circuit is closed after it has
been idle for a time period equal to the minimum plus the adaptation
factor times the number of available circuits, limited by the maximum
time. By administratively adjusting these variables, one has
considerable flexibility in adjusting the SVC utilization to meet the
needs of a specific gateway.
Acknowledgements
This paper is the outcome of discussions in the Performance and
Congestion Control Working Group between April 1988 and July 1989.
Both PCC WG and plenary IETF members gave us helpful reviews of
earlier drafts. Several of the ideas described here were contributed
by the members of the PCC WG. The Appendix was written by Art
Berggreen. We are particularly thankful also to Van Jacobson, Scott
Shenker, Bruce Schofield, Paul McKenney, Matt Mathis, Geof Stone, and
Lixia Zhang for participation and reviews.
References
[DKS89] Demers, A., Keshav, S., and S. Shenker, "Analysis and
Simulation of a Fair Queueing Algorithm", Proceedings of SIGCOMM '89.
[Fin89] Finn, G., "A Connectionless Congestion Control Algorithm",
Computer Communications Review, Vol. 19, No. 5, October 1989.
[Gar87] Gardner, M., "BBN Report on the ARPANET", Proceedings of the
McLean IETF, SRI-NIC IETF-87/3P, July 1987.
[GREQ87] Braden R., and J. Postel, "Requirements for Internet
Gateways", RFC 1009, USC/Information Sciences Institute, June 1987.
[HREQ89] Braden R., Editor, "Requirements for Internet Hosts --
Communications Layers", RFC 1122, Internet Engineering Task Force,
October 1989.
Performance and Congestion Control Working Group [Page 23]
RFC 1254 Gateway Congestion Control Survey August 1991
[Has90] Hashem, E., "Random Drop Congestion Control", M.S. Thesis,
Massachusetts Institute of Technology, Department of Computer
Science, 1990.
[Jac88] Jacobson, V., "Congestion Avoidance and Control", Proceedings
of SIGCOMM '88.
[Jac89] Jacobson, V., "Presentations to the IETF Performance and
Congestion Control Working Group".
[Jaf81] Jaffe, J., "Bottleneck Flow Control", IEEE Transactions on
Communications, COM-29(7), July, 1981.
[Jai86] Jain, R., "A Timeout-based Congestion Control Scheme for
Window Flow-controlled Networks", IEEE Journal on Selected Areas in
Communications, SAC-4(7), October 1986.
[JRC87] Jain, R., Ramakrishnan, K., and D. Chiu, "Congestion
Avoidance in Computer Networks With a Connectionless Network Layer",
Technical Report DEC-TR-506, Digital Equipment Corporation.
[Kle79] Kleinrock, L., "Power and Deterministic Rules of Thumb for
Probabilistic Problems in Computer Communications", Proceedings of
the ICC '79.
[KP87] Karn, P., and C. Partridge, "Improving Round Trip Estimates in
Reliable Transport Protocols", Proceedings of SIGCOMM '87.
[Man90] Mankin, A., "Random Drop Congestion Control", Proceedings of
SIGCOMM '90.
[McK90] McKenney, P., "Stochastic Fairness Queueing", Proceedings of
INFOCOM '90.
[Nag84] Nagle, J., "Congestion Control in IP/TCP Internetworks", RFC
896, FACC Palo Alto, 6 January 1984.
[Nag85] Nagle, J., "On Packet Switches With Infinite Storage", RFC
970, FACC Palo Alto, December 1985.
[NIST88] NIST, "Stable Implementation Agreements for OSI Protocols,
Version 2, Edition 1", National Institute of Standards and Technology
Special Publication 500-162, December 1988.
[Pos81a] Postel, J., "Internet Control Message Protocol - DARPA
Internet Program Protocol Specification", RFC-792, USC/Information
Sciences Institute, September 1981.
Performance and Congestion Control Working Group [Page 24]
RFC 1254 Gateway Congestion Control Survey August 1991
[Pos81b] Postel, J., "Transmission Control Protocol - DARPA Internet
Program Protocol Specification", RFC-793, DARPA, September 1981.
[RJC87] Ramakrishnan, K., Jain, R., and D. Chiu, "A Selective Binary
Feedback Scheme for General Topologies", Technical Report DEC-TR-510,
Digital Equipment Corporation.
[She89] Shenker, S., "Correspondence with the IETF Performance and
Congestion Control Working Group".
[Sp89] Spector, A., and M. Kazar, "Uniting File Systems", Unix
Review, Vol. 7, No. 3, March 1989.
[Zha89] Zhang, L., "A New Architecture for Packet Switching Network
Protocols", Ph.D Thesis, Massachusetts Institute of Technology,
Department of Computer Science, 1989.
Security Considerations
Security issues are not discussed in this memo.
Authors' Addresses
Allison Mankin
The MITRE Corporation
M/S W425
7525 Colshire Drive
McLean, VA 22102
Email: mankin@gateway.mitre.org
K.K. Ramakrishnan
Digital Equipment Corporation
M/S LKG1-2/A19
550 King Street
Littleton, MA 01754
Email: rama@kalvi.enet.dec.com
Performance and Congestion Control Working Group [Page 25]