[Docs] [txt|pdf] [Tracker] [WG] [Email] [Diff1] [Diff2] [Nits]
Versions: 00 01 02 03 RFC 3366
Internet Engineering Task Force Gorry Fairhurst
INTERNET DRAFT University of Aberdeen
Lloyd Wood
Cisco Systems Ltd
July 2001
draft-ietf-pilc-link-arq-issues-02.txt Expires: January 2002
Advice to link designers on link Automatic Repeat reQuest (ARQ)
STATUS OF THIS MEMO
This document is an Internet-Draft and is in full conformance with
all provisions of Section 10 of RFC2026.
Internet-Drafts are working documents of the Internet Engineering
Task Force (IETF), its areas, and its working groups. Note that
other groups may also distribute working documents as
Internet-Drafts.
Internet-Drafts are draft documents valid for a maximum of six
months and may be updated, replaced, or obsoleted by other documents
at any time. It is inappropriate to use Internet-Drafts as
reference material or to cite them other than as "work in progress".
The list of current Internet-Drafts can be accessed at
http://www.ietf.org/ietf/1id-abstracts.txt
The list of Internet-Draft Shadow Directories can be accessed at
http://www.ietf.org/shadow.html
ABSTRACT
This document provides advice to the designers of digital
communication equipment and link-layer protocols employing link
layer Automatic Repeat reQuest (ARQ) techniques. This document
presumes that the designers wish to support Internet protocols, but
may be unfamiliar with the architecture of the Internet and with the
implications of their design choices on the performance and
efficiency of Internet traffic carried over their links.
ARQ is described in a general way that includes its use over a wide
range of underlying physical media, including cellular wireless,
wireless LANs, RF links, and other types of bearer channel. This
document also describes issues relevant to supporting IP traffic
over physical channels where performance varies, and link ARQ is
likely to be used.
PILC WORKING GROUP INTERNET DRAFT expires January 2002 1
Advice to link designers on link Automatic Repeat reQuest July 2001
CONTENTS
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1 Link ARQ 4
1.2 Causes of packet loss on a link . . . . . . . . . . . . .5
1.3 Why use ARQ? 6
1.4 Commonly-used ARQ techniques . . . . . . . . . . . . . . 6
1.4.1 Stop-and-wait ARQ 7
1.4.2 Sliding-window ARQ. . . . . . . . . . . . . . . . .7
1.5 Causes of delay across a link 8
2. ARQ persistence. . . . . . . . . . . . . . . . . . . . . . . 9
2.1 Perfectly-persistent (reliable) ARQ protocols 10
2.2 High-persistence (highly-reliable) ARQ protocols. . . . 11
2.3 Low-persistence (partially-reliable) ARQ protocols 12
2.4 Choosing your persistency. . . . . . . . . . . . . . . .13
2.5 Impact of channel outages 13
3. Treatment of packets and flows. . . . . . . . . . . . . . . 15
3.1 Packet ordering 15
3.2 Using link ARQ to support multiple flows . . . . . . . .16
3.3 Differentiation of link service classes 17
4. Conclusions . . . . . . . . . . . . . . . . . . . . . . . . 19
5. Security implications 20
6. IANA considerations. . . . . . . . . . . . . . . . . . . . .21
7. Acknowledgements 21
8. References. . . . . . . . . . . . . . . . . . . . . . . . . 21
Authors' addresses 24
Full copyright statement. . . . . . . . . . . . . . . . . . . .25
1. INTRODUCTION
IP, the Internet Protocol [RFC791], forms the core protocol of the
global Internet and defines a simple "connectionless" packet-
switched network. Over the years, Internet traffic using IP has
been carried over a wide variety of links, of vastly different
capacities, delays and loss characteristics. In the future, IP
traffic can be expected to continue to be carried over a very wide
variety of new and existing link designs, again of varied
characteristics.
A companion document [DRAFTKARN01] describes the general issues
associated with link design. This document should be read in
conjunction with that and with other documents produced by the
Performance Implications of Link Characteristics (PILC) IETF
workgroup [DRAFTDAW01, RFC3135].
PILC WORKING GROUP INTERNET DRAFT expires January 2002 2
Advice to link designers on link Automatic Repeat reQuest July 2001
This document is intended for three distinct groups of readers:
a. Link designers wishing to configure (or tune) a link for the IP
traffic that it will carry, using standard link-layer mechanisms
such as the ISO High-level Data Link Control (HDLC) [ISO4335a] or
its derivatives.
b. Link implementers who may wish to design new link mechanisms that
perform well for IP traffic.
c. The community of people using or developing TCP, UDP and related
protocols, who may wish to be aware of the ways in which links
can operate.
The primary audiences are intended to be groups (a) and (b). Group
(c) should not need to be aware of the exact details of an ARQ
scheme across a single link, and should not have to consider it in
their implementations; much of the Internet runs across links that
do not use any form of ARQ. However, the TCP/IP community does need
to be aware that the IP protocol operates over a diverse range of
underlying subnetworks. This document may help to raise that
awareness.
Perfect reliability is not a requirement for IP networks or for
links [DRAFTKARN01]. IP networks may discard packets due to a
variety of reasons entirely unrelated to link errors, including lack
of queuing space, congestion management, faults, and route changes.
It has long been widely understood that perfect end-to-end
reliability can be ensured only at the transport layer [SALT81].
Some familiarity with TCP, the Transmission Control Protocol
[STE94], is presumed here. TCP provides a reliable byte-stream
transport service, building upon the best-effort service provided by
the Internet Protocol. TCP achieves this by dividing data into TCP
segments, and transporting these segments in IP packets. TCP
guarantees that a TCP session will retransmit the TCP segments
contained in any data packets that are lost along the Internet path
between endhosts. TCP normally performs retransmission using its
Fast Retransmit procedure, but if the loss fails to be detected (or
retransmission is unsuccessful), TCP falls back to a Retransmission
Time Out (RTO) retransmission using a timer [RFC2581]. (Link
protocols also implement timers to verify integrity of the link, and
to assist link ARQ.) TCP also copes with any duplication or
reordering introduced by the IP network. There are a number of
variants of TCP, with differing levels of sophistication in their
procedures for handling loss recovery and congestion avoidance. Far
from being static, the TCP protocol is itself subject to ongoing
gradual refinement and evolution, e.g. [RFC2488, RFC2760].
PILC WORKING GROUP INTERNET DRAFT expires January 2002 3
Advice to link designers on link Automatic Repeat reQuest July 2001
Internet networks may reasonably be expected to carry traffic from a
wide and evolving range of applications. Not all applications
require or benefit from using the reliable service provided by TCP.
In the Internet, these applications are carried by alternate
transport protocols, such as the User Datagram Protocol (UDP)
[RFC768].
1.1 LINK ARQ
At the link layer, ARQ operates on blocks of data, known as frames,
and attempts to deliver frames from the sender to the receiver over
a physical link. A link frame may contain one or more complete IP
packets.
Frames often have a small fixed or maximum size for convenience of
processing by Medium-Access Control (MAC) and link protocols. This
contrasts with the variable lengths of IP datagrams, or 'packets'.
An ARQ link mechanism relies on an integrity check for each frame
(e.g. strong link-layer CRC [DRAFTKARN01]) and a retransmission
process to retransmit lost or errored frames.
Links may be full-duplex (allowing two-way communication over
separate forward and reverse physical channels), half-duplex (where
two-way communication uses a shared forward and reverse physical
channel, e.g. IrDA, IEEE 802.11) or simplex (where one channel
allows communication in only one direction). ARQ requires a forward
and return path, and therefore link ARQ may be used over links that
employ full- or half-duplex links. When the channel is shared
between two or more link nodes, a link MAC protocol is required to
ensure all nodes requiring transmission get access to the shared
physical channel. Such schemes may add to the delay (jitter)
associated with transmission of packet data and ARQ control frames.
When using ARQ over a link where the probability of frame loss is
related to the frame size, there is an optimal frame size for any
specific target error rate. To allow for efficient use of the
channel, the maximum link frame size may therefore be considerably
lower than the maximum IP datagram size expressed in the IP Maximum
Transmission Unit (MTU). Each frame will then contain only a
fraction of an IP packet and transparent implicit fragmentation of
the IP datagram is used [DRAFTKARN01]. A smaller frame size
introduces more frame header overhead per payload byte transported.
Explicit network-layer IP fragmentation is undesirable for a variety
of reasons, and should be avoided [KEN88, DRAFTKARN01]. Its use can
be minimised with use of Path MTU discovery for TCP connections
[RFC1191, RFC1435].
Another way to reduce the frame loss rate (or reduce transmit signal
power for the same rate of frame loss) is to use coding, e.g.
PILC WORKING GROUP INTERNET DRAFT expires January 2002 4
Advice to link designers on link Automatic Repeat reQuest July 2001
Forward Error Correction (FEC) [LIN93]. FEC is commonly included in
the design of wireless links and may be used simultaneously with
link ARQ. FEC schemes also exist which combine modulation and
coding and may also be adaptive. Hybrid ARQ [LIN93] combines
adaptive FEC with link ARQ procedures to reduce the probability of
loss of retransmitted frames. Interleaving is also used to reduce
the probability of frame loss by dispersing the occurrence of errors
more widely in the channel to improve error recovery; this adds
delay to the channel.
The document does not consider the use of link ARQ to support a
broadcast or multicast service within a subnetwork, where a link may
send a single packet to more than one recipient using a single link
transmit operation. Although such schemes are supported in some
subnetworks, they raise a number of additional issues.
Links supporting stateful reservation-based quality of service (QoS)
according to the Integrated Services (intserv) model are also not
explicitly discussed.
1.2 CAUSES OF PACKET LOSS ON A LINK
Not all packets sent to a link are necessarily received successfully
by the receiver at the other end of the link. There are a number of
possible causes of packet loss. These may occur as frames travel
across a link, and include:
a. Loss due to channel noise, often characterised by random frame
loss.
b. Frame loss due to channel interference. This interference can
be random, structured, and in some cases even periodic.
c. A link outage, causing loss due to unavailability of the
physical link for a period of time. This is a common
characteristic of some types of link, e.g. mobile cellular
radio. During the outage, the link loses all or virtually all
frames, until the link is restored.
d. Loss of a frame transmitted in a shared physical channel that is
using a contention-aware MAC protocol (e.g. due to collision).
In this case, many protocols require that retransmission is
deferred to promote stability of the shared channel (i.e.
prevent link congestion collapse), discussed further in section
1.5.
Other forms of packet loss are not related to channel conditions,
but include:
PILC WORKING GROUP INTERNET DRAFT expires January 2002 5
Advice to link designers on link Automatic Repeat reQuest July 2001
i. Packet discards due to congestion. Queues will eventually
overflow as the arrival rate of new packets to send continues to
exceed the outgoing packet transmission rate over the link.
ii. Loss due to implementation errors, including hardware faults and
software errors. This is recognised as a common cause of packet
corruption detected in the endhosts [STONE00].
The levels of loss and patterns of loss experienced are functions of
the design of the link's physical and link layers. These vary
significantly across different link configurations. The performance
of a specific implementation may also vary considerably across the
same link configuration when operated over different types of
physical channel.
1.3 WHY USE ARQ?
Reasons that encourage considering the use of ARQ include:
a. ARQ across a single link has a faster control loop than TCP's
acknowledgement control loop, which takes place over the longer
end-to-end path over which TCP must operate. It is therefore
possible for ARQ to provide more rapid retransmission of TCP
segments lost on the link, at least for a reasonable number of
retries [DRAFTDAW01, SALT81].
b. Link ARQ can operate on individual frames, using implicit
transparent link fragmentation [DRAFTKARN01]. Frames may be
much smaller than IP packets, and repetition of smaller frames
containing lost or errored parts of an IP packet may improve the
efficiency of the ARQ process and the efficiency of the link.
A link ARQ procedure may be able to use local knowledge that is not
available to endhosts, in order to optimise delivery performance for
the current link conditions. This information can include
information about the state of the link and physical layer channel,
e.g. knowledge of the current available transmission rate, the
prevailing error environment, or available transmit power in
wireless links.
1.4 COMMONLY-USED ARQ TECHNIQUES
A link ARQ protocol uses a link protocol mechanism to allow the
sender to detect lost or errored frames and to schedule
retransmission. Detection of frame loss may be via a link protocol
timer, by detecting missing positive link acknowledgment frames,
PILC WORKING GROUP INTERNET DRAFT expires January 2002 6
Advice to link designers on link Automatic Repeat reQuest July 2001
receiving explicit negative acknowledgment frames and/or by polling
the link receiver status.
Whatever the mechanisms that are chosen, there are two easily-
described categories of ARQ retransmission process that are widely
used:
1.4.1 STOP-AND-WAIT ARQ
A sender using stop-and-wait ARQ (sometimes known as 'Idle ARQ'
[LIN93]) transmits a single frame and then waits for an
acknowledgement from the receiver for that frame. The sender then
either continues transmission with the next frame, or repeats
transmission of the same frame if the original frame was lost or
errored.
Stop-and-wait ARQ is simple, if inefficient, for protocol designers
to implement, and therefore popular, e.g. tftp [RFC1350] at the
transport layer. However, when stop-and-wait ARQ is used in the
link layer, it is well-suited only to links with low bandwidth-delay
products. This technique is not discussed further in this document.
1.4.2 SLIDING-WINDOW ARQ
A protocol using sliding-window link ARQ [LIN93] numbers every frame
with a unique sequence number, according to a modulus. The modulus
defines the numbering base for frame sequence numbers, and the size
of the sequence space. The largest sequence number value is viewed
by the link protocol as contiguous with the first (0), since the
numbering space wraps around.
TCP is itself a sliding-window protocol at the transport layer
[STE94], so similarities between a link-interface-to-link-interface
protocol and end-to-end TCP may be recognisable. A sliding-window
link protocol is much more complex in implementation than the
simpler stop-and-wait protocol described in the previous section,
particularly if per-flow ordering is preserved.
At any time the link sender may have a number of frames outstanding
and awaiting acknowledgement, up to the space available in its
transmission window. A sufficiently large link sender window
(equivalent to or greater than the number of frames sent, or larger
than the bandwidth*delay product capacity of the link) permits
continuous transmission of new frames. A smaller link sender window
causes the sender to pause transmission of new frames until a
timeout or a control frame, such as an acknowledgement, is received.
When frames are lost, a larger window, i.e. more than the link's
bandwidth*delay product, is needed to allow continuous operation
while frame retransmission takes place.
PILC WORKING GROUP INTERNET DRAFT expires January 2002 7
Advice to link designers on link Automatic Repeat reQuest July 2001
The modulus numbering space determines the size of the frame header
sequence number field. This sequence space needs to be larger than
the link window size, and if using selective repeat ARQ, larger than
twice the link window size. For continuous operation, the sequence
space should be larger than the product of the link capacity and the
channel persistence (discussed in section 2.), so that in-flight
frames can be identified uniquely.
As with TCP, which provides sliding-window delivery across an entire
end-to-end path, rather than across a single link, there are a large
number of variations on the basic sliding-window implementation,
with increased complexity and sophistication to make them suitable
for various conditions. Selective Repeat (SR), also known as
Selective Reject (SREJ), and Go-Back-N, also known as Reject (REJ),
are examples of ARQ techniques using protocols implementing sliding
window ARQ.
1.5 CAUSES OF DELAY ACROSS A LINK
Links and link protocols contribute to the total path delay
experienced between communicating applications on endhosts. Delay
has a number of causes, including:
a. Input packet queuing and frame buffering at the link head before
transmission over the link.
b. Retransmission back-off, an additional delay introduced for
retransmissions by some MAC schemes when operating over a shared
physical channel to prevent link congestion collapse. This
collapse may otherwise arise, if, for example, a set of link
receivers all retransmitted immediately after a collision on a
busy channel. Many protocols select a backoff delay, which
increases with the number of attempts taken to retransmit a
frame; analogies can be drawn with end-to-end TCP congestion
avoidance at the transport layer [RFC2581]. Retransmission
backoff is not required by protocols used on point-to-point
links, where retransmissions can be sent at the earliest
possible time.
c. Waiting for access to the allocated channel when the physical
channel is shared. There may be processing or protocol-induced
delay before transmission takes place [FER99, PAR00].
d. Frame serialisation and transmission processing. These are
functions of frame size and the transmission speed of the link.
PILC WORKING GROUP INTERNET DRAFT expires January 2002 8
Advice to link designers on link Automatic Repeat reQuest July 2001
e. Physical link propagation time, limited by the speed of
transmission of the signal in the physical medium of the
channel.
f. Per-frame processing including the cost of QoS scheduling,
packet interleaving, encryption, FEC and interleaving.
Interleaving and FEC also add substantial delay, and in some
cases additional jitter. Hybrid link ARQ schemes [LIN93] in
particular may incur significant receiver processing delay.
g. Packet processing, including buffering frame contents at the
receiver for packet reassembly, before onward transmission of
the packet.
When link ARQ is used, steps (b), (c), (d), (e), and (f) may be
repeated a number of times, for each time retransmission of a frame
occurs, increasing overall delay for the packet that the frame
partially carries. Adaptive ARQ schemes (e.g. hybrid ARQ using
adaptive FEC codes) may also incur extra per-frame processing for
retransmitted frames.
It is important to understand that applications and transport
protocols at the endhosts are unaware of the individual delays
contributed by each link in the path, and only see the overall path
delay. Application performance is therefore determined by the
cumulative delay of the entire end-to-end Internet path. This path
may include an arbitrary or even a widely-fluctuating number of
links, where each link may or may not use ARQ. As a result, it is
not possible to state fixed limits on the acceptable delay that a
link can add to a path; other links in the path will add an unknown
delay.
2. ARQ PERSISTENCE
ARQ protocols may be characterised by their persistency.
Persistence is the willingness of the protocol to retransmit lost
frames to ensure reliable delivery of traffic across the link.
A link's retransmission persistency defines how long the link is
allowed to delay an IP packet, in an attempt to transmit all the
frames carrying the packet and its content over the link, before
giving up and discarding the packet. This persistency can normally
be measured in milliseconds, but may, if the link propagation delay
is specified, be expressed in terms of the maximum number of link
retransmission attempts permitted. The latter does not always map
onto an exact time limit, for the reasons discussed in section 1.5.
PILC WORKING GROUP INTERNET DRAFT expires January 2002 9
Advice to link designers on link Automatic Repeat reQuest July 2001
An example of a reliable link protocol that is perfectly persistent
is the ISO HDLC protocol in the Asynchronous Balanced Mode (ABM)
[ISO4335a].
A protocol that only retransmits a number of times before giving up
is less persistent, e.g. Ethernet [FER99], IEEE 802.11, or GSM RLP
[RFC2757]. Here, lower persistence also ensures stability and fair
sharing of a shared subnetwork, even when many senders are
attempting retransmissions.
TCP, STCP [RFC2960] and a number of applications using UDP (e.g.
tftp [RFC1350]) implement their own end-to-end reliable delivery
mechanisms. Many TCP and UDP applications, e.g. streaming
multimedia, benefit from timely delivery from lower layers with
sufficient reliability, rather than perfect reliability with
increased link delays.
2.1 PERFECTLY-PERSISTENT (RELIABLE) ARQ PROTOCOLS
A perfectly-persistent ARQ protocol is one that attempts to provide
a reliable service, i.e. in-order delivery of packets to the other
end of the link, with no missing packets and no duplicate packets.
The perfectly-persistent ARQ protocol will repeat a lost or errored
frame an indefinite (and potentially infinite) number of times until
the frame is successfully received.
If traffic is going no further than across one link, and losses do
not occur within the endhosts, perfect persistence ensures
reliability between the two link ends without requiring any higher-
layer protocols. This reliability can become counterproductive for
traffic traversing multiple links, as it duplicates and interacts
with functionality in protocol mechanisms at higher layers [SALT81].
Arguments against perfect persistence for IP traffic include:
a. Variable link delay; the impact of ARQ introduces a degree of
jitter, a function of the link's physical delay and frame
serialisation and transmission times (see section 1.5), to all
flows sharing a link performing frame retransmission.
b. Perfect persistence does not provide a clear upper bound on the
maximum retransmission delay for the link. Significant changes
in path delay caused by excessive link retransmissions may lead
to timeout of TCP retransmission timers, although a high
variance in link delay and the resulting overall path delay may
also cause a large TCP RTO value to be selected [LUD99b, PAR00].
Both result in a decrease in TCP throughput.
PILC WORKING GROUP INTERNET DRAFT expires January 2002 10
Advice to link designers on link Automatic Repeat reQuest July 2001
c. Applications needing perfectly-reliable delivery can implement a
form of perfectly-persistent ARQ themselves, or use a reliable
transport protocol within the endhosts. Implementing perfect
persistence at each link along the path between the endhosts is
redundant, but cannot ensure the same reliability as end-to-end
transport [SALT81].
d. Link ARQ should not adversely delay the flow of end-to-end
control information. As an example, the ARQ retransmission of
data for one or more flows should not excessively extend the
protocol control loops. Excessive delay of duplicate TCP
acknowledgments (dupacks [STE94, BAL97]), SACK, or Explicit
Congestion Notification (ECN) indicators will reduce the
responsiveness of TCP flows to congestion events. Similar
issues exist for TCP Friendly Rate Control (TFRC [DRAFTHAN01])
when equation-based congestion control is used with UDP.
Perfectly-persistent link protocols that perform unlimited ARQ, i.e.
that continue to retransmit frames indefinitely until the frames are
successfully received, are seldom found in reality.
Most practical link protocols give up retransmission at some point,
but do not necessarily do so with the intention of bounding the ARQ
retransmission persistence. A protocol may, for instance, terminate
retransmission after a link connection failure, e.g. after no frames
have been successfully received within a pre-configured timer
period. The time a protocol retransmits a specific frame (or the
maximum number of retransmissions) therefore becomes a function of
many different parameters (ARQ procedure, link timer values, and
procedure for link monitoring), rather than being pre-configured.
Another common feature of this type of behaviour is that some
protocol implementers assume that after a link failure, queued data
is no longer significant and discard packets when giving up ARQ
retransmission.
Examples of ARQ protocols that are perfectly persistent include
ISO/ITU-T LAP-B [ISO7776] and ISO HDLC in the Asynchronously
Balanced Mode (ABM) [ISO4335a], e.g. using Multiple Selective Reject
(MSREJ [ISO4335b]). These protocols will retransmit a frame an
unlimited number of times until receipt of the frame is
acknowledged.
2.2 HIGH-PERSISTENCE (HIGHLY-RELIABLE) ARQ PROTOCOLS
High-persistence ARQ protocols limit the time (or number of
attempts) that ARQ may retransmit a particular frame before the
sender gives up on retransmission of the missing frame and moves on
to forwarding subsequent buffered in-sequence frames. Giving up
PILC WORKING GROUP INTERNET DRAFT expires January 2002 11
Advice to link designers on link Automatic Repeat reQuest July 2001
frame retransmission does not imply a lack of link connectivity and
does not cause a link protocol state change.
It has been recommended that a single IP packet should never be
delayed by the network for more than the Maximum Segment Lifetime
(MSL) of 120 seconds defined for TCP [RFC1122]. It is, however,
practically difficult to bound the maximum path delay of an Internet
path. One case where segment (packet) lifetime may be significant
is where alternate paths of different delays exist between endhosts
and route flapping or flow-unaware traffic engineering is used. Some
TCP packets may follow a short path, while others follow a much
longer path, e.g. using persistent ARQ over a link outage.
Failure to limit the maximum packet lifetime can result in TCP
sequence numbers wrapping at high transmission rates, where old data
segments may be confused with newer segments if the sequence number
space has been exhausted and reused in the interim. Some TCP
implementations use the Round Trip Timestamp Measurement (RTTM)
option in TCP packets to remove this ambiguity, using the Protection
Against Wrapped Sequence number (PAWS) algorithm [RFC1323].
In practice, the MSL is usually very large compared to the typical
TCP RTO. The calculation of TCP RTO is based on estimated round-
trip path delay. If the number of link retransmissions causes a
path delay larger than the value of RTO, the TCP retransmission
timer can expire, leading to a timeout and retransmission of a
segment (packet) by the TCP sender.
Although high persistency may benefit bulk flows, the additional
delay (and variations in delay) that it introduces may be highly
undesirable for other types of flows. Being able to treat flows
separately with different classes of link service is useful, and is
discussed in section 3.3.
Examples of high-persistence ARQ protocols include [BHA97, ECK98,
LUD99a, MEY99].
2.3 LOW-PERSISTENCE (PARTIALLY-RELIABLE) ARQ PROTOCOLS
The characteristics of a link using a low-persistence ARQ protocol
may be summarised as:
a. The link is not perfectly reliable and does not provide an
absolute guarantee of delivery, i.e. the transmitter will
discard some frames as it 'gives up' before receiving an
acknowledgement of successful transmission across the link.
b. There is a lowered limit on the maximum added delay that IP
packets will experience when travelling across the link
PILC WORKING GROUP INTERNET DRAFT expires January 2002 12
Advice to link designers on link Automatic Repeat reQuest July 2001
(typically lower than the TCP path RTO). This reduces
interaction with TCP timers or with UDP-based error-control
schemes.
c. The link offers a low bound for the time that retransmission for
any one frame can block completed transmission and assembly of
other correctly-received IP packets originally sent before the
missing frame. Limiting delay avoids aggravating contention or
interaction between different packet flows (see also section
3.2).
Examples of low-persistence ARQ protocols include [SAM96, WARD95,
CHE00].
2.4 CHOOSING YOUR PERSISTENCY
The TCP Maximum RTO is an upper limit on the maximum time that TCP
will wait until it performs a retransmission. Most TCP
implementations will generally have a TCP RTO of at least several
times the path delay.
Setting a lower link persistency (e.g. of the order 2-5
retransmission attempts) reduces interaction with the TCP RTO timer,
and may therefore reduce the probability of duplicate copies of the
same packet being present in the link transmit buffer under some
patterns of loss.
Links with a low propagation delay may allow tens of retransmission
attempts to deliver a single frame, and still satisfy a bound for
(b) in section 2.3. In this case, a low-delay link is defined as
one where the total packet transmission time is much less than 100
ms (this is a common value for the granularity of the internal TCP
system timer).
A packet may traverse a number of successive links on its total end-
to-end path. This is therefore an argument for much lower
persistency on any individual link, as delay due to persistency
accumulates along the path for each packet.
Some implementers have chosen a lower persistence, falling back on
the judgement of TCP or a UDP application to retransmit any packets
that are not recovered by the link ARQ protocol.
2.5 IMPACT OF CHANNEL OUTAGES
Channels experiencing a persistent loss event, where many
consecutive frames are corrupted over an extended time, may also
need to be considered. Examples of this type of channel behaviour
PILC WORKING GROUP INTERNET DRAFT expires January 2002 13
Advice to link designers on link Automatic Repeat reQuest July 2001
include fading, roaming, and some forms of interference. During the
loss event, there is an increased probability that a retransmission
request may be corrupted, and/or an increased probability that a
retransmitted frame will also be lost. This type of loss event is
often known as an 'transient outage'.
If the transient outage extends for longer than the TCP RTO, the TCP
sender will also perform transport layer retransmission. At the
same time, the TCP sender will reduce its cwnd to 1 segment
(packet), recalculate its RTO, and wait for an ACK packet. If none
is received, TCP will retransmit again, up to a retry limit. TCP
only determines that the outage is over (i.e. that path capacity is
restored) by receipt of an ACK. If the link ARQ protocol
persistency causes the link to discard the ACK, the TCP sender must
wait for the next RTO retransmission to discover the link is
restored. This can be many seconds after the end of the loss event.
When a link layer is able to differentiate a set of link service
classes (see section 3.3), a link ARQ persistency longer than the
largest link loss event may benefit a TCP session. This would allow
TCP to rapidly restore transmission without the need to wait for a
retransmission time out, generally improving TCP performance in the
face of transient outages. Implementation of such schemes remains a
research issue.
When an outage occurs for a sender sharing a common physical channel
with other nodes, uncontrolled high persistence can continue to
consume transmission resources for the duration of the outage. This
may be undesirable, since it reduces the capacity available for
other nodes sharing the channel, which do not necessarily experience
the same outage. These nodes could otherwise use the channel for
more productive transfers. The persistence is often limited by
another controlling mechanism in this case. To counter such
effects, ARQ protocols may delay retransmission requests, or defer
retransmission of requested frames until the outage ends for the
sender.
An alternate suggested approach for a link layer that is able to
identify separate flows is to use low link persistency (section 2.3)
along with a higher-level mechanism, for example one that attempts
to deliver one TCP packet (whole TCP segment) per TCP flow after a
loss event [DRAFTKARN01]. This is intended to ensure that TCP
transmission is restored rapidly. Algorithms to implement this
remain an area of future research.
PILC WORKING GROUP INTERNET DRAFT expires January 2002 14
Advice to link designers on link Automatic Repeat reQuest July 2001
3. TREATMENT OF PACKETS AND FLOWS
3.1 PACKET ORDERING
A common debate is whether a link should be allowed to forward
packets in an order different to that in which they were originally
received at its transmit interface.
IP networks are not required to deliver all IP packets in order,
although generally networks do deliver most IP packets in their
original transmission order. Routers supporting class-based queuing
do reorder received packets, by reordering packets in different
flows, but these usually retain per-flow ordering.
Policy-based queuing, allowing fairer access to the link, may also
reorder packets. There is still much debate on optimal algorithms,
and on optimal queue sizes for particular link speeds. This,
however, is not related to use of link ARQ and applies to any
(potential) bottleneck router.
Although small amounts of reordering are common in IP networks
[BEN00], significant re-ordering within a flow is undesirable as it
can have a number of effects:
a. Reordering will increase packet jitter for real-time
applications. This may lead to application data loss if a small
play-out buffer is used by the receiving application.
b. Reordering will interleave arrival of TCP segments, leading to
generation of duplicate ACKs (dupacks), leading to assumptions
of loss. A sequence of three identical dupacks causes the TCP
sender to trigger fast retransmission and recovery, a form of
congestion avoidance, since TCP always presumes loss due to
congestion [RFC2581, STE94]. This reduces TCP throughput
efficiency as far as the application is concerned, but it should
not impact data integrity.
In addition, reordering may negatively impact processing by some
existing poorly-implemented TCP/IP stacks, by leading to unwanted
side-effects in poorly-implemented IP fragment reassembly code,
poorly-implemented IP demultiplexing (filter) code, or poorly-
implemented UDP applications.
Ordering effects must also be considered when breaking the end-to-
end paradigm and evaluating transport-level relays such as split-TCP
implementations or Protocol Enhancing Proxies [RFC3135].
As with total path delay, TCP and UDP flows are impacted by the
cumulative effect of reordering along the entire path. Link
protocol designers must not assume that their link is the only link
PILC WORKING GROUP INTERNET DRAFT expires January 2002 15
Advice to link designers on link Automatic Repeat reQuest July 2001
undertaking packet reordering, as some level of reordering may be
introduced by other links along the same path, or by router
processing within the network [BEN00]. Ideally, the link protocol
should not contribute to reordering within a flow, or at least
ensure that it does not significantly increase the level of
reordering within the flow. To achieve this, buffering is required
at the link receiver. The total amount of buffering required is a
function of the link's bandwidth*delay product and the level of ARQ
persistency, and is bounded by the link window size.
A number of experimental ARQ protocols have allowed out-of-order
delivery [BAL95, SAM96, WARD95].
3.2 USING LINK ARQ TO SUPPORT MULTIPLE FLOWS
Most links can be expected to carry more than one IP flow at a time.
Some high-capacity links are expected to carry a very large number
of simultaneous flows, often from and to a large number of different
endhosts. With use of multiple applications at an endhost, multiple
flows can be considered the norm rather than the exception, even for
last-hop links.
When packets from several flows are simultaneously in transit within
a link ARQ protocol, ARQ may cause a number of additional effects:
a. ARQ introduces variable delay (jitter) to a TCP flow sharing a
link with another flow experiencing loss. This additional
delay, introduced by the need for a link to provide in-sequence
delivery of packets, may adversely impact other applications
sharing the link, and can increase the duration of the initial
slow-start period for TCP flows for these applications. This is
significant for short-lived TCP flows (e.g. those used by
HTTP/1.0 and earlier), which spend most of their lives in slow-
start.
b. ARQ introduces jitter to UDP flows that share a link with
another flow experiencing loss. An end-to-end protocol may not
require reliable delivery, particularly if it is supporting a
delay-sensitive application.
c. High-persistence ARQ may delay packets long enough to cause
premature timeout of another TCP flow's RTO timer, although
modern TCP implementations should not experience this since
their computed RTO values should leave sufficient margin over
path RTTs to cope with reasonable amounts of jitter.
Reordering of packets belonging to different flows may be desirable
[LUD99b, CHE00] to achieve fair sharing of the link between
PILC WORKING GROUP INTERNET DRAFT expires January 2002 16
Advice to link designers on link Automatic Repeat reQuest July 2001
established bulk-data transfer sessions and sessions that transmit
less data but would benefit from lower link transit delay.
Preserving ordering within each individual flow, to avoid the
effects of reordering described earlier, is worthwhile.
3.3 DIFFERENTIATION OF LINK SERVICE CLASSES
High ARQ persistency is generally considered unsuitable for many
applications using UDP, where reliable delivery is not always
required and where it may introduce unacceptable jitter, but may
benefit bulk data transfers under certain link conditions. A scheme
that differentiates packet flows into two or more classes, to
provide different service to each class, is therefore desirable.
Observation of flow behaviour can tell you which flows are
controlled and congestion-sensitive, or uncontrolled and not, so
that you can treat them differently and ensure fairness. However,
this cannot tell you whether a flow is intended as reliable or
unreliable by its application, or what the application requires for
best operation.
Supporting different link services for different classes of flows
therefore requires that the link is able to distinguish the
different flows from each other. This generally needs an explicit
indication of the class associated with each flow.
Some potential schemes for indicating the class of a packet include:
a. Using the Type of Service (ToS) bits in the IP header [RFC791].
The IETF has replaced these globally-defined bits, which were
not widely used, with the differentiated services model
(diffserv [RFC2475]). In diffserv, each packet carries a
differentiated service code point (DSCP) which indicates which
one of a set of diffserv classes the flow belongs to. Each
router maps the DSCP value of a received IP packet to one of a
set of Per Hop Behaviours (PHBs) as the packet is processed
within the network. Diffserv uses include policy-based routing,
class-based queuing, and support for other QoS metrics,
including IP packet priority, delay, reliability, and cost.
b. Inspecting the network packet header and viewing the IP protocol
type [RFC791] to gain an idea of the transport protocol used and
thus guess its needs. This is not possible when carrying an
encrypted payload, e.g. using the IP security extensions (IPSec)
with Encapsulation Security Payload (ESP) [RFC1827] payload
encryption.
c. By inspecting the transport packet header information to view
the TCP or UDP headers and port numbers (e.g. [PAR00, BAL95]).
PILC WORKING GROUP INTERNET DRAFT expires January 2002 17
Advice to link designers on link Automatic Repeat reQuest July 2001
This is not possible when using payload encryption, e.g. IPSec
with ESP [RFC1827] payload encryption, and incurs processing
overhead in routers.
There are, however, some drawbacks to these schemes:
i. The ToS/Differentiated Services Code Point (DSCP) values
[RFC2475] may not be set reliably, and may be overwritten by
intermediate routers along the packet's path. These values may
be set by an ISP, and do not necessarily indicate the level of
reliability required by the end application. The link must be
configured with knowledge of the local meaning of the values.
ii. Tunnelling of traffic (e.g. GRE, MPLS, L2TP, IP-in-IP
encapsulation) can aggregate flows of different transport
classes, complicating individual flow classification with
schemes (b) and (c) and incurring further header processing if
tunnel contents are inspected.
iii. Use of the TCP/UDP port number makes assumptions about
application behaviour and requirements. New applications or
protocols can invalidate these assumptions, as can the use of
e.g. Network Address Port Translation, where port numbers are
remapped [RFC3022].
iv. In IPv6, locating the transport layer protocol type requires
parsing the entire IPv6 header, adding complexity to header
inspection. Again, this assumes that IPSec payload encryption
is not used.
Despite the difficulties in providing a framework for accurate flow
identification, approach (a) may be beneficial, and is preferable to
adding optimisations that are triggered by inspecting the contents
of specific IP packets.
Flow management is desirable; clear flow identification increases
the number of tools available for the link designer, and permits
more complex ARQ strategies that may otherwise make misassumptions
about traffic requirements and behaviour without flow
identification.
Links that are unable to distinguish clearly and safely between
delay-sensitive flows, e.g. real-time multimedia, DNS queries or
telnet, and delay-insensitive flows, e.g. bulk ftp transfers or
reliable multicast file transfer, cannot separate link service
classes safely. All flows should therefore experience the same link
behaviour.
In general, if separation of flows according to class is not
practicable, a low persistency is best for link ARQ.
PILC WORKING GROUP INTERNET DRAFT expires January 2002 18
Advice to link designers on link Automatic Repeat reQuest July 2001
4. CONCLUSIONS
A number of techniques may be used by link protocol designers to
counter the effects of errors / loss on links. One of these is
Automatic Repeat ReQuest, ARQ, which has been and continues to be
used on links that carry IP traffic. An ARQ protocol retransmits
link frames that have been corrupted on the physical link. Link ARQ
may significantly improve the probability of successful transmission
of IP packets over links prone to occasional loss.
A lower rate of packet loss generally benefits transport protocols
and endhost applications. Applications using TCP generally benefit
from Internet paths with little or no loss and low round trip path
delay. This reduces impact on applications, allows more rapid
growth of TCP's congestion window during slow start, and ensures
prompt reaction to end-to-end protocol exchanges (e.g.
retransmission, congestion indications). Applications using other
transport protocols, e.g. UDP or SCTP, also benefit from low loss
and timely delivery.
A side-effect of link ARQ is that link transit delay is increased
when frames are retransmitted. At low error rates, many of the
details of ARQ, such as degree of persistence or resulting out-of-
order delivery, become unimportant. Most frame losses will be
resolved in one or two retransmission attempts, and this is
generally unlikely to cause significant impact to e.g. TCP.
However, on shared high-delay links, the impact of ARQ on other UDP
or TCP flows may lead to unwanted jitter.
For links where error rates are highly variable, high ARQ
persistence may provide good performance for a single TCP flow.
However, interactions between flows can arise when many flows share
capacity on the same link. A link ARQ procedure that provides flow
management will be beneficial. Lower ARQ persistence may also have
merit, and is preferable for applications using UDP. The reasoning
here is that the link can perform useful work forwarding some
complete packets, and that blocking all flows by retransmitting the
frames of a single packet with high persistence is undesirable.
During a link outage, interactions between ARQ and multiple flows
are less significant; the ARQ protocol is likely to be equally
unsuccessful in retransmitting frames for all flows. High
persistence may benefit TCP flows, by enabling prompt recovery once
the channel is restored.
Low ARQ persistence is particularly useful where it is difficult or
impossible to classify traffic flows, and hence treat each flow
independently, and where the link capacity can accommodate a large
number of simultaneous flows.
PILC WORKING GROUP INTERNET DRAFT expires January 2002 19
Advice to link designers on link Automatic Repeat reQuest July 2001
Link ARQ designers should consider the implications of their design
on the wider Internet. Effects such as increased transit delay,
jitter, and re-ordering are cumulative when performed on multiple
links along an Internet path. It is therefore very hard to say how
many ARQ links may exist in series along an arbitrary Internet path
between endhosts, especially as the path taken and its links may
change over time.
In summary, when links cannot classify traffic flows and treat them
separately, low persistence is generally desirable; preserving
packet ordering is generally desirable.
Extremely high persistence and perfect persistence are generally
undesirable; highly-persistent ARQ is a bad idea unless flow
classification and detailed and accurate knowledge of flow
requirements make it possible to deploy high persistency where it
will be beneficial.
There is currently insufficient experience to recommend a specific
ARQ scheme for any class of link. It is also important to realise
that link ARQ is just one method of error recovery, and that other
complementary physical-layer techniques may be used instead of, or
together with, ARQ to improve overall link throughput for IP
traffic.
The choice of potential schemes includes adapting the data rate,
adapting the signal bandwidth, adapting the transmission power,
adaptive modulation, and adaptive information redundancy / forward
error control, and interleaving. All of these schemes can be used
to improve the received signal energy per bit, and hence reduce
error, frame loss and resulting packet loss rates given specific
physical channel conditions.
There is a need for more research to more clearly identify the
importance of and trade-offs between the above issues over various
types of link. It would be useful if researchers and implementers
clearly indicated the loss model, link capacity and characteristics,
link and end-to-end path delays, details of TCP, and the number (and
details) of flows sharing a link when describing their experiences.
In each case, it is recommended that specific details of the link
characteristics and mechanisms are also considered; solutions vary
with conditions.
5. SECURITY IMPLICATIONS
No security implications have been identified as directly impacting
IP traffic. However, an unreliable link service may adversely
impact some link-layer key management distribution protocols if
encryption is used over such a link.
PILC WORKING GROUP INTERNET DRAFT expires January 2002 20
Advice to link designers on link Automatic Repeat reQuest July 2001
Denial-of-service attacks exploiting the behaviour of the link
protocol, e.g. using knowledge of its retransmission behaviour and
propagation delay to cause a particular form of jamming, may be
specific to an individual link scenario.
6. IANA CONSIDERATIONS
No assignments from the IANA are required.
7. ACKNOWLEDGMENTS
Much of what is described here has been developed from a summary of
a subset of the discussions on the archived IETF PILC mailing list.
We thank the contributors to PILC for vigorous debate.
In particular, the authors would like to thank Spencer Dawkins,
Aaron Falk, Dan Grossman, Gary Kenwood, Reiner Ludwig and Jean
Tourrilhes for their detailed comments.
8. REFERENCES
References of the form RFCnnnn are Internet Request for Comments
(RFC) documents available online at http://www.rfc-editor.org/.
[BAL95] Balakrishnan, H., Seshan, S. and R. H. Katz, Improving
Reliable Transport and Handoff Performance in Cellular Wireless
Networks, ACM MOBICOM, Berkeley, 1995.
[BAL97] Balakrishnan, H., Padmanabhan, V. N., Seshan, S. and R. H.
Katz, A Comparison of Mechanisms for Improving TCP Performance over
Wireless Links, IEEE/ACM Transactions on Networking, 5(6), pp. 756-
759, 1997.
[BEN00] Bennett, J. C., Partridge, C. and N. Schectman, Packet
Reordering is Not Pathological Network Behaviour, IEEE/ACM
Transactions on Networking, 7(6), pp. 789-798, 2000.
[BHA97] Bhagwat, P., Bhattacharya, P, Krishna A. and S. K. Tripathi,
Using channel state dependent packet scheduling to improve TCP
throughput over wireless LANs, ACM/Baltzer Wireless Networks
Journal, (3)1, 1997.
PILC WORKING GROUP INTERNET DRAFT expires January 2002 21
Advice to link designers on link Automatic Repeat reQuest July 2001
[CHE00] Cheng, H S., G. Fairhurst et al., An Efficient Partial
Retransmission ARQ Strategy with Error Codes by Feedback Channel,
IEE Proceedings - Communications, (147)5, pp. 263-268, 2000.
[DRAFTDAW01] Dawkins, S., Montenegro, G., Kojo, M., Magret, V. and
N. Vaidya, End-to-end Performance Implications of Links with Errors,
draft-ietf-pilc-error-08.txt, to be published as a BCP RFC, 2001.
[DRAFTKARN01] Karn, P. (editor) et al., Advice for Internet
Subnetwork Designers, draft-ietf-pilc-link-design-nn.txt, work in
progress as internet-draft, 2001.
[DRAFTHAN01] Handley, M., Floyd, S. and J. Widmer, TCP Friendly Rate
Control (TFRC): Protocol Specification, draft-ietf-tsvwg-tfrc-nn
.txt, work in progress as internet-draft, 2001.
[ECK98] Eckhardt, D. A., and P. Steenkiste, Improving Wireless LAN
Performance via Adaptive Local Error Control, IEEE ICNP, 1998.
[FER99] A. Ferrero, The Eternal Ethernet, Addison-Wesley, 1999.
[ISO4335a] HDLC Procedures: Specification for Consolidation of
Elements of Procedures, ISO 4335 and AD/1, International
Standardization Organization, 1985.
[ISO4335b] HDLC Procedures: Elements of Procedures, Amendment 4:
Multi-Selective Reject Option, ISO 4335/4, International Standards
Organization, 1991.
[ISO7776] Specification for X.25 LAPB-Compatible DTE Data Link
Procedures, ISO 4335/4, International Standards Organization, 1985.
[KEN88] Kent, C. A. and J. C. Mogul, Fragmentation Considered
Harmful, Proceedings of ACM SIGCOMM, pp. 390-401, 1988.
[LIN93] Lin, S. and D. Costello, Error Control Coding: Fundamentals
and Applications, Prentice Hall, 1993.
[LUD99a] Ludwig, R., Rathonyi, B., Konrad, A., Oden, K., and A.
Joseph, Multi-Layer Tracing of TCP over a Reliable Wireless Link,
ACM SIGMETRICS, pp. 144-154, 1999.
[LUD99b] Ludwig, R., Konrad, A., Joseph, A. and R. Katz, Optimizing
the End-to-End Performance of Reliable Flows over Wireless Links,
ACM MobiCOM, 1999.
[MEY99] M. Meyer, TCP Performance over GPRS, IEEE WCNC, 1999.
[PAR00] Parsa, C. and J. J. Garcia-Luna-Aceves, Improving TCP
Performance over Wireless Networks at the Link Layer, Mobile
Networks and Applications, (1)5, pp. 57-71, 2000.
PILC WORKING GROUP INTERNET DRAFT expires January 2002 22
Advice to link designers on link Automatic Repeat reQuest July 2001
[RFC768] J. Postel, User Datagram Protocol, 1980.
[RFC791] J. Postel, Internet Protocol, RFC 791, 1981.
[RFC1122] R. Braden et al., Requirements for Internet Hosts --
Communication Layers, RFC 1122, 1989.
[RFC1191] Mogul, J. and S. Deering, Path MTU Discovery, RFC 1191,
1990.
[RFC1323] Jacobson, V. and R. Braden, TCP Extensions for High
Performance, RFC 1323, 1992.
[RFC1350] K. Solins, The TFTP Protocol (Revision 2), RFC 1350, 1992.
[RFC1435] S. Knowles, IESG Advice from Experience with Path MTU
Discovery, RFC 1435, 1993.
[RFC1827] R. Atkinson, IP Encapsulating Security Payload (ESP),
RFC 1827, 1995.
[RFC2475] Blake, S., Black, D., M. Carlson et al., An Architecture
for Differentiated Services, RFC 2475, 1998.
[RFC2488] Allman, M., Glover, D. and L. Sanchez, Enhancing TCP Over
Satellite Channels using Standard Mechanisms, RFC 2488, 1999.
[RFC2581] Allman, M., Paxson, V. and W. Stevens, TCP Congestion
Control, RFC 2581, 1999.
[RFC2757] Dawkins, S., Kojo, M., Magret V. and N. Vaidya, Long Thin
Networks, RFC 2757, 2000.
[RFC2760] Allman, M., Dawkins, S., Glover, D., J. Griner et al.,
Ongoing TCP Research Related to Satellites, RFC 2760, 1999.
[RFC2960] Stewart, R., Xie, Q., Morneault, K., Sharp, C.,
Schwarzbauer, H., Taylor, T., Rytina, I., Kalla, M., Zhang, L. and
V. Paxson, Stream Control Transmission Protocol, RFC 2960, 2000.
[RFC3022] Srisuresh, P., and K. Egevang, Traditional IP Network
Address Translator (Traditional NAT), RFC 3022, 2001.
[RFC3135] Border, J., Kojo, M., Griner, J., Montenegro, G. and Z.
Shelby, Performance Enhancing Proxies Intended to Mitigate Link-
Related Degradations, RFC 3135, 2001.
PILC WORKING GROUP INTERNET DRAFT expires January 2002 23
Advice to link designers on link Automatic Repeat reQuest July 2001
[SALT81] Saltzer, J. H., Reed, D. P. and D. Clark, End-to-End
Arguments in System Design, Second International Conference on
Distributed Computing Systems, pp. 509-512, 1981. Published with
minor changes in ACM Transactions in Computer Systems (2)4,
pp. 277-288, 1984.
[SAM96] Samaraweera, N. and G. Fairhurst, Robust Data Link Protocols
for Connection-less Service over Satellite Links, International
Journal of Satellite Communications, 14(5), pp. 427-437, 1996.
[SAM98] Samaraweera, N. and G. Fairhurst, Reinforcement of TCP/IP
Error Recovery for Wireless Communications, ACM CCR, 28(2), pp. 30-
38, 1998.
[STE94] W. R. Stevens, TCP/IP Illustrated, Volume 1, Addison-Wesley,
1994.
[STONE00] Stone, J. and C. Partridge, When the CRC and TCP Checksum
Disagree, Proceedings of SIGCOMM 2000, ACM Computer Communications
Review pp. 309-321, September 2000.
[WARD95] C. Ward et al., A Data Link Control Protocol for LEO
Satellite Networks Providing a Reliable Datagram Service, IEEE/ACM
Transactions on Networking, 3(1), 1995.
AUTHORS' ADDRESSES
Gorry Fairhurst (gorry@erg.abdn.ac.uk)
http://www.erg.abdn.ac.uk/users/gorry/
Department of Engineering, University of Aberdeen,
Aberdeen AB24 3UE, United Kingdom.
Lloyd Wood (lwood@cisco.com)
http://www.ee.surrey.ac.uk/Personal/L.Wood/
Cisco Systems UK Ltd, 4 The Square, Stockley Park,
Uxbridge UB11 1BY, United Kingdom.
PILC WORKING GROUP INTERNET DRAFT expires January 2002 24
Advice to link designers on link Automatic Repeat reQuest July 2001
FULL COPYRIGHT STATEMENT
Copyright (C) The Internet Society (2000-2001).
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.
PILC WORKING GROUP INTERNET DRAFT expires January 2002 25
Html markup produced by rfcmarkup 1.129b, available from
https://tools.ietf.org/tools/rfcmarkup/