Network Working Group P. Cameron
Request for Comments: 1692 Xylogics, International Ltd.
Category: Standards Track D. Crocker
Silicon Graphics, Inc.
D. Cohen
Myricom
J. Postel
ISI
August 1994
Transport Multiplexing Protocol (TMux)
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.
Abstract
One of the problems with the use of terminal servers is the large
number of small packets they can generate. Frequently, most of these
packets are destined for only one or two hosts. TMux is a protocol
which allows multiple short transport segments, independent of
application type, to be combined between a server and host pair.
Acknowledgments
This specification is the result of the merger of two documents: the
original TMux proposal which was the result of several discussions
and related initiatives through IETF working groups; and IEN 90 [1]
originally proposed by Danny Cohen and Jon Postel in May 1979.
Applicability Statement
The TMux protocol is intended to optimize the transmission of large
numbers of small data packets that are generated in situations where
many interactive Telnet and Rlogin sessions are connected to a few
hosts on the network. In these situations, TMux can improve both
network and host performance. TMux is not intended for multiplexing
long streams composed of large blocks of data that are typically
transmitted by sUCh applications as FTP.
The TMux protocol may be applicable to other situations where small
packets are generated, but this was not considered in the design.
The use of the TMux protocol in any other situation may require some
modification.
1. Introduction
When network designers consider which protocols generate the most
load, they naturally tend to consider protocols which transfer large
blocks of data (e.g., FTP, NFS). What is often not considered is the
load generated by Telnet and Rlogin because of the assumption that
users type slowly and the packets are very small. This is a grave
underestimation of the load on networks and hosts which have many
Telnet and Rlogin ports on multiple terminal servers.
The problem stems from the fact that the work a host must do to
process a 1-octet packet is very nearly as much as the work it must
do to process a 1500-octet packet. That is, it is the overhead of
processing a packet which consumes a host's resources, not the
processing of the data.
In particular, communication load is not measured only in bits per
seconds but also in packets per seconds, and in many situation the
latter is the true performance limit, not the former. The proposed
multiplexing is aimed at alleviating this situation.
If one assumes that most users connected to a terminal server will be
connecting to only a few hosts, then it should be obvious that the
network and host load could be greatly reduced if traffic from
multiple users, destined for the same host, could be sent in the same
packet.
TMux is designed to improve network utilization and reduce the
interrupt load on hosts which conduct multiple sessions involving
many short packets. It does this by multiplexing transport traffic
onto a single IP datagram [2], thereby resulting in fewer, larger
packets. TMux is highly constrained in its method of accomplishing
this task, seeking simplicity rather than sophistication.
2. Protocol Design
IP hosts may engage in the use of TMux transparently, and may even
switch back and forth between use of TMux and carriage of transport
segments in the usual, independent IP datagrams.
TMux operates by placing a set of transport segments into the same IP
datagram. Each segment is preceded by a TMux mini-header which
specifies the segment length and the actual segment transport
protocol. The receiving host demultiplexes the individual transport
segments and presents them to the transport layer as if they had been
received in the usual IP/transport packaging. The transport layer
is, therefore, unaware of the special encapsulation which was used.
Hence, a TMux message appears as:
IP hdr TM hdr Tport segment TM hdr Tport segment ...
Where:
TM hdr is a TMux mini-header and specifies the following
Tport segment.
Tport segment refers to the entire transport segment, including
transport headers.
The TMux Protocol is defined to allow the combining of transmission
units of different higher level protocols in one transmission unit of
a lower level protocol. Only segments with the same Internet Protocol
(IP) header, (with the possible exception of the protocol and check-
sum fields) may be combined. For example, the segment (H1, B1) and
the segment (H2, B2), where Hi and Bi are the headers and the bodies
of the segment, respectively, may be combined (multiplexed) only if
H=H1=H2. The combined TMux message is either (H, B1, B2) or (H, B2,
B1).
The receiver of this combined message should treat it as if the two
original segments, (H,B1), and (H,B2), arrived separately. It is
recommended, though not a requirement, that the segments in the TMux
message should be processed in the same order that they are in the
TMux message.
The multiplexing is achieved by combining the individual segments,
(H,B1) through (H,Bn), into a single message. This single message
has an IP header which is equal to H, but having in the PROTOCOL
field the value 18 which is the protocol number of the TMux protocol.
This IP header is followed by all the segments, B1 through Bn. Each
segment, Bi, is preceded by a 4 octet TMux mini header. This contains
the number of the protocol to which this segment is addressed. It
also contains the total length of this segment, including this mini
header. Since this mini header is not otherwise protected by a check-
sum, it also includes a checksum field which just covers this mini
header.
2.1. IP Protocol field value
TMux is indicated in an IP datagram by the Protocol (ID) value of 18
(22 octal), see [3].
2.2. Header Format
Each 4 octet TMux mini-header has the following general format:
+-------------------------------+
Length high
+-------------------------------+
Length low
+-------------------------------+
Protocol ID
+-------------------------------+
Checksum
+-------------------------------+
Transport segment
...
...
The LENGTH field specifies the octet count for this mini header and
the following transport segment, from 0-65535 octets. Hence, the
length field has a minimum value of 4. For segments that are larger
than the maximum allowed for TMux (see section 5.1), individual IP
datagrams should be sent.
The Protocol ID field contains the value that would normally have
been placed in the IP header Protocol field.
The 'Checksum' field is the XOR of the first 3 octets.
To ensure that TCP, UDP and other segments keep their 32 bit
alignment, where the segments being multiplexed are not a multiple of
32 bits long, extra octets will be added to re-align the end of the
segment, and hence the next segment. These octets will be ignored on
input. This padding will not affect the LENGTH field, it will still
contain the real length of the segment.
2.3. Sending Data
Host endpoints may choose to use TMux at any time and in either (or
both) directions. They also may switch back and forth between use of
TMux packaging and the usual individual IP datagrams for individual
transport associations. The only barrier to the use of TMux is for
the sender to know whether TMux is supported by the receiver. This
is important, since early use of TMux is likely to be limited.
The easiest way to detect TMUX support is to only send TMux messages
to hosts from which a valid TMux message has already been received.
This then leaves the problem of one host starting the TMux
connection. This is most easily accomplished by the host sending an
IP datagram with no data (i.e., with the IP total length field of
20), but with an IP Protocol field value of 18 for TMux. This is
referred to as a TMux ENQ (enquiry) message. The host receiving this
message then knows that the originator supports TMux, and can start
to send TMux messages. This will in turn cause the originator of the
ENQ message to start to use TMux. If for any reason the receiver
does not intend to send TMux messages to the originator, but is
prepared to accept them, then it can reply with another ENQ message.
If an ENQ message does not get a response, then it is reasonable to
resend the ENQ a while later in case the original ENQ message was
lost. If this again is lost, the ENQ may be repeated as often as
needed, but the time between requests should increase eXPonentially
up to a limit of about 1 hour. Suitable times between ENQs would be
15 seconds, 30 seconds, 60 seconds, 120 seconds etc.
Note that this checking process does not need to impede any of the
transport (user) data, which may be sent as convenient, albeit in its
less-efficient IP datagram form.
The only problem with this scheme is that a host which supports TMux
may stop supporting it, as might happen when the host is re-booted.
Other hosts need to learn of this change. The solution to this is to
maintain a Time To Live (TTL) value for hosts from which TMux
messages have been received. This TTL is a timed TTL, rather than a
count as used in the IP TTL field, and this time stamp is updated
every time a TMux message is received. This can then be used to
expire the information held by TMux on the host after a suitable
time, e.g., 1 minute.
This TTL time stamp is used as follows. When TMux is passed a segment
to be sent to a host, a check is made to see if the time to live has
expired. If the TTL has not expired, the segment is sent in a TMux
message as normal. If the TTL has expired, the host is marked as
being unable to TMux, but the segment is STILL sent as a TMux message
(i.e., with the normal delay to allow other segments to be
multiplexed). If the host is really unable to TMux anymore (a rare
occurrence) then this segment will be timed out and retried by the
transport provider i.e., TCP. Because the host was marked as not
able to TMux, the retry will be sent as a normal IP datagram. If the
remote host is still able to TMux then it should send back TMux
traffic (even if it has been rebooted), typically a TCP window
update, and the local host will mark it as able to TMux again. This
way of operating removes any performance problem caused by
continually dropping out of TMuxing and having to send probe
messages. If the IP datagram to be sent is from UDP, then the remote
host may not send anything in reply. So for UDP this scheme will not
be any better than just stopping sending TMux messages to the host,
but it is also no worse.
3. Protocol Behavior
3.1. Transport Flow Control
TMux operates as an extension to the IP datagram protocol. Hence, it
has no impact on most flow control mechanisms, since they operate at
the transport layer -- above TMux.
3.2. Connection Management
The concept of a connection pertains to certain transport protocols,
but not to IP or to TMux. Hence, when connection management is
required by a transport protocol using TMux, it occurs in the same
fashion as it does for IP. In fact, the transport protocol is not to
be aware that TMux is being used.
3.3 Multiplexed Message Construction
When a transport provider (e.g., TCP or UDP) sends a segment, TMux
first removes the IP header (if present) and adds a TMux mini-header
and the segment to the Multiplexed Message under construction for the
host specified by the destination address of the segment.
When the first message to be transmitted is placed into the
Multiplexed Message under construction, a timer is started. When the
timer expires, the Multiplexed Message under construction is
transmitted. This ensures that all segments available for sending
before the timer expires are sent in a single Multiplexed Message.
If, during construction of the Multiplexed Message, the buffer
holding the message fills, the Multiplexed Message is transmitted
immediately.
The delay time should be user configurable; a reasonable time is 20
to 30 milliseconds. The time period should be large enough to give a
reasonable probability of sending multiple segments but not so large
that the echo response time becomes a problem. This suggests that
the upper limit for the timer is probably 1/10th second. As the cost
of using timeouts on many systems is quite large, it is recommended
that a single timer be used and that all TMux messages under
construction are sent when the timer expires.
Additionally, configuration options may limit the number of included
data segments or the maximum size of the Multiplexed Message before
it is transmitted. It is also suggested that larger segments (e.g.,
those over 700 octets) should be sent as standard IP datagrams, and
not multiplexed. This is to ensure that the delay caused by the TMux
timer does not put a delay on those segments for which it is
inadvisable. The size of the largest segments to be multiplexed
should (if possible) be configurable.
4. Protocol Example
This example shows a TMux message consisting of three multiplexed
segments:
A TCP segment consisting of a 20 octet TCP header, 5 octets of data
and 3 octets of padding. Thus the length field is
Mini header + TCP header + data
= 4 + 20 + 5
= 29
The padding is NOT included in the length.
A TCP segment consisting of a 20 octet TCP header, 4 octets of data.
This segment does not require padding.
A UDP segment consisting of a 4 octet UDP header, 41 octets of data
and 3 octets of padding.
+-------------------------------+
Length = 29
(2 octets)
+-------------------------------+
Protocol ID = 6 (TCP)
+-------------------------------+
Checksum
+-------------------------------+
TCP Header
(20 octets)
+-------------------------------+
TCP data
(5 octets)
+-------------------------------+
Padding
(3 octets)
+-------------------------------+
Length = 28
(2 octets)
+-------------------------------+
Protocol ID = 6 (TCP)
+-------------------------------+
Checksum
+-------------------------------+
TCP Header
(20 octets)
+-------------------------------+
TCP data
(4 octets)
+-------------------------------+
Length = 49
(2 octets)
+-------------------------------+
Protocol ID = 17 (UDP)
+-------------------------------+
Checksum
+-------------------------------+
UDP Header
(4 octets)
+-------------------------------+
UDP data
(41 octets)
+-------------------------------+
Padding
(3 octets)
+-------------------------------+
5. Implementation Suggestion
5.1 Maximum TMux Message Size
In section 3.3, a note is made about sending messages immediately if
the limit on TMux message size is reached. On systems where Path MTU
Discovery (as per RFC1191 [4]) has been implemented this should be
used to discover the maximum message size that can be transmitted,
and this should be used as the maximum TMux message size.
5.2 Deciding Which Segments to Multiplex
It is the responsibility of the sender to decide which segments
should be TMux'd and which should not. For example, segments sent by
FTP should not normally be multiplexed. In many situations, it may
be sensible to restrict the sessions that can be multiplexed to just
those involved in interactive traffic (Telnet and Rlogin) by
examining the source and destination TCP port numbers. However, if a
segment that would not normally be multiplexed is to be sent and a
TMux message is already under construction, then the extra segment
can be added to the TMux message under construction, and this
complete message should be sent immediately, rather than waiting for
the timer to expire.
6. Implementation notes
The following notes are the result of experience gained during the
testing of early implementations of TMux. Whilst they do not form
part of the actual standard, they should be followed if possible to
ensure compatibility with other implementations.
Because the TMux mini-header does not contain a TOS field, only
segments with the same IP TOS field should be contained in a single
TMux message. As most systems do not use the TOS feature, this is
not a major restriction. Where the TOS field is used, it may be
desirable to hold several messages under construction for a host, one
for each TOS value.
Segments containing IP options should not be multiplexed.
Only unicast addresses should be considered for multiplexing.
Segments addressed to the loopback address (127.0.0.1) are not
candidates for multiplexing.
Only segments with a source or destination port that is for an
interactive session (i.e., Telnet and Rlogin) should be considered
for multiplexing using TMux.
If an error is discovered in a checksum of a TMux header, the rest of
the message, starting there, is ignored. If an unknown PROTOCOL
field is discovered in any TMux header, this segment, and only this
one, is ignored.
If the TMux implementation is continually sending TMux messages
containing exactly one segment (because is there is little traffic to
multiplex), then TMux may be turned off. This implies that TMux may
be switched off when there is no congestion.
To prevent intermediate nodes from fragmenting and reconstructing
TMux frames, implementations may want to set the "do not fragment"
flag in the IP datagram of TMux messages.
If host B receives a TMux ENQ message from host A, but does not have
any data for host A, then it may also send back an ENQ message.
However, host A may send another ENQ message in response to this, so
causing B to respond and so on. Thus if this facility is used, code
must be included to prevent this looping behavior happening. Sending
an ENQ in response to an ENQ is not recommended, except in special
circumstances.
It is recommended that the following ASPects of the TMux protocol be
user configurable:
The maximum size of a segment that can be multiplexed by TMux.
The delay between the first segment being placed into the message
under construction and the message being sent.
7. Security Considerations
Because TMux is effectively an extension to IP, it does not have any
more impact on site security than does IP. Security should be dealt
with by upper layer protocols.
Because some routers filter packets on the TCP port numbers, any
segments sent using TMux will not be subject to this filtering as it
will obscure the TCP port number However, larger segments for the
same TCP connection will still be sent as IP datagrams, and so will
be subject to filtering, thus giving rise to a potential problem.
For this reason, any routers that do not support TMux, but which do
support this type of filtering should not allow TMux messages through
(in either direction). This will cause both hosts to think the other
does not support TMux, so all segments will be sent as IP datagrams,
thus eliminating this problem.
A better solution to this problem, is for routers to understand the
TMux protocol, and to inspect each of the multiplexed segments and
remove those segments that fail the filtering.
8. References
[1] Cohen, D., and Postel, J., "Multiplexing Protocol", IEN 90,
USC/Information Sciences Institute,, May 1979.
[2] Postel, J., "Internet Protocol", STD 5, RFC791, USC/Information
Sciences Institute, September 1981.
[3] Reynolds, J. and J. Postel, "Assigned Numbers", STD 2, RFC1340,
USC/Information Sciences Institute, March 1990.
[4] Mogul, J., and S. Deering, "Path MTU discovery", RFC1191, DECWRL
and Stanford University, November 1990.
9. Authors' Addresses
Peter Cameron
Xylogics International, Ltd.
Featherstone Rd.
Wolverton Mill
Milton Keynes
MK12 5RD
United Kingdom
Phone: +44 908 222112
Fax: +44 908 222115
EMail: cameron@xylint.co.uk
David Crocker
Silicon Graphics, Inc.
2011 N. Shoreline Blvd.
P.O. Box 7311
Mountain View, CA 94039-7311
USA
Phone: +1 415 390 1804
Fax: +1 415 962 8404
EMail: dcrocker@sgi.com
Danny Cohen
Myricom
325 N. Santa Anita Ave.
Arcadia, CA 91006
USA
Phone: +1 818 821 5555
EMail: Cohen@myricom.com
Jon Postel
USC/Information Sciences Institute
4676 Admiralty Way
Marina del Rey, CA 90292-6695
USA
Phone: +1 310 822 1511
Fax: +1 310 823 6714
EMail: Postel@ISI.EDU
10. Discussion List
There is a discussion list for this protocol, which for
historical reasons is called:
cmp-id@xylint.co.uk
Requests to join the list should be sent to:
