Network Working Group S. Jackowski
Request for Comments: 2688 Deterministic Networks
Category: Standards Track D. Putzolu
Intel Architecture Labs
E. Crawley
Argon Networks
B. Davie
Cisco Systems
September 1999

Integrated Services Mappings for Low Speed Networks

Status of this Memo

This document specifies an Internet standards track protocol for the
Internet community, and requests discussion and suggestions for
improvements. Please refer to the current edition of the "Internet
Official Protocol Standards" (STD 1) for the standardization state
and status of this protocol. Distribution of this memo is unlimited.

Copyright Notice

Copyright (C) The Internet Society (1999). All Rights Reserved.

Abstract

A set of companion documents describe an architecture for providing
integrated services over low-bitrate links, sUCh as modem lines, ISDN
B-channels, and sub-T1 links [1, 2, 3, 4]. The main components of the
architecture are: a set of real-time encapsulation formats for
asynchronous and synchronous low-bitrate links, a header compression
architecture optimized for real-time flows, elements of negotiation
protocols used between routers (or between hosts and routers), and
announcement protocols used by applications to allow this negotiation
to take place.

This document defines the service mappings of the IETF Integrated
Services for low-bitrate links, specifically the controlled load [5]
and guaranteed [6] services. The approach takes the form of a set of
guidelines and considerations for implementing these services, along
with evaluation criteria for elements providing these services.

1. Introduction

In addition to the "best-effort" services the Internet is well-known
for, other types of services ("integrated services") are being
developed and deployed in the Internet. These services support
special handling of traffic based on bandwidth, latency, and other
requirements that cannot usually be met using "best-effort" service.

This document defines the mapping of integrated services "controlled
load" [5] and "guaranteed" [6] services on to low-bandwidth links.
The architecture and mechanisms used to implement these services on
such links are defined in a set of companion documents. The
mechanisms defined in these documents include both compression of
flows (for bandwidth savings) [4,10] and a set of extensions to the
PPP protocol which permit fragmentation [2] or suspension [3] of
large packets in favor of packets from flows with more stringent
service requirements.

1.1. Specification Language

The key Words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
"SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this
document are to be interpreted as described in RFC2119 [11].

2. Issues for Providing Controlled and Guaranteed Service

Unlike other link layers, the links referred to in this document
operate only over low speed point to point connections. Examples of
the kinds of links addressed here include dial-up lines, ISDN
channels, and low-speed (1.5Mbps or less) leased lines. Such links
can occur at different positions within the end-to-end path:

- host to directly connected host.
- host to/from network Access device (router or switch).
- Edge device (subnet router or switch) to/from router or switch.
- In rare circumstances, a link from backbone router to backbone
router.

These links often represent the first or last wide area hop in a true
end to end service. Note that these links may be the most bandwidth
constrained along the path between two hosts.

The services utilized in mapping integrated services to these links
are only provided if both endpoints on the link support the
architecture and mechanisms referenced above. Support for these
mechanisms is determined during the PPP negotiation. The non-shared

nature of these links, along with the fact that point-to-point links
are typically dual simplex (i.e., the send and receive channels are
separate) allows all admission control decisions to be made locally.

As described in [2] and [3], for systems that can exert real time
control of their transmission at a finer grain than entire HDLC
frames, the suspend/resume approach optimizes the available bandwidth
by minimizing header overhead associated with MLPPP pre-fragmentation
and can provide better delay. However, this comes at the eXPense of
preparing all outgoing data and scanning all incoming data for
suspend/resume control information. The fragmentation approach can
be implemented without additional scanning of the data stream (beyond
bit-/byte-stuffing, which may be in hardware) and is applicable to
systems which provide only frame-oriented transmission control.
Choice of suspend/resume versus fragmentation should be made based on
the level of transmission control, the element's capability to handle
the HDLC-like framing described in [2], and the system overhead
associated with byte by byte scanning (required by suspend/resume).

To provide controlled load or guaranteed service with the
suspend/resume approach, when a packet for an admitted flow (QoS
packet) arrives during transmission of a best effort packet and
continued transmission of the best effort packet would violate delay
constraints of the QoS service flows, the best effort packet is
preempted, the QoS packet/fragments are added to the transmission,
and the best effort packet transmission is then resumed: usually all
in one transmission. The receiving station separates the best effort
packet from the embedded QoS packet's fragments. It is also
conceivable that one QoS flow's packet might suspend another flow's
packet if the delivery deadline of the new packet is earlier than the
current packet.

For systems which use fragmentation, any packets longer than the
maximum tolerable delay for packets from enhanced service flows are
fragmented prior to transmission so that a short packet for another
flow can be interleaved between fragments of a larger packet and
still meet the transmission deadline for the flow requiring enhanced
services.

Note that the fragmentation discussed in this document refers to
multilink PPP (MLPPP) fragmentation and associated MCMLPPP
modifications as described in [2], not IP or other layer 3
fragmentation. MLPPP fragmentation is local to the PPP link, and
does not affect end-to-end (IP) MTU.

2.1 Calculating "Acceptable Delay" for Int-serv flows

A router which provides Controlled Load or Guaranteed Service over a
low speed serial link needs to have some notion of the "acceptable
delay" for packets that belong to int-serv flows. If using
fragmentation, a router needs to know what size to fragment packets
to; if using suspend/resume, it needs to know when it is appropriate
to suspend one packet to meet the delay goals of another.

Unfortunately, there is no hard and fast way for a single delay bound
to be determined for a particular flow; while the end-points of a
flow have enough information to determine acceptable end-to-end delay
bounds and to make reservation requests of the network to meet those
bounds, they do not communicate a "per-hop" delay to routers.

In the case of Guaranteed Service [6], one approach is to let the
network operator configure parameters on the router that will
directly affect its delay performance. We observe that guaranteed
service allows routers to deviate from the ideal fluid flow model and
to advertise the extent of the deviation using two error terms C and
D, the rate-dependent and rate-independent error terms, defined in
[6]. A network operator can configure parameters of the low speed
link in such a way that D is set to a value of her choice.

If link-level fragmentation is used, the router controlling a low-
speed link can be configured with a certain fragment size. This will
enable a component of the error term D to be calculated based on the
time to send one fragment over the link. (Note that D may have other
components such as the speed of light delay over the link.) Details
of the calculation of D are described below. Similarly, if
suspend/resume is used, the router may be configured with a delay
parameter, which would enable it to decide when it was appropriate to
suspend a packet.

For Controlled Load, there are no error terms, and the router must
decide how best to meet the requirements of the admitted reservations
using only the information in their TSpecs. Since the definition of
Controlled Load states that a CL flow with Tspec rate r should
receive treatment similar to an unloaded network of capacity r, CL
packets should not generally experience end-to-end delays
significantly greater than b/r + propagation delays. Clearly a router
connected to a low speed link should not introduce a delay greater
than b/r due to transmission of other fragments; ideally it should
introduce substantially less delay than b/r, since other hops on the
end-to-end path may introduce delay as well. However, this may be
difficult for flows with very small values of b.

It is expected that implementers will make their own tradeoffs as to
how low to make the delay for Controlled Load flows. Similarly, it
may not be possible or desirable to configure the parameters
affecting D to arbitrarily small values, since there is a cost in
overhead in fragmenting packets to very small sizes. Conversely, if D
is too large, some applications may find that they cannot make a
reservation that will meet their delay objectives.

For the remainder of this document, we assume that a router has some
notion of the acceptable delay that it may introduce before beginning
transmission of a packet. This delay is in addition to any delay that
a packet might be subjected to as a result of the "ideal" queuing
algorithm that the router uses to schedule packets.

3. Controlled Load and Guaranteed Service Class Mapping

Supporting integrated services over PPP links which implement MCML or
RTF can be accomplished in several ways. Guidelines for mapping
these services to PPP links and to the classes provided by the
suspend/resume and fragmentation mechanisms are presented below.
Note that these guidelines assume that some sort of signaling
protocol is used to indicate desired quality of service to both the
sender and receiver of a flow over a PPP link.

3.1 Predefined Class Mappings

A relatively simple method of class mapping that MAY be used is one
where class values correspond to predefined levels of service. In
this arrangement, all admitted flows are grouped into one of several
buckets, where each bucket roughly corresponds to the level of
service desired for the flows placed in it. An example set of
mappings appears below:

MCML Short MCML Long RTF Service

0b00 0b0000 0b000 Best Effort
NA 0b0001 0b001 Reserved
0b01 0b0010 0b010 Delay Sensitive, no bound
NA 0b0011 0b011 Reserved
NA 0b0100 0b100 Reserved
0b10 0b0101 0b101 Delay Sensitive, 500ms bound
NA 0b0110 0b110 Delay Sensitive, 250ms bound
0b11 0b0111 0b111 Network Control

Table 1: Example Mappings of Classes to Services

Note that MCML has two formats, short sequence numbers, and long
sequence numbers, that allow for 2 and 4 bits of class identification.
RTF allows for 3 bits of class identification in all formats.

Using a default-mapping method of assigning classes to flows in a
fixed fashion comes with certain limitations. In particular, all flows
which fall within a particular bucket (are assigned to a particular
class) will be scheduled against each other at the granularity of
packets, rather than at the finer grained level of fragments. This
can result in overly conservative admission control when the number of
available classes is small such as in MCML short sequence number
format.

3.2 Predefined Class Mappings and Prefix Elision

In the case where fewer reservations are expected than the total
number of classes negotiated for a PPP link, it is possible to assign
individual flows to fixed class numbers. This assignment is useful in
the case where the protocol identifier associated with one or more
flows is known at LCP negotiation time and the bandwidth of the
connection is relatively small. If these conditions hold true, then
for those flows that are known, a specific class can optionally be
assigned to them and the prefix elision PPP option [2] can be used for
those classes to achieve a small bandwidth savings.

3.3 Dynamic Class Mappings

In the case where predefined class mappings are not satisfactory, an
implementer MAY map class values to individual packets rather than
assigning flows to fixed classes. This can be done due to the fact
that the classes that MCML and RTF provide can be viewed purely as
PPP-specific segmentation/fragmentation mechanisms. That is, while the
class number MUST remain constant on an intra-packet basis, it MAY
vary on an inter-packet basis for all flows transiting a PPP
link. Actual assignment of particular flows to fixed classes is
unnecessary, as the class numbers are NOT REQUIRED to have any meaning
other than in the context of identifying the membership of
fragments/segments as part of a single packet. This point is
sufficiently important that an example is provided below.

Consider a PPP link using the MCML short sequence number fragment
format (that is, four classes are provided). Assume that in addition
to carrying best effort traffic, this link is carrying five guaranteed
service flows, A, B, C, D, and E. Further assume that the link
capacity is 100kbit/s and the latency is 100ms. Finally, assume the BE
traffic is sufficient to keep the pipe full at all times and that GS
flows A-E are each 10kbit/s and all have delay bounds of 145ms.

Time(ms) Action
0 BE traffic is queued up
0 2kbit fragment from 10kbit packet of BE traffic sent, cls 0 (...)
8 2kbit fragment from BE sent, cls 0 (10kbit BE packet done)
9 8kbit packet from flow A arrives
10 2kbit fragment from A sent, cls 1 (8kbit flow A packet start)
11 8kbit packet from flow B arrives
12 2kbit fragment from B sent, cls 2 (8kbit flow B packet start)
13 8kbit packets from flows C, D, and E arrive
14 2kbit fragment from C sent, cls 3 (8kbit flow C packet start)
16 2kbit fragment from D sent, cls 0 (8kbit flow D packet start)
18 2kbit fragment from A sent, cls 1
20 2kbit fragment from B sent, cls 2
22 2kbit fragment from A sent, cls 1
24 2kbit fragment from A sent, cls 1 (8kbit flow A packet done)
26 2kbit fragment from E sent, cls 1 (8kbit flow E packet start)
27 8kbit packet from flow A arrives
28 2kbit fragment from B sent, cls 2
30 2kbit fragment from C sent, cls 3
32 2kbit fragment from E sent, cls 1
34 2kbit fragment from B sent, cls 2 (8kbit flow B packet done)
36 2kbit fragment from E sent, cls 1
38 2kbit fragment flow A sent, cls 2 (8kbit flow A packet start)
(etc.)

This example shows several things. First, multiple flows MAY share
the same class, particularly in the case where there are more flows
than classes. More importantly, there is no reason that a particular
flow must be assigned to a fixed class - the only requirement is that
each packet, when fragmented, MUST have the same class value assigned
to all fragments. Beyond this requirement the link scheduler may
assign individual to changing class numbers as necessary to meet
reservation requirements.

One suggestion to implementers of integrated services on MCML and RTF
links using dynamic mappings is that all BE traffic SHOULD be
logically separated from QoS traffic, and mapped to a fragmentable
(MCML classes 0-3 in short sequence number fragment format, 0-15 in
long sequence number fragment format) or suspendable (RTF classes 0-
6) class. Since BE traffic will in most implementations not be
scheduled for transmission except when a link is empty (that is, no
CL or GS traffic is ready for transmission), implementers MAY choose
to make use of class number 0 for BE traffic.

3.4 Non-Conformant Traffic

Treatment of non-conformant QoS traffic is largely determined by the
appropriate service specifications, but the detailed implementation
in the context of this draft allows for some flexibility. Policing
of flows containing non-conformant traffic SHOULD always be done at
the level of granularity of individual packets rather than at a finer
grained level. In particular, in those cases where a network element
scheduling flows for transmission needs to drop non-conformant
traffic, it SHOULD drop entire packets rather than dropping
individual fragments of packets belonging to non-conformant traffic.
In those cases where a network element forwards non-conformant
traffic when link bandwidth is available rather than dropping the
traffic, the implementation SHOULD fragment packets of such traffic
as if it were best effort traffic.

Whether BE and non-conformant traffic are treated differently in
regards to transmission (e.g., BE is given priority access over non-
conformant traffic to the link) or whether within each type of
traffic special treatment is afforded to individual flows (e.g., WFQ,
RED, etc.) is service dependent.

4. Guidelines for Implementers

4.1. PPP Bit and Byte Stuffing Effects on Admission Control

An important consideration in performing admission control for PPP
links is reductions in effective link rate due to bit stuffing.
Typical bit stuffing algorithms can result in as much as 20%
additional overhead. Thus, admission control implementations for
guaranteed service over links where bit stuffing is used SHOULD take
the RSpec rate of all flows and multiply by 1.2, to account for the
20% overhead from bit stuffing, when determining whether a new flow
can be admitted or not. Admission control implementations for
controlled load reservations may use a similar algorithm using the
TSpec peak rate or may attempt to measure the actual degree of
expansion occurring on a link due to bit stuffing. This
characterization can then be used to adjust the calculated remaining
link capacity. Such measurements must be used cautiously, in that the
degree of bit stuffing that occurs may vary significantly, both in an
inter- and intra-flow fashion.

Byte stuffing is also used on many PPP links, most frequently on POTS
modems when using the v.42 protocol. Byte stuffing poses a difficult
problem to admission control, particularly in the case of guaranteed
service, due to its highly variable nature. In the worse case, byte
stuffing can result in a doubling of frame sizes. As a consequence, a
strict implementation of admission control for guaranteed load on

byte stuffed PPP links SHOULD double the RSpec of link traffic in
making flow admission decisions. As with bit stuffing,
implementations of controlled load service admission control
algorithms for links with byte stuffing MAY attempt to determine
average packet expansion via observation or MAY use the theoretical
worst case values.

4.2. Compression Considerations

The architecture for providing integrated services over low bandwidth
links uses several PPP options to negotiate link configuration as
described in [4, 8, 10]. When deciding whether to admit a flow,
admission control MUST compute the impact of the following on MTU
size, rate, and fragment size:

Header compression: Van Jacobson or Casner-Jacobson [4,8,10].
Prefix Elision.
CCP.
Fragment header option used.
Fragmentation versus suspend/resume approach.

If any of the compression options are implemented for the connection,
the actual transmission rate, and thus the bandwidth required of the
link, will be reduced by the compression method(s) used.

Prefix elision can take advantage of mapping flows to MLPPP classes
to elide prefixes which cannot be compressed at higher layers. By
establishing agreement across the link, the sender may elide a prefix
for a certain class of traffic and upon receiving packets in that
class, the receiver can restore the prefix.

Both compression gain and elision gain MUST be included as described
in the admission control section below. Note that the ability to
perform compression at higher layers (e.g. TCP or RTP/UDP) may depend
on the provision of a hint by the sender, as described in [9].

4.3. Admission Control

Admission control MUST decide whether to admit a flow based on rate
and delay. Assume the following:

LinkRate is the rate of the link.
MTU is the maximum transmission unit from a protocol.
MRU is the maximum receive unit for a particular link.
CMTU is the maximum size of the MTU after compression is applied.
eMTU is the effective size at the link layer of an MTU-sized packet
after link layer fragmentation and addition of the fragment headers.
FRAG is the fragment size including MLPPP header/trailers.

Header is the size of the header/trailers/framing for MLPPP/Fragments.
pHeader is the additional header/framing overhead associated with
suspend/resume. This should include FSE and worst case stuffing
overhead.
pDelay is the time take to suspend a packet already "in flight",
e.g. due to the delay to empty the output FIFO.
b is the bucket depth in bytes
R is the requested Rate.
Dlink is the fixed overhead delay for the link (Modem, DSU,
speed-of-light, etc).
eRate is the effective rate after compression and fragmentation.

The Dlink term MAY be configured by an administrative tool once the
network is installed; it may be determined by real-time measurement
means; or it MAY be available from hardware during link setup and/or
PPP negotiation. Refer to Appendix A for more considerations on PPP
link characteristics and delays.

Admission control MUST compute CMTU, eMTU, and eRate for Controlled
Load Service, and it MUST compute CMTU, eMTU, eRate, and D for
Guaranteed Service:

To determine whether the requested rate is available, Admission
Control MUST compute the effective rate of the request (eRate) -
worst case - as follows:

#_of_Fragments = CMTU div (FRAG-Header) [Integer divide]
Last_Frag_Size = CMTU mod (FRAG-Header

If Last_Frag_Size != 0
eMTU = (#_of_Fragments) * FRAG + Last_Frag_Size + Header
Else
eMTU = (#_of_Fragments) * FRAG

eRate = eMTU/CMTU * R [floating point divide]

Admission control SHOULD compare the eRate of the request against the
remaining bandwidth available to determine if the requested rate can
be delivered.

For Controlled Load Service, a flow can be admitted as long as there
is sufficient bandwidth available (after the above computation) to
meet the rate requirement, and if there is sufficient buffer space
(sum of the token bucket sizes does not exceed the buffer capacity).
While some statistical multiplexing could be done in computing
admissibility, the nature of the low-bitrate links could make this
approach risky as any delay incurred to address a temporary
overcommitment could be difficult to amortize.

4.4 Error Term Calculations

Guaranteed Service requires the calculation of C and D error terms. C
is a rate-dependent error term and there are no special
considerations affecting its calculation in the low-speed link
environment. The D term is calculated from the inherent link delay
(Dlink) plus the potential worst case delay due to transmission of
another fragment or suspend/resume overhead. Thus, D should be
calculated as

D = Dlink + FRAG/LinkRate

in the case of a fragementing implementation and

D = Dlink + pHeader + pDelay

for a suspend/resume implementation.

4.5 Scheduling Considerations

We may think of the link scheduler as having two parts, the first of
which schedules packets for transmission before passing them to the
second part of the scheduler -- the link level scheduler -- which is
responsible for fragmenting packets, mapping them to classes, and
scheduling among the classes.

In the dynamic class mapping mode of Section 3.3, when deciding which
class to assign a packet to, the link level scheduler should take
account of the sizes of other packets currently assigned to the same
class. In particular, packets with the tightest delay constraints
should not be assigned to classes for which relatively large packets
are in the process of being transmitted.

In either the dynamic or the static class mapping approach, note that
the link-level scheduler SHOULD control how much link bandwidth is
assigned to each class at any instant. The scheduler should assign
bandwidth to a class according to the bandwidth reserved for the sum
of all flows which currently have packets assigned to the class. Note
that in the example of Section 3.3, when packets from flows A and E
were assigned to the same class (class 1), the scheduler assigned
more bandwidth to class 1, reflecting the fact that it was carrying
traffic from reservations totaling 20kbit/s while the other classes
were carrying only 10kbit/s.

5. Security Considerations

General security considerations for MLPPP and PPP links are addressed
in RFC1990 [12] and RFC1661 [13], respectively. Security
considerations relevant to RSVP, used as the signaling protocol for
integrated services, are discussed in RFC2209 [14].

A specific security consideration relevant to providing quality of
service over PPP links appears when relying on either observed or
theoretical average packet expansion during admission control due to
bit- or byte-stuffing. Implementations based on these packet-
expansion values contain a potential vulnerability to denial of
service attacks. An adversary could intentionally send traffic that
will result in worst case bit- or byte stuffing packet expansion.
This in turn could result in quality of service guarantees not being
met for other flows due to overly permissive admission control. This
potential denial of service attack argues strongly for using a worst
case expansion factor in admission control calculations, even for
controlled load service.

Beyond the considerations documented above, this document introduces
no new security issues on top of those discussed in the companion
ISSLL documents [1], [2] and [3] and AVT document [4]. Any use of
these service mappings assumes that all requests for service are
authenticated appropriately.

6. References

[1] Bormann, C., "Providing Integrated Services over Low-bitrate
Links", RFC2689, September 1999.

[2] Bormann, C., "The Multi-Class Extension to Multi-Link PPP", RFC
2686, September 1999.

[3] Bormann, C., "PPP in a Real-time Oriented HDLC-like Framing",
RFC2687, September 1999.

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

[5] Wroclawski, J., "Specification of the Controlled-Load Network
Element Service", RFC2211, September 1997.

[6] Partridge, C. and R. Guerin, "Specification of Guaranteed
Quality of Service", RFC2212, September 1997.

[7] Shenker, S. and J. Wroclawski, "General Characterization
Parameters for Integrated Service Network Elements", RFC2215,
September 1997.

[8] Jacobson, V., "TCP/IP Compression for Low-Speed Serial Links",
RFC1144, February 1990.

[9] B. Davie et al. "Integrated Services in the Presence of
Compressible Flows", Work in Progress (draft-davie-intserv-
compress-00.txt), Feb. 1999.

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

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

[12] Sklower, K., Lloyd, B., McGregor, G., Carr, D. and T.
Coradettim, "The PPP Multilink Protocol (MP)", RFC1990, August
1996.

[13] Simpson, W., Editor, "The Point-to-Point Protocol (PPP)", STD
51, RFC1661, July 1994.

[14] Braden, R. and L. Zhang, "Resource ReSerVation Protocol (RSVP)
-- Version 1 Message Processing Rules", RFC2209, September
1997.

7. Authors' Addresses

Steve Jackowski
Deterministic Networks, Inc.
245M Mt Hermon Rd, #140
Scotts Valley, CA 95060
USA

Phone: +1 (408) 813 6294
EMail: stevej@DeterministicNetworks.com

David Putzolu
Intel Architecture Labs (IAL)
JF3-206-H10
2111 NE 25th Avenue
Hillsboro, OR 97124-5961
USA

Phone: +1 (503) 264 4510
EMail: David.Putzolu@intel.com

Eric S. Crawley
Argon Networks, Inc.
25 Porter Road
Littleton, MA 01460
USA

Phone: +1 (978) 486-0665
EMail: esc@argon.com

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

Phone: +1 (978) 244 8921
EMail: bdavie@cisco.com

Acknowledgements

This document draws heavily on the work of the ISSLL WG of the IETF.

Appendix A. Admission Control Considerations for POTS Modems

The protocols used in current implementations of POTS modems can
exhibit significant changes in link rate and delay over the duration
of a connection. Admission control and link scheduling algorithms
used with these devices MUST be prepared to compensate for this
variability in order to provide a robust implementation of integrated
services.

Link rate on POTS modems is typically reported at connection time.
This value may change over the duration of the connection. The v.34
protocol, used in most POTS modems, is adaptive to link conditions,
and is able to recalibrate transmission rate multiple times over the
duration of a connection. Typically this will result in a small
(~10%) increase in transmission rate over the initial connection
within the first minute of a call. It is important to note, however,
that other results are possible as well, including decreases in
available bandwidth. Admission control algorithms MUST take such
changes into consideration as they occur, and implementations MUST be
able to gracefully handle the pathological case where link rate
actually drops below the currently reserved capacity of a link.

Delay experienced by traffic over POTS modems can vary significantly
over time. Unlike link rate, the delay often does not converge to a
stable value. The v.42 protocol is used in most POTS modems to
provide link-layer reliability. This reliability, which is
implemented via retransmission, can cause frames to experience
significant delays. Retransmissions also implicitly steal link
bandwidth from other traffic. These delays and reductions in link
bandwidth make it extremely difficult to honor a guaranteed service
reservation. On a link that is actually lightly or moderately loaded,
a controlled load service can to some extent accept such events as
part of the behavior of a lightly loaded link. Unfortunately, as
actual link utilization increases, v.42 retransmissions have the
potential of stealing larger and larger fractions of available link
bandwidth; making even controlled load service difficult to offer at
high link utilization when retransmissions occur.

9. Full Copyright Statement

Copyright (C) The Internet Society (1999). All Rights Reserved.

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

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

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

Acknowledgement

Funding for the RFCEditor function is currently provided by the
Internet Society.