The Fundamentals of Asynchronous Transfer Mode

Peter W. Owings
December 10, 1994

Final Paper
 MSCS: Introduction to Telecommunications
Prof. C. C. Lee


1.0 Introduction

The whole point of this paper is to try to get a handle on ATM, try to figure out how it works, 
and at least have some sort of perspective with respect to implementing this technology at 
Northwestern University.

The majority of this paper is devoted to trying to understand ATM conceptually and to some 
degree of detail.  The current state of this technology ultimately is revealed the closer we 
examine the details.

At the end of this paper I look at the existing network at Northwestern University, and the 
foreseeable demands on the network infrastructure, and try to give some sort of perspective on 
whether or not ATM should be implemented at the university, to what extent, how it fits in 
with existing network schemes, and where it fits in the future.

2.0  Understanding ATM

The purpose of this first part of this paper is to try to understand the details of ATM.  
Therefore these next several pages are an attempt by me to define ATM, describe it and possibly 
as an end result of this exercise hopefully better understand it.

B-ISDN Principles

The ultimate goal is to provide a broadband transport mechanism which provides for a few 
basic networking needs.  First, the goal is to create an universal network scheme which supports 
both wide-area and local-area networking needs.  The adopted scheme should support a large 
range of services with widely varying demands for bandwidth.  B-ISDN is designed to provide 
subscriber communication services over a wide range of bit rates.  Ideally what is desired is a 
service-independent network which is flexible and will be scalable as applications change and 
demands grow.  The network would also be efficient in its use of resources such that all services 
are able to share all available resources.  This allows for an optimization of statistical sharing 
of the provided network resources.  A system which allows all services to access all available 
resources should produce lower costs for the overall operation of the network.

Originally the most efficient and flexible method of transporting information was based on 
packet switching.  Unfortunately the level of complexity associated with packet switching 
does not allow existing technologies to scale well at higher speeds.  In addition, packet 
switching methods cannot guarantee the level of support needed by time-dependent 
applications.  If traffic is heavy on a packet switching network, time sensitive packets may be 
delayed.

Therefore a new network paradigm was required to satisfy the needs of a broadband services.  A 
few basic principles were identified around which the new system was designed.  First, 
functions must not be repeated in the network several times if service can be guaranteed when 
functions are only implemented once at the network's boundary.  Second, correct delivery of 
transmitted information (semantic transparency), and the continuity of the bit stream (time 
transparency) would be guaranteed.

In 1987, the CCITT selected ATM to be the underlying telecommunication service for Broadband 
Integrated Services Digital Network (B-ISDN) as the switching technology for future 
telephone and data networks.  ATM needs minimal network node functionality, allowing for 
very high speeds and is application independent.

ATM  Overview

ATM and the Telephone Companies

Since its adoption in 1987, ATM has been and is still being developed by the CCITT as part of B-
ISDN.   The ATM Forum is a group of users and vendors attempting to speed up the process of 
creating standards defining ATM.  Currently ATM supports up to 622 Mb/s data rates but ATM is 
theoretically capable of Gigabit speeds.

ATM is being defined to support telephony.  A large amount of data communication takes place 
inside of the telephone networks to provide the necessary communications between the 
telephone networks' computers.  This data traffic is carried over separate data networks that 
run parallel with the existing voice networks.  The phone companies would like to have both 
this data and voice traffic transmitted over the same network.

At the same time the implementation of fiber has led to the consolidation of transmission lines.  
Because of this multiplexing over the same links is both difficult and expensive.  It is believed 
that at high speeds it will be easier to multiplex separately addressed packets then to 
multiplex mixes of individual bits.

Therefore the telephone companies are designing ATM to meet internal needs and also to give 
them the ability to offer new services to customers.  Because these are the primary goals of the 
telephone companies, the needs of the data communications community are not necessarily being 
addressed.  It is clear that ATM will be implemented by the phone company; how it will 
support the needs of the data communications community is not as clearly defined.

Fundamental Principles of ATM

The principles behind ATM were born out of packet switching and time division multiplexing 
(TDM).  Modifications were necessary to support the needs for semantic transparency (the 
correct delivery of transmitted information), and time transparency (the continuity of the bit 
stream).  

ATM defines that:

-       no error control or flow control is done inside the ATM network, but is done only at the 
network boundary,
-       all information is transferred in a virtual circuit,
-       packets have a small, fixed, size (called cells), and
-       a header is used to associate the cell to a virtual circuit.

All time slotted operations of TDM are done at the network's edge.  No time relation is 
maintained inside of the ATM network.  Instead the cell is only associated with a virtual 
circuit.  Therefore, ATM shifts the semantic and time transparency to the boundary of the 
network.  

ATM is a cell-based switching and multiplexing technology designed to be a general purpose 
connection-oriented method of transporting information.  It is designed to handle both 
connection-oriented and connectionless traffic, and support either constant bit rate or variable 
bit rate connections.  Each cell contains addressing information associated with a virtual 
connection, and virtual connections can either be permanent (PVC) or switched (SVC).  ATM is 
asynchronous because transmitted cells do not have to be periodic.

ATM is view as a standard network architecture defining multiplexing and switching, with 
SONET and STM provided as the physical transport standard.  ATM supports multiple 
qualities of service (QOS) classes for different application requirements on delay and loss.  The 
vision is that a single network architecture can support voice, packet data, video, imaging and 
circuit emulation.  ATM is designed to provide bandwidth on demand.

ATM is designed as a hierarchy with local services connected to regional ATM providers who 
are in turn connected to national and international ATM providers.  ATM is a connection-
oriented service.  A cell is not sent until a connection (channel) is established.  ATM is fiber 
based and therefore provides a low bit error rate (BER) on the order of 10-12.  The goal is also to 
provide support for low cost attachments connected to the ATM network.

This leads to hierarchical principles.  ATM is designed to work as a hierarchical network with 
user-network interface (UNI) and a network-network interface (NNI).  It uses a 28 bit channel 
identifier which itself is hierarchical with a channel identifier and a path identifier.  To gain 
access to a user both a path and channel identifier is used; when data is sent across a backbone 
only the path identifier is necessary.

The BER is defined as that of the BER of fiber.  This is done by assigning priorities to cells 
transmitted on the network.  High priority cells are guaranteed the defined BER, at the 
expense of low priority cells if network traffic is excessive.

In order to keep ATM attachments low cost, cell reordering is prohibited.  This greatly 
simplifies buffering, and reduces the cost of hardware.

Basics of Cell Networking

In cell networks, all data should be transmitted in small, fixed-length packets, called cells.  
Data packets from other networks (ethernet) are broken up into smaller packets of data, a 
header is attached resulting in a cell.  On the other end, the cells are reassembled into data 
packets native to the receivers network topology.

ATM cells are 53 bytes, 48 bytes are data, 5 bytes are the header for the cell.  The last cell of 
the stream representing the original data packet will normally not be filled.  The small cell 
size minimizes waste.  Serialization delay, the time it takes to put a cell onto the network, 
depends on the cell size.  The smaller cell size allows for better delivery of real-time data.  
Large data packets which may delay time dependent data are broken up into cell length data, 
and the time-sensitive data will only be delayed the time period it takes to put a single cell 
onto the network.  Cell size is dictated by the available bandwidth.  The higher the 
bandwidth, the larger the cell size can be.  Analog-to-digital conversion is also a factor.  It is 
necessary to buffer converted signals until enough data is available to fill a cell. The larger the 
cell size the more delay in transmission.  Forwarding delay is the time it takes to receive a 
packet and then begin sending the packet on.  Store-and-forward receives the entire packet 
before it is sent on.  Each device performing store-and-forward function adds delay.  An 
alternative is to use cut through forwarding.  Cut through forwarding starts to forward while 
still receiving the packet.  This is faster but more difficult to implement.  It only works when 
the next network link is free.  It also requires very fast routing software.  Cell networking, 
however, can be as fast as cut through if the cell size is appropriate.

ATM Cell Format

An ATM cell is made up of 53 bytes, 48 bytes are user data, the remaining 5 bytes are the header 
used by User-Network Interfaces (UNIs), Network-Network Interfaces (NNIs), and switches 
for routing purposes.  The 48 byte user payload is formatted in the ATM Adaptation Layer 
(AAL).  The size of the cell is a compromise between the data and voice camps who ultimately 
designed the formats.

Cell Header Format

There are two forms of the cell header, one for the UNIs and one for NNIs.  For the NNI there 
is a 12-bit virtual path identifier (VPI) and a 16-bit virtual circuit identifier (VCI).  The VPI 
and VCI provide two levels of addressing for routing.  In general the VPI provides regional 
addressing whereas the VCI provides local addressing.  The switch looks to see if the cell is a 
local cell or a long distance cell.  If it is local both the VPI and VCI are used to determine the 
local routing.  If it is a long distance cell it is sent to an identified long distance switch.  The VPI 
and VCI are local identifiers only; they are unique on a hop-by-hop basis.  Each time the cell 
passes through a switch the VPI and VCI may be regenerated for the next leg of the cells 
journey.  Each switch changes the VPI and VCI for each local hop, or only changes the VPI if it 
is routing between backbone switches; in this case the VCI remains unchanged.  This 
regeneration of identifiers eliminates the possibility of collisions.





The Payload Type Field is used to distinguish between user traffic and operation and 
management (OAM) traffic.  If the first bit is 0, the payload of the cell contains user data.  The 
second bit (of user type) determines if congestion was encountered (0 for no congestion, 1 for 
encountered congestion).  The third bit of the payload type when referring to user data is a 
signaling bit defined in AAL5, which if the bit is set, the end of a datagram is signified.

The CLP bit or Cell Loss Priority bit defines the priority of the cell.  If the bit is set to 1, the 
cell will be dropped before a cell with the CLP=0.  This provides two levels of traffic priority 
which give protection against abusive users and allows users to sometimes send additional 
traffic.  ATM guarantees levels of service.  An end user will subscribe to some level of service 
with some agreed upon data rate.  In order to encourage users to use the full data rate they have 
subscribed to, it is possible to allow users to exceed this data rate on occasion.  Data that 
exceeds the agreed upon data rate will have the CLP bit set, therefore if any congestion occurs 
on the network, those cells will be dropped first.

In addition the CLP protects the service provider from abusive users.  If a user transmits data at 
rate exceeding the agreed upon rate, the network will assign priorities such that the rate of 
high priority cells is within the agreed upon data rate.

The CRC is for the header only, not for the data.  It is necessary to check the integrity of the 
header each time the cell passes through a switch since the header is replaced at each hop.  
The algorithm for the CRC is x8+x2+x+1 and can correct single bit errors in the header as well 
as being able to detect a large class of multiple bit errors.

UNI Header Format





The user-network interface header is slightly different from the network-network interface 
header.  With the UNI header, part of the VPI is replaced with the Generic Flow Control 
(GFC).  This field is used to negotiate with shared access networks about how to multiplex the 
shared network among cells of various ATM connections.  Types of networks include DQDB 
networks.  Currently this is set to all zeros since no definitions exist.  In addition the ATM UNI 
is not fully defined.  Problems still exist and the standards are still being defined specifically 
related to how ATM interfaces with the user device.

ATM Layer Model

The information flow in ATM is defined by B-ISDN Protocol Reference Model.  There are three 
defined planes, the User Plane, Control Plane, and Management Plane.  Each plane represents 
protocol layers associated with three types of traffic flows, user information, signaling 
information (used for call/connection control), and management information used for efficient 
operation of the ATM network.

The User Plane has a hierarchical structure and is made up of the Physical Layer, ATM Layer, 
ATM Adaptation Layer and the Service Layer.  The Physical Layer can be either unformatted 
(cell based) or formatted using SONET.  The ATM Adaptation Layer (AAL) has two sublayers, 
Segmentation and Reassembly (SAR), and the Convergence Sublayer (CS).  The SAR maps user 
information into 48-byte payloads of ATM cells before entry into the ATM network, then 
reassembles the payloads into user information after delivery.  The CS provides adaptation 
functions to the Services Layer that are service-specific.







ATM Physical Layer

The physical layer is concerned with taking cells, converting them into bits and putting the bits 
on the wire.  It provides for the transmission of cells over a medium connecting two ATM 
devices.  Three levels exist within the physical layer, the regeneration section, digital section 
and the transmission path.  The physical layer is also broken up into two sublayers, the 
Physical Medium Dependent (PMD) sublayer an the Transmission Convergence (TC) sublayer.

Physical Medium Dependent Sublayer

The PMD sublayer provides the actual clocking of the bit transmission over the medium.  Three 
single-mode optical interfaces for the ATM UNI are defined.  All are based on ATM-SONET 
transmission methods.  These interfaces include: 

        STS1 (51 Mbps), 
        STS3c (155 Mbps), and 
        STS12c (622 Mbps).  

In addition, several electrical interfaces are defined including:

        DS1 (1.544 Mbps),
        DS2 (6.312 Mbps), and
        DS3 (44.736 Mbps).

Finally, the ATM Forum has defined physical layer interface rates to ease interoperability 
between existing data networks and ATM.  These include:

        FDDI (100 Mbps, multimode fiber), and 
        STP (155 Mbps, over UTP category 3 and category 5).

Transmission Convergence Sublayer

The TC sublayer converts between the bit stream on the physical medium and cells.  It maps the 
cells into a TDM frame format (during transmission) and delineates cells in the receiving bit 
stream.  It uses the Header Error Control (HEC) in the ATM header to correct and detect errors 
within the addressing of the ATM cell.  The HEC is 1 byte and can correct single bit errors, or 
detect other types of errors.  It is also used to detect the delineation of cells in the TDM frames.  
The transmission convergence sublayer also performs speed matching functions by introducing 
null cells into the bit stream when necessary.

ATM Layer

The ATM Layer is responsible for end-to-end sequence-preserving transfer of ATM cell streams.
The ATM Layer is independent of the physical layer.  It uses Virtual Channels (VCs)for 
unidirectional transport of ATM cells associated by a common unique identifier called the 
Virtual Channel Identifier or VCI.  It also uses the Virtual Path (VP), a bundle of VCs, where 
the VP has itself a unique identifier, or VPI.

The ATM layer performs the multiplexing and demultiplexing of cells of different connections 
onto a single cell stream.  The cell header is extracted before the cell is delivered to the ATM 
Adaptation Layer (AAL) or is added after the cell is received from the AAL.  The translation 
of the VCI might be required at switching nodes.  Access flow control mechanism may be 
implemented on the user network interface.

The main functions of the ATM Layer is:

        - cell construction
        - cell reception and header validation
        - cell relaying and forwarding, copying
        - cell multiplexing, demultiplexing
        - cell payload type discrimination
        - cell loss prioritization
        - support for Quality of Service (QOS) classes, and 
        - connection assignment and removal.

The key concepts within the ATM layer are the Virtual Paths and the Virtual Channels.  The 
ATM layer is only concerned with the transmission path.  The ATM device may be either an 
endpoint or a connecting point for the VP or the VC.  The VPC or the VCC exists only between 
endpoints.  The VP link or VC link can exist between endpoint and connecting point or between 
two connecting points.  A VPC or VCC is an ordered list of VP or VC links.

Virtual Channel Level

The cell header contains the Virtual Channel Identifier (VCI).  This identifies a single virtual 
channel on a particular virtual path.  Switching is done based on the VP and the VCI.  The VC 
link is a unidirectional flow of ATM cells with the same VCI between a VC connecting point and 
the VC end point or another VC connecting point.  The Virtual Channel Connection (VCC) is a 
concatenation of VC links and defines the unidirectional flow of cells from one user to another 
user or a group of users.

Virtual Path

The virtual path (VP) is set up in a similar manner except that the virtual path contains 
virtual channels and at switches may change the virtual path, virtual channel or both, 
depending on the congestion of the switch.  Switching at a VP connection point is based solely on 
the VPI, where the VCI is ignored since it only reflects the local destination of the cell.


ATM Adaptation Layers (AALs)

ATM sends data packaged into cells over a connection.  The protocols for packaging data into 
cells are called the ATM Adaptation Layers or AALs.  Each AAL defines how to package a 
specific type of data (voice, video, data, etc.) into cells.  The AALs are needed for 
interoperability between diverse devices and applications and define how the data is 
formatted.  Different AALs are used to optimize the packaging of different types of data.  It is 
the AAL which links services provided by higher layers to the ATM level.

Originally there were four classes of applications which the AALs are defined around.

AAL1 - constant bit-rate applications (bounded delay)
AAL2 - variable bit-rate applications (still needs bounded delay)
AAL3 - connection-oriented data applications (supports historical connection-oriented apps.)
AAL4 - connectionless data applications (datagram networking)

AAL2 is currently undefined.  Variable-rate time-dependent data is currently being developed, 
so there does appear to be a future use for this protocol.  AAL3 and AAL4 were combined because 
both connection-oriented data and connectionless data can be treated as the same, so these two 
AALs are now one, or AAL3/4.  Additionally a 5th AAL was added, AAL5, which is a 
redefinition of how to treat datagrams because it was believed that AAL3/4 was not an 
effective implementation.

Each AAL is broken into two sub-layers, the SAR and the CS.  The SAR or Segmentation and 
Reassembly sublayer is responsible for breaking data into cell length segments and then 
reassembling the cells into user data.  The CS (Convergence sublayer) helps manage the flow of 
data to and from the SAR.

Definition of the AALs

AAL1 is used to support video and phone services.  It would be used to replace leased line 
services currently in place.  AAL1 specifies how TDM-type circuits can be emulated over ATM.  
Emulation exists for DS1 and DS3 services and is defined for others.



AAL1 SAR

One byte of the 48 bytes for the payload is used as a header. 

CSI - provided to the Convergence Sublayer for signaling purposes.
Sequence Count - keeps track of the order of the received cells, checks for out of sequence cells.
CRC field - this is a checksum which is calculated across the first four bits (Sequence Number 
fields).  The sequence number is critical since the order of the cells must be maintained.
Parity bit - Even parity is maintained across the AAL1 SAR.

The AAL1 is very much like fractional T1.  For video, blocks of 128 cells are used (an array of 
128 cells by 47 bytes of data)  where the last four cells of this block are used for error correction 
over the block of data.  The data is carried in a block of 47 bytes by 124 cells.

AAL3/4

AAL3/4 supports both connection-oriented and connectionless traffic.

AAL3/4 SAR

The SAR takes the packet from the Convergence layer and segments it into a series of cells for 
transmission and then reassembles the packet at the receiver.  The AAL3/4 SAR requires 4 
bytes, a 2-byte header and 2-byte trailer, out of the available 48 bytes.



T - Type determines whether the cell is the beginning of the packet (BOM), a cell in the middle 
of the packet (COM), the last cell in the packet (EOM), or a single cell packet (SSM).
Sequence Number - keeps track of the order of the received cells, checks for out of sequence cells, 
4 bits long.
MID - Multiplex Identification, 10 bits, used to determine how multiple packets are interwoven 
on the same ATM connection, if multiplexing is used.
Length - the total number of bytes of data sent.
CRC field - can correct a single bit in the cell.

AAL3/4 Convergence Sublayer

This sublayer define the format for all packets to be sent using the AAL3/4 SAR.  

The AAL3/4 requires that each cell be processed as it arrives.  This requirement creates a great 
deal of overhead, especially if data is sent in large chunks.  AAL5 was designed to reduce the 
amount of overhead for datagram type delivery.

AAL5

AAL5 was designed to provide a more efficient AAL for data transmissions.  It was designed to 
reduce the overhead, and is designed to emulate existing data interfaces like FDDI or Ethernet 
so that it is easier to port data from these types of networks to ATM and back again.  Finally, it 
was designed to minimize the cost for devices for handling cells.

AAL5 SAR





The SAR for the AAL5 is just one bit located in the ATM header.  Therefore all 48 bytes are 
available for user data.  In the last cell in the sequence, the data is padded to 40 bytes, and 
then an 8-byte trailer is inserted.

Each receiving device now only needs to look at the ATM header to see if the cell is the last cell 
in a sequence.  If it is not, the cells are stored until the last cell is received.  At that time the 
CRC is performed on the last cell to determine the integrity of the transmission.

AAL5 Convergence Sublayer

The last cell of the packet contains 8 bytes of convergence layer information.

UU and CPI are not used.
Length field - the number of bytes in the packet
CRC - a 32 bit CRC over the entire convergence layer.  This can detect lost or misordered cells.

User, Control and Management Planes

The User Plane is made up of the Physical Layer, ATM Layer, ATM Adaptation Layer and the 
Service Layers.   The AALs and higher layers provide meaningful interfaces and services to the 
end user applications such as frame relay, SMDS, IP and other protocols.  They also provide 
service to APIs.  These services represent the User Plane.  Applications supported by the user 
plane protocols include video, multicast LAN support, LAN emulation as well as others.

The Control and Management planes are primarily used to support services provided by the 
User Plane.  The Control Plane performs functions needed in the creation of Switched Virtual 
Connections (SVCs) and Private Virtual Connections (PVCs).  The Management Plane is used for 
overall plane management, management of each of the User and Control plane components.

In addition to the User, Control and Management Planes, there is Layer Management which is 
responsible for monitoring the User and Control Planes for faults, and can generate alarms and 
take corrective action when problems arise.  Also the Layer Management monitors compliance 
with the performance stated in the traffic contract with end user devices.

Switching

ATM switches have several characteristics defining their operation:

        - The architecture (blocking or non-blocking), 
        - the switching fabric, and
        - the buffering method.

There are two main categories of switches, internally blocking and internally non-blocking.  
Internally non-blocking switching is where two or more packets at different input ports can be 
simultaneously forwarded to the requested output port provided that they have distinct output 
port addresses.  An internally blocking switch cannot do this due to contentions within the 
switching fabric, for example when more than one packet has to access the same internal link.

If a switch is non-blocking then an input port can be connected to any unoccupied output port up to 
the point where all input ports are occupied.  If this is the case then the switch is considered 
strictly non-blocking.  If blocking exists, then the result is that cell loss occurs.  In an ATM 
switch arriving virtual connection traffic is uniformly and randomly distributed across output 
ports, therefore up to a certain input load a very low cell loss ratio occurs.  If this is the case the 
switch is considered non-blocking.

Switch architectures or switching fabrics are characterized by their complexity, maximum 
overall speed, scalability, ease of support of multicast transmission, and blocking level.  
Switch buffering also plays a key role in blocking (cell loss) performance.  Switch fabrics built 
with internal queuing (buffering) have the potential to scale to large sizes.

One of the lingering difficulties surrounding ATM is trying to build fast packet switches which 
are able to match the high speeds of input links as well as the high performance requirement 
guarantees.

ATM needs high speed hardware switches.  However, due to the use of statistical 
multiplexing, buffering is also required in order to avoid packet loss whenever multiple input 
packets arrive simultaneously on different input ports and are destined for the same output 
port.  Only one packet at time can be transmitted over an output link, so the rest must be stored 
in a buffer.

The proper placement of the buffering system has a dramatic impact on switch performance. 
[Garcia-Haro, 1994].  Assume that the switch is only non-blocking, that it has dedicated buffers 
and that it performs a first-in, first-out (FIFO) operation.  The switch operates at the same 
speed as the input/output lines, and therefore it is not required to speed up the internal 
operations of the switch, and therefore the hardware complexity of the switching fabric is 
lower.  The problem arises when one cell must wait to be processed.  When this happens, all 
successive cells must also wait.  Therefore the throughput is reduced if output contention arises.

One solution is that for each output an output buffer exists such that it can receive up to N cells 
where N is the size of the switch.  However what this means is that the switch must operate N 
times faster than the input/output speed of the network.

Another solution to this problem is to implement output buffering but to use shared memory.  In 
shared-memory type of switches without blocking, all input/output ports have access to the 
shared-memory module which is able to write up to N incoming cells and to read N outgoing 
cells in a single switching time cycle.  Therefore a smaller number of buffers are needed and the 
throughput/delay is optimized.  This type of switch is also easily modified to handle several 
service classes.  The best shared-memory switch is a single-stage common-memory switch.  This 
type of switch provides the lowest delay, and the highest link utilization.  The goal is to build 
a single-stage switch with the maximum achievable size.  Shared-memory switches are also 
attractive because they can use existing VLSI chips and therefore would be less expensive than 
other methods.

Problems which still must be addressed are what method of switch design should be 
implemented, the need for selecting an optimum fabric structure for point-to-point as well as 
multicast connections, what growability limits exist and the development of appropriate 
control methods.  These issues are still being worked on.

Control Issues

Control issues may be more complicated and costly for ATM switches compared to current 
telephone circuit switches [Chen, 1994].  The switching fabric provides the essential routing 
and buffering functions.  ATM switching systems will also require management and control 
functions necessary for the efficient operation of the network.  The main problems are 
associated with operational details.  The control of multiple type of traffic with different 
service requirements, standards on operation and management (OAM) and traffic control 
principles have not yet been finalized.

These problems mean increased cost for implementing systems.  The process and control must be 
exercised on the levels of individual ATM cells as well as virtual connections.  Also, various 
types of traffic (voice, data, video) with different quality of service (QOS) requirements will 
be mixed within the network.  In addition, because of the high speeds of the network, even brief 
losses or degradation of service may result in serious losses of user information.

Outstanding Issues Surrounding ATM

There are many issues which remain unresolved with ATM.  Some of these have become 
apparent in the previous discussion.  Others include:

        - agreement on standards,
        - final definitions for as of yet unimplemented yet promised features,
        - support for interoperability with existing technologies,
        - support for existing protocols and applications, and
        - native applications which can take full advantage of ATM.

It is currently possible to buy ATM of the shelf, however the closer one looks at the details, the 
clearer it becomes that ATM is not ready to provide the services it was intended to offer.  There 
is however overwhelming support and impetus behind ATM, therefore one would believe that 
these outstanding issues will be addressed, and ATM, if not providing all the services it says it 
will provide, then at the very least it will be a valuable part of networking in the not too 
distant future.

3.0  Northwestern University and ATM

At this point it may be useful to look at Northwestern's network and try to see if ATM will or 
should play a role in the future.

Northwestern's Current Network Infrastructure

NUNet currently is based on two network technologies.  The campus backbone is FDDI and the 
local connections, connections to users' workstations is via Ethernet.  Therefore workstations 
normally access a shared 10Mbps access locally, and the campus (or segments of the campus 
since multiple rings are deployed) shares a 100Mbps backbone.  The supported protocols are 
TCP/IP, AppleTalk and IPX to support the existing communities of UNIX workstations, 
Macintoshes and IBM compatibles supported by Novell's NetWare for local file sharing.

The primary applications and files used at the university are either local on an individual's 
workstation or accessed from a local file server.  IPX and AppleTalk services are available 
campus-wide, and TCP/IP provides access to both campus and world-wide network services.

Currently the applications which are in wide-spread use are easily supported by the existing 
network infrastructure.  The only weak link on campus is the intercampus connection between 
Evanston and Chicago which currently is a T1 line.  However this will be replaced by a 
155Mbps SONET connection within several months. 

Finally, NUNet, even though currently based on FDDI, was designed so that future 
technologies (and ATM was at the time the clear direction for future network design) could 
easily be implemented if at some point there was a need to improve the existing infrastructure.  
To that end, fiber is deployed to every building running in a star-like fashion from central 
routers which are themselves designed to support emerging technologies, specifically ATM.  
Therefore, even though there is no apparent need for ATM at this point, and even though ATM 
is as of yet not standardized, and not truly implemented even today, Northwestern's expanded 
network design of two years ago was designed to allow for an easy migration from the currently 
implemented FDDI backbone to an ATM environment.

Future Computing Needs at NU

Applications, the speed of the CPU, disk capacity and the speed at which data can be read 
from and written to disk, as well as network protocols all affect the speed of the overall 
system.  A high speed environment includes processing, storing, displaying, and moving data all 
at speeds in excess of 1 GB/s.

Currently these types of speeds are obtainable from the key components of the system.  Fast 
technologies include:

CPUs which currently can process gigabits of data
RAID (Redundant Arrays of Inexpensive Disks) disk arrays which can support disk I/O at 
gigabit speeds
Graphic systems can display gigabits of data per second (1024x1024x32 bit displays just over a 
gigabit per second)

Unfortunately these systems rely on existing technologies which currently do not support such 
high speeds.  Memory, disk I/O from single disks and computer buses still are bottlenecks.  In 
addition, even though the speed of an application relies on CPU power, applications still are 
unable to deal with data at such high speeds.

It is the applications which dictate what technologies will be needed.  Even though everyone 
has been waiting for multimedia, full-motion video, large scale imaging systems to flood the 
networks, as of yet the applications still are unable to provide the services network engineers 
have been fearing.

Basic Assumptions

Clearly, even though all the parts which make up a high speed networked environment are not 
quite ready, it would be short-sighted not to think that at some point, and chances are that 
point is no more than one to two years away, the promised applications and hardware 
necessary to support and provide bandwidth-intensive data will exist.

Therefore we can make some general assumptions about the needs of the future.  First, 
applications which support the distribution of multimedia, video, and high resolution images 
will exist.  In addition to this, these services will be desired in a multicast environment.  Other 
applications which clearly are desired include collaborative computing applications 
(groupware), videoconferencing, and distributed data access.  It is ultimately the goal to have 
data accessible (and that data could be in any form) on demand from any location.

By looking at these types of services, providing access to information, it may be possible to 
define what type of infrastructure may be needed.  The information described above can be 
characterized as either data of fixed size that must be transferred from on point to another, or 
as a stream of data defined by a certain data rate or bandwidth.

A network which supported 10Mbps transfer rates can only provide real-time access (if no 
contention existed) for a single session.  A 100Mbps network could provide near real-time 
delivery of data in the range of still images (10-100Mb files) again without contention.  If real-
time data is desired, the existing networking infrastructure could support very few concurrent 
sessions.  If video were to be multicast, assuming 10Mbps was needed to support a single session, 
only a handful of concurrent sessions could be supported even with the existing backbone speeds.

It is clear that if and when applications supporting the distribution of video, the existing 
infrastructure would be unable to support large scale use of such applications.  It is also clear 
that ATM is designed to support these types of services and will be needed in such an 
environment.

The User Community

The last part of this is to look at the needs and abilities of the user community and the support 
available to that community.  Even though computers are everywhere, not everyone is using 
computers, and a large majority of those using computers are not sophisticated users.  The 
network at Northwestern has just now been brought to everyone who wants (or perhaps can 
afford) access to it.  A large number of those users who are now just beginning to use the network 
use only basic applications.  Although there are desires for more sophisticated applications 
which provide a multitude of services to the end-user, these applications require the end-user 
to be more sophisticated as well.  The faster technologies race ahead, the greater the gap will 
be between these technologies and the average end-user.  Therefore, the community which is 
able to take advantage of advanced applications will be a small minority at least initially 
and therefore it may not be necessary to provide 155Mbps services to every end-user, but only to 
a few over the next several years.

Conclusions

ATM has a great deal of support and if it is not reality now, it will be in the near future.  Even 
though many details need to be addressed for it to be able to offer the full services it promises, 
limited implementations exist not just as test cases, but in production environments.  It will 
however be several years before the general public has access to this type of high speed 
service.

At Northwestern the existing network infrastructure is capable of supporting the general-user's 
needs within the university for several years.  Specialized needs for higher bandwidth will 
inevitably arise sooner than that, and it will be useful to attempt to implement small scale 
ATM networks if not to support existing needs then at the very least to better understand the 
technology.

Northwestern's existing network is designed so that when it is needed the transition to ATM 
will be a relatively easy migration from existing technologies.  When it is time to make that 
transition, the university will be ready to do so.


4.0  References

Biagioni, E., Cooper, E., Sansom, R., "Designing a Practical ATM LAN," IEEE Network, March 
1993, pp. 32-39.
De Prycker, M., Asynchronous Transfer Mode:  Solution for Broadband ISDN, Second Edition, 
1993.
De Prycker, M., Peschi, R., Van Landegem, T., "B-ISDN and the OSI Protocol Reference Model," 
IEEE Network, March 1993, pp. 10-18.
Garcia-Haro, J., Jajszczyk, A., "ATM Shared-Memory Switching Architectures," IEEE 
Network, July/August 1994, pp. 18-26.
Leslie, I., McAuley, D., Tennenhouse, D., "ATM Everywhere?" IEEE Network, March 1993, pp. 
40-46.
Lyles, J., Swinehart, D., "The Emerging Gigabit Environment and the Role of Local ATM," IEEE 
Communications Magazine, April 1992, pp. 52-58.
McDysan, D., Spohn, D., ATM, Theory and Application, 1995.
Newman, P., "ATM Technology for Corporate Networks," IEEE Communications Magazine, 
April 1992, pp. 90-101.
Partridge, C., Gigabit Networking, 1994.
Page