Network Working Group C. Bormann
Request for Comments: 2689 Universitaet Bremen TZI
Category: Informational September 1999

Providing Integrated Services over Low-bitrate Links

Status of this Memo

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

Copyright Notice

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

Abstract

This document describes an architecture for providing integrated
services over low-bitrate links, sUCh as modem lines, ISDN B-
channels, and sub-T1 links. It covers only the lower parts of the
Internet Multimedia Conferencing Architecture [1]; additional
components required for application services such as Internet
Telephony (e.g., a session initiation protocol) are outside the scope
of this document. The main components of the architecture are: a
real-time encapsulation format 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.

1. Introduction

As an extension to the "best-effort" services the Internet is well-
known for, additional types of services ("integrated services") that
support the transport of real-time multimedia information are being
developed for, and deployed in the Internet. Important elements of
this development are:

- parameters for forwarding mechanisms that are appropriate for
real-time information [11, 12],

- a setup protocol that allows establishing special forwarding
treatment for real-time information flows (RSVP [4]),

- a transport protocol for real-time information (RTP/RTCP [6]).

In addition to these elements at the network and transport levels of
the Internet Multimedia Conferencing Architecture [1], further
components are required to define application services such as
Internet Telephony, e.g., protocols for session initiation and
control. These components are outside the scope of this document.

Up to now, the newly developed services could not (or only very
inefficiently) be used over forwarding paths that include low-bitrate
links such as 14.4, 33.6, and 56 kbit/s modems, 56 and 64 kbit/s ISDN
B-channels, or even sub-T1 links. The encapsulation formats used on
these links are not appropriate for the simultaneous transport of
arbitrary data and real-time information that has to meet stringent
delay requirements. Transmission of a 1500 byte packet on a 28.8
kbit/s modem link makes this link unavailable for the transmission of
real-time information for about 400 ms. This adds a worst-case delay
that causes real-time applications to operate with round-trip delays
on the order of at least a second -- unacceptable for real-time
conversation. In addition, the header overhead associated with the
protocol stacks used is prohibitive on low-bitrate links, where
compression down to a few dozen bytes per real-time information
packet is often desirable. E.g., the overhead of at least 44
(4+20+8+12) bytes for HDLC/PPP, IP, UDP, and RTP completely
overshadows typical audio payloads such as the 19.75 bytes needed for
a G.723.1 ACELP audio frame -- a 14.4 kbit/s link is completely
consumed by this header overhead alone at 40 real-time frames per
second total (i.e., at 25 ms packetization delay for one stream or 50
ms for two streams, with no space left for data, yet). While the
header overhead can be reduced by combining several real-time
information frames into one packet, this increases the delay incurred
while filling that packet and further detracts from the goal of
real-time transfer of multi-media information over the Internet.

This document describes an approach for addressing these problems.
The main components of the architecture are:

- a real-time encapsulation format 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.

2. Design Considerations

The main design goal for an architecture that addresses real-time
multimedia flows over low-bitrate links is that of minimizing the
end-to-end delay. More specifically, the worst case delay (after
removing possible outliers, which are equivalent to packet losses
from an application point of view) is what determines the playout
points selected by the applications and thus the delay actually
perceived by the user.

In addition, any such architecture should obviously undertake every
attempt to maximize the bandwidth actually available to media data;
overheads must be minimized.

An important component of the integrated services architecture is the
provision of reservations for real-time flows. One of the problems
that systems on low-bitrate links (routers or hosts) face when
performing admission control for such reservations is that they must
translate the bandwidth requested in the reservation to the one
actually consumed on the link. Methods such as data compression
and/or header compression can reduce the requirements on the link,
but admission control can only make use of the reduced requirements
in its calculations if it has enough information about the data
stream to know how effective the compression will be. One goal of
the architecture therefore is to provide the integrated services
admission control with this information. A beneficial side effect
may be to allow the systems to perform better compression than would
be possible without this information. This may make it worthwhile to
provide this information even when it is not intended to make a
reservation for a real-time flow.

3. The Need for a Concerted Approach

Many technical approaches come to mind for addressing these problems,
in particular a new form of low-delay encapsulation to address delay
and header compression methods to address overhead. This section
shows that these techniques should be combined to solve the problem.

3.1. Real-Time Encapsulation

The purpose of defining a real-time link-layer encapsulation protocol
is to be able to introduce newly arrived real-time packets into the
link-layer data stream without having to wait for the currently
transmitted (possibly large) packet to end. Obviously, a real-time
encapsulation must be part of any complete solution as the problem of
delays induced by large frames on the link can only be solved on this
layer.

To be able to switch to a real-time packet quickly in an interface
driver, it is first necessary to identify packets that belong to
real-time flows. This can be done using a heuristic approach (e.g.,
favor the transmission of highly periodic flows of small packets
transported in IP/UDP, or use the IP precedence fields in a specific
way defined within an organization). Preferably, one also could make
use of a protocol defined for identifying flows that require special
treatment, i.e. RSVP. Of the two service types defined for use with
RSVP now, the guaranteed service will only be available in certain
environments; for this and various other reasons, the service type
chosen for many adaptive audio/video applications will most likely be
the controlled-load service. Controlled-load does not provide
control parameters for target delay; thus it does not unambiguously
identify those packet streams that would benefit most from being
transported in a real-time encapsulation format. This calls for a
way to provide additional parameters in integrated services flow
setup protocols to control the real-time encapsulation.

Real-time encapsulation is not sufficient on its own, however: Even
if the relevant flows can be appropriately identified for real-time
treatment, most applications simply cannot operate properly on low-
bitrate links with the header overhead implied by the combination of
HDLC/PPP, IP, UDP, and RTP, i.e. they absolutely require header
compression.

3.2. Header Compression

Header compression can be performed in a variety of elements and at a
variety of levels in the protocol architecture. As many vendors of
Internet Telephony products for PCs ship applications, the approach
that is most obvious to them is to reduce overhead by performing
header compression at the application level, i.e. above transport
protocols such as UDP (or actually by using a non-standard,
efficiently coded header in the first place).

Generally, header compression operates by installing state at both
ends of a path that allows the receiving end to reconstruct
information omitted at the sending end. Many good techniques for
header compression (RFC1144, [2]) operate on the assumption that the
path will not reorder the frames generated. This assumption does not
hold for end-to-end compression; therefore additional overhead is
required for resequencing state changes and for compressed packets
making use of these state changes.

Assume that a very good application level header compression solution
for RTP flows could be able to save 11 out of the 12 bytes of an RTP
header [3]. Even this perfect solution only reduces the total header
overhead by 1/4. It would have to be deployed in all applications,

even those that operate on systems that are attached to higher-
bitrate links.

Because of this limited effectiveness, the AVT group that is
responsible for RTP within the IETF has decided to not further pursue
application level header compression.

For router and IP stack vendors, the obvious approach is to define
header compression that can be negotiated between peer routers.

Advanced header compression techniques now being defined in the IETF
[2] certainly can relieve the link from significant parts of the
IP/UDP overhead (i.e., most of 28 of the 44 bytes mentioned above).

One of the design principles of the new IP header compression
developed in conjunction with IPv6 is that it stops at layers the
semantics of which cannot be inferred from information in lower layer
(outer) headers. Therefore, this header compression technique alone
cannot compress the data that is contained within UDP packets.

Any additional header compression technique runs into a problem: If
it assumes specific application semantics (i.e., those of RTP and a
payload data format) based on heuristics, it runs the risk of being
triggered falsely and (e.g. in case of packet loss) reconstructing
packets that are catastrophically incorrect for the application
actually being used. A header compression technique that can be
operated based on heuristics but does not cause incorrect
decompression even if the heuristics failed is described in [7]; a
companion document describes the mapping of this technique to PPP
[10].

With all of these techniques, the total IP/UDP/RTP header overhead
for an audio stream can be reduced to two bytes per packet. This
technology need only be deployed at bottleneck links; high-speed
links can transfer the real-time streams without routers or switches
eXPending CPU cycles to perform header compression.

4. Principles of Real-Time Encapsulation for Low-Bitrate Links

The main design goal for a real-time encapsulation is to minimize the
delay incurred by real-time packets that become available for sending
while a long data packet is being sent. To achieve this, the
encapsulation must be able to either abort or suspend the transfer of
the long data packet. As an additional goal is to minimize the
overhead required for the transmission of packets from periodic
flows, this strongly argues for being able to suspend a packet, i.e.
segment it into parts between which the real-time packets can be
transferred.

4.1. Using existing IP fragmentation

Transmitting only part of a packet, to allow higher-priority traffic
to intervene and then resuming its transmission later on, is a kind
of fragmentation. Fragmentation is an existing functionality of the
IP layer: An IPv4 header already contains fields that allow a large
IP datagram to be fragmented into small parts. A sender's "real-time
PPP" implementation might simply indicate a small MTU to its IP stack
and thus cause all larger datagrams to be fragmented down to a size
that allows the Access delay goals to be met (this assumes that the
IP stack is able to priority-tag fragments, or that the PPP
implementation is able to correlate the fragments to the initial one
that carries the information relevant for prioritizing, or that only
initial fragments can be high-priority). (Also, a PPP implementation
can negotiate down the MTU of its peer, causing the peer to fragment
to a small size, which might be considered a crude form of
negotiating an access delay goal with the peer system -- if that
system supports priority queueing at the fragment level.)

Unfortunately, a full, 20 byte IP header is needed for each fragment
(larger when IP options are used). This limits the minimum size of
fragments that can be used without too much overhead. (Also, the
size of non-final fragments must be a multiple of 8 bytes, further
limiting the choice.) With path MTU discovery, IP level
fragmentation causes TCP implementations to use small MSSs -- this
further increases the per-packet overhead to 40 bytes per fragment.

In any case, fragmentation at the IP level persists on the path
further down to the datagram receiver, increasing the transmission
overheads and router load throughout the network. With its high
overhead and the adverse effect on the Internet, IP level
fragmentation can only be a stop-gap mechanism when no other
fragmentation protocol is available in the peer implementation.

4.2. Link-Layer Mechanisms

Cell-oriented multiplexing techniques such as ATM that introduce
regular points where cells from a different packet can be
interpolated are too inefficient for low-bitrate links; also, they
are not supported by chips used to support the link layer in low-
bitrate routers and host interfaces.

Instead, the real-time encapsulation should as far as possible make
use of the capabilities of the chips that have been deployed. On
synchronous lines, these chips support HDLC framing; on asynchronous
lines, an asynchronous variant of HDLC that usually is implemented in
software is being used. Both variants of HDLC provide a delimiting
mechanism to indicate the end of a frame over the link. The obvious
solution to the segmentation problem is to combine this mechanism
with an indication of whether the delimiter terminates or suspends
the current packet.

This indication could be in an octet appended to each frame
information field; however, seven out of eight bits of the octet
would be wasted. Instead, the bit could be carried at the start of
the next frame in conjunction with multiplexing information (PPP
protocol identifier etc.) that will be required here anyway. Since
the real-time flows will in general be periodic, this multiplexing
information could convey (part of) the compressed form of the header
for the packet. If packets from the real-time flow generally are of
constant length (or have a defined maximum length that is often
used), the continuation of the suspended packet could be immediately
attached to it, without expending a further frame delimiter, i.e.,
the interpolation of the real-time packet would then have zero
overhead. Since packets from low-delay real-time flows generally
will not require the ability to be further suspended, the
continuation bit could be reserved for the non-real-time packet
stream.

One real-time encapsulation format with these (and other) functions
is described in ITU-T H.223 [13], the multiplex used by the H.324
modem-based videophone standard [14]. It was investigated whether
compatibility could be achieved with this specification, which will
be used in future videophone-enabled (H.324 capable) modems.
However, since the multiplexing capabilities of H.223 are limited to
15 schedules (definitions of sequences of packet types that can be
identified in a multiplex header), for general Internet usage a
superset or a more general encapsulation would have been required.
Also, a PPP-style negotiation protocol was needed instead of using
(and necessarily extending) ITU-T H.245 [15] for setting the
parameters of the multiplex. In the PPP context, the interactions
with the encapsulations for data compression and link layer
encryption needed to be defined (including operation in the presence
of padding). But most important, H.223 requires synchronous HDLC
chips that can be configured to send frames without an attached CRC,
which is not possible with all chips deployed in commercially
available routers; so complete compatibility was unachievable.

Instead of adopting H.223, it was decided to pursue an approach that
is oriented towards compatibility both with existing hardware and
existing software (in particular PPP) implementations. The next
subsection groups these implementations according to their
capabilities.

4.3. Implementation models

This section introduces a number of terms for types of
implementations that are likely to emerge. It is important to have
these different implementation models in mind as there is no single
approach that fits all models best.

4.3.1. Sender types

There are two fundamental approaches to real-time transmission on
low-bitrate links:

Sender type 1
The PPP real-time framing implementation is able to control the
transmission of each byte being transmitted with some known,
bounded delay (e.g., due to FIFOs). For example, this is
generally true of PC host implementations, which directly access
serial interface chips byte by byte or by filling a very small
FIFO. For type 1 senders, a suspend/resume type approach will be
typically used: When a long frame is to be sent, the attempt is to
send it undivided; only if higher priority packets come up during
the transmission will the lower-priority long frame be suspended
and later resumed. This approach allows the minimum variation in
access delay for high-priority packets; also, fragmentation
overhead is only incurred when actually needed.

Sender type 2
With type 2 senders, the interface between the PPP real-time
framing implementation and the transmission hardware is not in
terms of streams of bytes, but in terms of frames, e.g., in the
form of multiple (prioritized) send queues directly supported by
hardware. This is often true of router systems for synchronous
links, in particular those that have to support a large number of
low-bitrate links. As type 2 senders have no way to suspend a
frame once it has been handed down for transmission, they
typically will use a queues-of-fragments approach, where long
packets are always split into units that are small enough to
maintain the access delay goals for higher-priority traffic.
There is a trade-off between the variation in access delay
resulting from a large fragment size and the overhead that is
incurred for every long packet by choosing a small fragment size.

4.3.2. Receiver types

Although the actual work of formulating transmission streams for
real-time applications is performed at the sender, the ability of the
receiver to immediately make use of the information received depends
on its characteristics:

Receiver type 1
Type 1 receivers have full control over the stream of bytes
received within PPP frames, i.e., bytes received are available
immediately to the PPP real-time framing implementation (with some
known, bounded delay e.g. due to FIFOs etc.).

Receiver type 2
With type 2 receivers, the PPP real-time framing implementation
only gets hold of a frame when it has been received completely,
i.e., the final flag has been processed (typically by some HDLC
chip that directly fills a memory buffer).

4.4. Conclusion

As a result of the diversity in capabilities of current
implementations, there are now two specifications for real-time
encapsulation: One, the multi-class extension to the PPP multi-link
protocol, is providing the solution for the queues-of-fragments
approach by extending the single-stream PPP multi-link protocol by
multiple classes [8]. The other encapsulation, PPP in a real-time
oriented HDLC-like framing, builds on this specification end extends
it by a way to dynamically delimit multiple fragments within one HDLC
frame [9], providing the solution for the suspend/resume type
approach.

5. Principles of Header Compression for Real-Time Flows

A good baseline for a discussion about header compression is in the
new IP header compression specification that was designed in
conjunction with the development of IPv6 [2]. The techniques used
there can reduce the 28 bytes of IPv4/UDP header to about 6 bytes
(depending on the number of concurrent streams); with the remaining 4
bytes of HDLC/PPP overhead and 12 bytes for RTP the total header
overhead can be about halved but still exceeds the size of a G.723.1
ACELP frame. Note that, in contrast to IP header compression, the
environment discussed here assumes the existence of a full-duplex PPP
link and thus can rely on negotiation where IP header compression
requires repeated transmission of the same information. (The use of
the architecture of the present document with link layer multicasting
has not yet been examined.)

Additional design effort was required for RTP header compression.
Applying the concepts of IP header compression, of the (at least) 12
bytes in an RTP header, 7 bytes (timestamp, sequence, and marker bit)
would qualify as RANDOM; DELTA encoding cannot generally be used
without further information since the lower layer header does not
unambiguously identify the semantics and there is no TCP checksum
that can be relied on to detect incorrect decompression. Only a more
semantics-oriented approach can provide better compression (just as
RFC1144 can provide very good compression of TCP headers by making
use of semantic knowledge of TCP and its checksumming method).

For RTP packets, differential encoding of the sequence number and
timestamps is an efficient approach for certain cases of payload data
formats. E.g., speech flows generally have sequence numbers and
timestamp fields that increase by 1 and by the frame size in
timestamp units, resp.; the CRTP (compressed RTP) specification makes
use of this relationship by encoding these fields only when the
second order difference is non-zero [7].

6. Announcement Protocols Used by Applications

As argued, the compressor can operate best if it can make use of
information that clearly identifies real-time streams and provides
information about the payload data format in use.

If these systems are routers, this consent must be installed as
router state; if these systems are hosts, it must be known to their
networking kernels. Sources of real-time information flows are
already describing characteristics of these flows to their kernels
and to the routers in the form of TSpecs in RSVP PATH messages [4].
Since these messages make use of the router alert option, they are
seen by all routers on the path; path state about the packet stream
is normally installed at each of these routers that implement RSVP.
Additional RSVP objects could be defined that are included in PATH
messages by those applications that desire good performance over low-
bitrate links; these objects would be coded to be ignored by routers
that are not interested in them (class number 11bbbbbb as defined in
[4], section 3.10).

Note that the path state is available in the routers even when no
reservation is made; this allows informed compression of best-effort
traffic. It is not quite clear, though, how path state could be torn
down quickly when a source ceases to transmit.

7. Elements of Hop-By-Hop Negotiation Protocols

The IP header compression specification attempts to account for
simplex and multicast links by providing information about the
compressed streams only in the forward direction. E.g., a full
IP/UDP header must be sent after F_MAX_TIME (currently 3 seconds),
which is a negligible total overhead (e.g. one full header every 150
G.723.1 packets), but must be considered carefully in scheduling the
real-time transmissions. Both simplex and multicast links are not
prevailing in the low-bitrate environment (although multicast
functionality may become more important with wireless systems); in
this document, we therefore assume full-duplex capability.

As compression techniques will improve, a negotiation between the two
peers on the link would provide the best flexibility in
implementation complexity and potential for extensibility. The peer
routers/hosts can decide which real-time packet streams are to be
compressed, which header fields are not to be sent at all, which
multiplexing information should be used on the link, and how the
remaining header fields should be encoded. PPP, a well-tried suite
of negotiation protocols, is already used on most of the low-bitrate
links and seems to provide the obvious approach. Cooperation from
PPP is also needed to negotiate the use of real-time encapsulations
between systems that are not configured to automatically do so.
Therefore, PPP options that can be negotiated at the link setup (LCP)
phase are included in [8], [9], and [10].

8. Security Considerations

Header compression protocols that make use of assumptions about
application protocols need to be carefully analyzed whether it is
possible to subvert other applications by maliciously or
inadvertently enabling their use.

It is generally not possible to do significant hop-by-hop header
compression on encrypted streams. With certain security policies, it
may be possible to run an encrypted tunnel to a network access server
that does header compression on the decapsulated packets and sends
them over an encrypted link encapsulation; see also the short mention
of interactions between real-time encapsulation and encryption in
section 4 above. If the security requirements permit, a special RTP
payload data format that encrypts only the data may preferably be
used.

9. References

[1] Handley, M., Crowcroft, J., Bormann, C. and J. Ott, "The
Internet Multimedia Conferencing Architecture", Work in
Progress.

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

[3] Scott Petrack, Ed Ellesson, "Framework for C/RTP: Compressed
RTP Using Adaptive Differential Header Compression",
contribution to the mailing list rem-conf@es.net, February
1996.

[4] Braden, R., Zhang, L., Berson, S., Herzog, S. and S. Jamin,
"Resource ReSerVation Protocol (RSVP) -- Version 1 Functional
Specification", RFC2205, September 1997.

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

[6] Schulzrinne, H., Casner, S., Frederick, R. and V. Jacobson,
"RTP: A Transport Protocol for Real-Time Applications", RFC
1889, January 1996.

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

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

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

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

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

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

[13] ITU-T Recommendation H.223, "Multiplexing protocol for low bit
rate multimedia communication", International Telecommunication
Union, Telecommunication Standardization Sector (ITU-T), March
1996.

[14] ITU-T Recommendation H.324, "Terminal for low bit rate
multimedia communication", International Telecommunication
Union, Telecommunication Standardization Sector (ITU-T), March
1996.

[15] ITU-T Recommendation H.245, "Control protocol for multimedia
communication", International Telecommunication Union,
Telecommunication Standardization Sector (ITU-T), March 1996.

10. Author's Address

Carsten Bormann
Universitaet Bremen FB3 TZI
Postfach 330440
D-28334 Bremen, GERMANY

Phone: +49.421.218-7024
Fax: +49.421.218-7000
EMail: cabo@tzi.org

Acknowledgements

Much of the early discussion that led to this document was done with
Scott Petrack and Cary Fitzgerald. Steve Casner, Mikael Degermark,
Steve Jackowski, Dave Oran, the other members of the ISSLL subgroup
on low bitrate links (ISSLOW), and in particular the ISSLL WG co-
chairs Eric Crawley and John Wroclawski have helped in making this
architecture a reality.

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.