This article will appear at the IEEE 19th Conference on Local Computer
Networks, October 1994 in Minneapolis, Minnesota. Re-printed with the
permission of the authors.

========================================================================

EMERGING TRENDS - FULL-DUPLEX AND THE SWITCHED LAN


Ken Christensen, Franc Noel, and Norm Strole

IBM Corporation
Networking Hardware Division
Research Triangle Park, NC 27709


ABSTRACT

Ethernet, Token-Ring, and FDDI are well established protocols for
regulating the access to a common transmission medium among a large number
of stations.  All of the attached stations share this common medium and its
bandwidth.  Sharing of a common medium reduces the bandwidth that is
available to an individual station.  To increase bandwidth to the end user,
emerging LAN topologies are departing from shared-media, shared-bandwidth
methods in favor of dedicated-media and dedicated-bandwidth methods.
Dedicated-bandwidth switched LANs can take advantage of full-duplex
operation of attached stations.  This is in contrast to the normal
half-duplex operation of LAN stations on a shared-bandwidth LAN.  This
paper describes the evolution of shared-media, shared-bandwidth LANs into
dedicated-media, dedicated-bandwidth switched LANs with full-duplex
operation of the attached stations.


1.0 INTRODUCTION

Early LANs were typically small - both in network span and in the number of
stations.  In these early LANs the attached stations were much less
powerful than today's Personal Computers and workstations.  As a result,
there were few concerns over bandwidth requirements.  These early LANs were
typified by Ethernet 10Base5 and 10Base2 and can be categorized as having
shared media and shared bandwidth.  With bandwidth being shared on a common
medium, the LAN adapters were half-duplex in operation.  Figure 1 shows an
Ethernet 10Base5 LAN segment where multiple stations share a common
transmission medium and its bandwidth.

During the 1980's, LAN data rates increased from 2 Mbps for IBM's PC-Net
LAN up to 100 Mbps for FDDI.  The most popular LANs during this time were
10-Mbps Ethernet and 4- and 16-Mbps Token-Ring.  Cable types evolved from
coax (for example, Ethernet 10Base5 and 3270 display systems) towards
shielded and then unshielded twisted-pair cables.  Optical fiber cabling
was also installed in some locations to allow for future very high-speed
technologies.  As the strategic importance of LANs grew they became larger.
This increased importance, along with growing station count and increased
geographic coverage, encouraged the growth of systematic cabling schemes.
The IBM Cabling System and the EIA/TIA-568 Commercial Building
Telecommunications Cabling Standard (see [1]) describe systematic
star-wired cabling systems for multiproduct and multivendor environments.
Ethernet 10BaseT and Token-Ring are examples of LANs that often employ a
wiring closet for concentrating each star-wired LAN segment.  Star-wiring
simplifies problem determination on a LAN.  A wiring closet provides one
place where all station attachments can be accessed.  Thus, a defective
station or cable, once identified, can easily be removed and repaired
without the need to literally "walk the wire."  Figure 2 (on the next
page) shows a star-wired Token-Ring LAN segment with a central wiring
closet.


         Figure 1 - Shared-media, shared-bandwidth Ethernet


In this paper, section 2 describes the evolution from shared media to
dedicated media LANs.  Section 3 describes the next step - the evolution to
dedicated band- width LANs.  Section 4 describes full-duplex operation of a
LAN adapter and reviews proposals made to IEEE 802.5 for full-duplex
operation of Token-Ring stations.  Section 5 describes performance
characteristics of full-duplex dedicated-bandwidth LANs, discusses
motivating applications, and briefly discusses migration and management.
Finally, section 6 is a summary.


         Figure 2 - A star-wired Token-Ring LAN segment


2.0 SHARED MEDIA TO DEDICATED MEDIA

Initially, wiring closets contained little intelligence - simple repeaters
for Ethernet or unpowered concentrators containing insertion relays for
Token-Ring.  On these shared-media and shared-bandwidth LANs, any cable or
station failure would likely affect all stations on the same LAN segment.
For example, forgetting to install terminators on both ends of an Ethernet
10Base2 segment would cause the entire segment to be inoperable.

With the continued growth and business importance of LANs and the
decreasing cost of electronics, wiring closets started to become
"intelligent".  Intelligent repeaters for Ethernet and intelligent
concentrators for Token-Ring are capable of automatically isolating and
removing failed stations.  For example, an intelligent Token-Ring
concentrator can prevent a 4-Mbps station from inserting into a 16-Mbps LAN
segment.  This minimizes outages and increases the reliability of a LAN.
LANs built with intelligent wiring closets can be classified as dedicated
media and shared bandwidth.  Bandwidth is still shared among all users,
however each station's media is "dedicated."  Because of the star-wired
cabling topology, station failures are most often isolated to a single
station and its media.  Intelligent repeaters and concentrators can also
contain additional features such as remote network monitoring and reporting
mechanisms, connection between different media types, data security
facilities, and LAN bridging.  Figure 3 shows a five-port Token-Ring
intelligent concentrator.  The active logic shown in the figure
automatically removes failing stations from the main ring so, for example,
a station attempting to insert at a data rate different from the main ring
data rate will not be permitted to join the network.


         Figure 3 - A Token-Ring intelligent concentrator


3.0 THE NEXT STEP - DEDICATED BANDWIDTH

Shared bandwidth with dedicated media increases the manageability and
reliability of LAN segments, but does not increase the amount of bandwidth
available per station.  In a shared-bandwidth LAN all transmitted frames
pass by all stations, even those frames not destined for a particular
station.  LAN bridges or port-switching hubs can be used to partition large
LAN segments into smaller interconnected segments.A port-switching hub
contains multiple backplanes, each corresponding to a LAN segment.
Compared to a single large LAN segment, average station bandwidth may be
increased by localizing stations that primarily communicate between each
other on the same (smaller) LAN segment.

By using parallel switching techniques in an intelligent concentrator, the
amount of bandwidth per station is increased.  For example, a cross-bar
switch can allow dedicated frame-by-frame connections between switch ports.
When a frame is received at a port on a switch, it is automatically
switched "on the fly" to the appropriate port for output based on the
destination address contained in the frame.  A table at each switch port
determines to which port a given destination address should be switched.
Frames addressed to the all-stations address (that is, broadcast frames)
are sent out on all ports.  Thus, each LAN segment attached to a switch
port receives only the frames with destination addresses on that LAN
segment.  A switch port can also contain frame buffering to temporarily
"hold" frames that cannot be immediately forwarded.  For example, in an
Ethernet LAN if the destination Ethernet segment is currently busy, a
switched frame would be temporarily stored in the switch port buffers.
Figure 4 shows a four-port cross-bar LAN switch.  Each switch port contains
a Media Access Control (MAC) unit to support attachment to a LAN segment.
Attached to each switch port can be a dedicated-media single station or
shared-media LAN segment with multiple stations.


         Figure 4 - A four-port cross-bar switch


Detailed descriptions of LAN switch architectures are beyond the intended
scope of this paper.  Manufacturer's sales literature from references[4]
and [6] can be consulted.  In[3] the evolution of 10-Mbps "private
Ethernet", then 100-Mbps dedicated networking, and finally 155-Mbps
Asynchronous Transfer Mode (ATM) is described.

4.0 FULL-DUPLEX OPERATION

All LAN adapters support a MAC protocol.  The purpose of the protocol is to
regulate access to a common media (for example, via tokens as in
Token-Ring).  In a switched LAN with a single link per station, the
requirements for a MAC protocol are greatly minimized.  No longer must a
station "contend" with other stations for permission to transmit onto the
media.  Thus, in a switched LAN, stations can transmit whenever a frame is
queued in the adapter.A station can also receive at any time.  Full-duplex
operation is now possible - a station can transmit and receive
independently of other stations on the switched LAN.

Full-duplex operation has been proposed for Ethernet in IEEE 802.3 (see[2]
) and for Token-Ring in IEEE 802.5 (see[5] ). Full-duplex offers the
following benefits over half-duplex operation:

  ­ Increases the bandwidth of switch-attached single station LAN segments.

  ­ Provides migration to higher bandwidths as existing and emerging
    applications require additional bandwidth.

  ­ Capitalizes on emerging LAN switching technology.

  ­ Positions the desktop Personal Computer or workstation for the future
    high-speed enterprise Asynchronous Transfer Mode (ATM) backbone.

  ­ Preserves the investment in existing LAN adapters and infrastructure
    (for example, cabling, software applications, and network management
    platforms and protocols).

4.1 Full-Duplex Token-Ring

The Token-Ring MAC protocol is usually executed within a Token-Ring
adapter.  When operating in standard Token-Ring mode, an adapter normally
transmits a frame or receives a frame, but not both simultaneously.  Figure
5 shows the normal mode of operation of a Token-Ring adapter.  The figure
shows the repeat path for all frames, whether received or not, within an
adapter.  The repeat path is required in half-duplex Token-Ring operation
to allow each transmitted frame to circulate around the entire ring.


         Figure 5 - Normal operation of a Token-Ring adapter


If the repeat path is removed and other changes made that are discussed
later, a Token-Ring adapter can transmit a frame on the outgoing link while
also receiving a frame on the inbound link, see Figure 6 (on the next
page).  This is full-duplex operation.  In this full-duplex mode, frames
are transmitted on the outgoing link immediately upon being queued for
transmission at the MAC interface.  Depending upon the capabilities of the
adapter, its station, and the switch to which it is attached, sustained
transmit rates of 16 Mbps are possible.  Even more significant, the adapter
can also simultaneously receive frames at the 16 Mbps rate on the receive
link, for an aggregate capacity of 32 Mbps.  Typically, LAN applications
(for example, client/server) will generate traffic in bursts rather than in
continuous streams.  Thus, the achievable data rate of an aggregate 32 Mbps
will be required in bursts, rather than in a continuous mode of operation.

When in full-duplex mode, a station does not employ token protocols,
including:

  ­ Token access and token recovery protocols

  ­ Frame repeating, frame stripping, and end-of-frame status update

A key aspect of full-duplex operation is that the Token-Ring frame format
is preserved without change.  The Frame Control field, priority bits,
destination and source address fields, and source routing subfields are all
maintained.  Thus, existing IEEE802.5 end-user applications software and
network infrastructure (for example, bridges) are preserved making the
full-duplex mode of operation transparent to the user.


         Figure 6 - Full-duplex Token-Ring adapter


A single Token-Ring adapter may support both normal (half-duplex)
Token-Ring and full-duplex operation.  There are two logical connectivity
options that an adapter should satisfy:

  1. Activate in full-duplex mode when attached to a compatible full-duplex
     switch port.

  2. Otherwise, activate in Token-Ring mode when attached to a standard
     Token-Ring concentrator or switch port that does not support
     full-duplex operation.


A station cannot physically determine whether it is attached to a
half-duplex Token-Ring concentrator or to a full-duplex switch.  This
determination must be accomplished via a logical protocol during the
activation procedure.  Figure 7 shows an example activation sequence for an
adapter to a switch or concentrator port.  In state S2 the normal
Token-Ring insertion procedure of internal diagnostics and lobe media test
is executed.  State S3 advertises, for example via a MAC frame, the
capability of the adapter to operate in full-duplex mode.  If the port is
capable of full-duplex operation, transition is made to state S6
(full-duplex operation), otherwise transition is made to state S5 (normal
Token-Ring operation).

Following activation for full-duplex mode of operation, the token access
protocols can be suspended, while other functions, such as those used to
detect and report errors, continue to be active.  When in full-duplex mode,
each station or switch port must transmit using the local crystal within
that station since there is no longer an active monitor providing clocking.
The existing differential Manchester encoding scheme is maintained.
Stations receive incoming signal via the acquired clock of the incoming
link.  When there are no frames to transmit, a station transmits (and
expects to receive) a continuous idle stream (for example, all zeros).
This idle stream provides a continuous signal to the receiving port for
clock synchronization and serves as a means of detecting and/or reporting a
link failure or disconnection.


         Figure 7 - Example activation sequence


All frames can be transmitted immediately upon proper setup to the MAC
interface without having to wait for token access.  A maximum frame length
of 4500 bytes is recommended for full-duplex operation, primarily to both
reduce buffer requirements in the LAN switch and end-to-end frame transit
delays due to blocking or store-and-forward queueing delays.  A station may
maintain multiple transmit queues in order to prioritize local access to
the full-duplex link.  However, first-in first-out frame order within a
given priority must be maintained within a station or LAN switch.  The
low-order three bits of the Token-Ring Frame Control field have been
designated as the priority bits of a given frame.  These bits may be used
for priority transmit queueing.A full-duplex station will receive all
frames that are properly formatted and contain the appropriate destination
address.

When in full-duplex mode, the majority of the normal Token-Ring MAC level
protocols are no longer required and can thus be suspended.  Since many of
these MAC protocols depend upon specific MAC frames, stations must not
transmit these frames, nor should LAN switches forward these frames, when
in full-duplex mode.  The Active Monitor and Standby Monitor Present MAC
frames could be used to provide a periodic "heartbeat" between a station
and a switch port on a full-duplex link to indicate a logical presence to
each other.  When in full-duplex mode, a switch port can learn the address
of its attaching adapter during the activation protocol.  This is
particularly applicable to mapping a station address to a physical port on
an interconnect switch.  Beacon MAC frames are still applicable in
full-duplex mode to indicate temporary or permanent hard errors.  Report
Soft Error MAC frames can also still be used, but some of the error
conditions that are reportable in half-duplex Token-Ring mode do not apply
to full-duplex mode.

Full-duplex operation should be less sensitive to many of the disruptions
that can occur in normal Token-Ring operation.  For example, all errors
associated with maintaining token flow are no longer a concern.  Also, link
errors no longer affect data on the entire ring, but only on an isolated
segment.  However, there can still be situations where the first indication
of a problem is the appearance of soft errors in data frames.  Errors in
received frames may be indicative of a poor transmission link or of
external electromagnetic interference.  These errors will normally appear
as Frame Check Sequence (FCS) errors.  Improperly formatted or incomplete
frames must also be handled.  A frame that exceeds the maximum allowed
frame size for a full-duplex link could be handled as an improperly
formatted frame.  In general, the receiving station is responsible for
detecting frame errors.  Frames that exceed the maximum frame length can be
detected via the activation of a valid-frame timer (for example, 2.25
milliseconds for a maximum length 4500 byte frame at 16 Mbps) at the
beginning of frame reception.  If a valid End Delimiter is not detected
before the timer expires, the frame reception is terminated.  The detection
of an Abort Delimiter by a LAN switch during the forwarding of a frame can
also immediately terminate frame forwarding.

5.0 PERFORMANCE CHARACTERISTICS

This section describes the performance characteristics of a full-duplex
switched LAN.  In existing LANs the total bandwidth of a common medium is
shared between all attached stations.  For example, on a 16-Mbps Token-Ring
with 50 stations, each station can expect an average bandwidth of 16 Mbps /
50 = 320 Kbps.  Of course, the peak bandwidth available to a station is the
full 16 Mbps assuming that no other stations are transmitting data onto the
LAN.  Similar per-station bandwidth degradation is typical of other LANs,
such as 10-Mbps Ethernet and 100-Mbps FDDI.  If a single station is
attached to a switch port, then that single station has the full, or
dedicated, bandwidth of the switch port.  In a shared-bandwidth LAN
segment, the addition of each new station decreases the average bandwidth
available for every station.  For a dedicated bandwidth LAN, the addition
of a new station does not imply less bandwidth for every station.  Figure 8
is a graph showing the average bandwidth available to stations on shared
and dedicated-bandwidth LANs.  This graph shows a uniform 100% load for
shared (half-duplex) and full-duplex Token-Ring.  This graph assumes a
maximum achievable shared (half-duplex) Ethernet utilization of 80%.


         Figure 8 - Shared and dedicated bandwidth performance


Access delay is the length of time a station must wait for a queued frame
to be transmitted.  On a shared-bandwidth LAN the access delay varies
according to how much traffic is on the LAN.  For a Token-Ring LAN the
access delay is the amount of time a station waits for a token.  This token
wait time varies according to how many other upstream stations also have
data ready to transmit onto the LAN.  For a Token-Ring LAN segment with
very low utilization levels, token wait time is negligible.  For a
Token-Ring LAN segment with higher utilization levels, token wait times
would rarely exceed several milliseconds.  With a full-duplex link to a
switch, a station has dedicated transmit bandwidth.A full-duplex link
yields a full 16 Mbps for transmit and a simultaneous 16 Mbps for receive.
Thus, the access delay for transmitting a frame is effectively zero since
there is no token wait delay.  Some frame queueing delays may, however, be
experienced within the LAN switch due to the bursty nature of the traffic
or the load distribution among the various stations.

Table 1 shows the bandwidth characteristics of a 16-Mbps shared-bandwidth
LAN and a 16-Mbps full-duplex dedicated-bandwidth switched LAN.  Table 1
also shows access delay where the value 2.25 milliseconds represents the
delay caused by the transmission of a 4500 byte frame at 16 Mbps.  Each LAN
is assumed to have N stations attached.  The minimum bandwidth is the worst
case bandwidth that a single station can expect to obtain over a period of
time.  The peak bandwidth is the maximum instantaneous bandwidth a station
can achieve.  Aggregate bandwidth is the total bandwidth available in the
LAN.  Finally, access delay is the time a station must wait before it can
transmit a frame.  For maximum full-duplex performance, the aggregate
bandwidth of the switch must be able to handle peak load situations.  For
example, a peak load situation occurs when half of the switch ports are
sending frames to the other half while also receiving frames from the other
half.  In other words, a peak load situation occurs when all switch ports
are busy in simultaneous full-duplex transmit and receive of frames.


         Table 1 - Performance of normal and full-duplex


5.1 Applications Motivating Switched LANs

Two applications that can benefit from a full-duplex switched LAN are
client/server computing and multimedia.  In client/server computing a
powerful dedicated server is used to store common data and programs
accessed by multiple client stations.  Multimedia encompasses a large
number of applications that typically combine text, images, audio, and
video.  Multimedia applications can operate as either client/server or
client/client.  Delivery on demand of video clips is an example of a
client/server multimedia application.  Video conferencing is an example of
a client/client multimedia application.

For client/server computing the server can be a performance bottleneck.
That is, multiple clients accessing a single server can overwhelm the
processing capabilities or network bandwidth of the server.  If network
bandwidth is the bottleneck, then putting the server on a dedicated
full-duplex link can improve overall performance.  Figure 9 shows a
switched LAN with two full-duplex file servers.  The client stations are on
shared-bandwidth LAN "micro-segments" attached to the same switch.  A
micro-segment is a LAN segment with very few stations (for example, five to
six stations) as compared to a typical LAN segment of thirty to fifty
stations.  For this configuration, the two file servers have up to four
times the available bandwidth than if they and all the client stations were
attached to a single shared-bandwidth LAN.  With full-duplex and dedicated
bandwidth it is possible to improve performance for client stations by
changing only the location of the file server(s) via direct attachment to a
LAN switch.  On a shared-bandwidth LAN, to improve the data throughput of a
single station (for example, a file server), all stations have to be
upgraded, at a high cost, to a faster LAN adapter.  Cabling infrastructure
may also have to be changed.  An example of this type of upgrade would be
changing from a 10-Mbps Ethernet LAN to a 100-Mbps FDDI LAN.


         Figure 9 - Switched LAN with full-duplex file servers


Multimedia applications that contain real-time audio and video require
guaranteed bandwidth and, in some cases, non-varying access delay.  Data
traffic does not, in general, require guaranteed bandwidth.  Response times
for file transfers, for example, simply take longer when less bandwidth is
available.  However, without guaranteed bandwidth, real-time traffic such
as video will suffer in quality.  This reduced quality will be observed by
a user as pauses in both video and audio.  Typically, full-motion video
requires several megabits per second of bandwidth for each video stream.
If many users on a LAN desire to view independent video streams, a large
amount of aggregate bandwidth is needed.  Full-duplex and switching can
provide the needed bandwidth to enable many emerging multimedia
applications on a LAN.

5.2 Migration and Management of Switched LANs

The evolution to LAN switching allows existing LAN adapters and
infrastructure to be migrated towards higher bandwidths.  The first step in
this migration is to isolate the few high-bandwidth stations (for example,
file servers) to dedicated full-duplex links.  Stations with lower
bandwidth requirements can remain on larger LAN segments.  As client
bandwidth demands increase, the first step towards dedicated-bandwidth can
be micro-segmentation.  In this case, each LAN switch port has a LAN
segment consisting of more than one station attached to it.  The final step
is individually attaching client stations directly to a LAN switch.

By keeping existing LAN adapters, the existing network infrastructure can
also remain largely intact.  This includes existing cabling layouts,
network management platforms and protocols, and software applica tions.
This then lays the groundwork for attaching to high-speed backbones such as
ATM.

6.0 SUMMARY

This paper has described the evolution of small shared-media LANs into
large LANs with dedicated media.  This first evolutionary step had
significant benefits in reliability.  The second and emerging step is
towards dedicated media and dedicated bandwidth.  Migration to
dedicated-bandwidth switched LANs is motivated by demands for increased
bandwidth and improved management.  Switched LANs enable full-duplex
operation of LAN stations.  With switched LANs the existing install base of
LAN adapters and infrastructure can be incrementally migrated (from
shared-bandwidth to dedicated-bandwidth) for increased client/server
traffic demands and many future multimedia applications.

Acknowledgments

The authors would like to acknowledge Mike Siegel, Tom Toher, Ray Zeisz
(all of Networking Hardware Division, IBM Corporation), and many others for
their contributions to full-duplex Token-Ring and this paper.

  REFERENCES

  [1] Commercial Building Telecommunications Wiring Standard, EIA/TIA 568 ,
      Electronics Industries Association, July 1991.

  [2] Cullerot, D. "Full Duplex Switched Ethernet (FDSE)."  IEEE 802.3
      Plenary , November 1992.

  [3] McQuillan, J. "The End of Shared-Medium LANs."  Business
      Communications Review , October 1992.

  [4] "MultiNet - The Future is Here (Video-Ready Switching LAN Hubs)."
      Sales literature from LANNET Data Communications, Inc., Huntington
      Beach, California, 1993.

  [5] Strole, N. "Full-Duplex Option for 16-Mbps Token-Ring Stations."
      IEEE 802.5 High Perform- ance Architecture Study Group , September
      1993.

  [6] "The Full Duplex Switching Advantage."  Sales literature from
      Kalpana, Sunnyvale, California, 1993.

                                           ­­­­­­­ãí­­­­­­¨­­ÊËÈÈ­ÍÎÏ­[