3
Network Technology
This chapter concerns the generic aspects of network technology that are important in
providing transport services and giving them certain qualities of performance. We define
a set of generic control actions and concepts that are deployed in today’s communication
networks. Our aim is to explain the workings of network technology and to model those
issues of resource allocation that are important in representing a network as a production
plant for service goods.
In Section 3.1 we outline the main issues for network control. These include the timescale
over which control operates, call admission control, routing control, flow control and
network management. Tariffing and charging mechanisms provide one important type of
control and we turn to these in Section 3.2. Sections 3.3 and 3.4 describe in detail many
of the actual network technologies in use today, such as Internet and ATM. We relate these
examples of network technologies to the generic control actions and concepts described
in earlier sections. In Section 3.5 we discuss some of the practical requirements that must
be met by any workable scheme for charging for network services. Section 3.6 presents a
model of the business relations amongst those who participant in providing Internet services.
3.1 Network control
A network control is a mechanism or procedure that the network uses to provide services.
The more numerous and sophisticated are the network controls, the greater and richer can
be the set of services that the network can provide. Control is usually associated with the
procedures needed to set up new connections and tear down old ones. However, while a
connection is active, network control also manages many other important aspects of the
connection. These include the quality of the service provided, the reporting of important
events, and the dynamic variation of service contract parameters.
Synchronous services provided by synchronous networks have the simple semantics of a
constant bit rate transfer between two predefined points. They use simple controls and all
bits receive the same quality of service. Asynchronous networks are more complex. Besides
providing transport between arbitrary points in the network, they must handle unpredictable
traffic and connections of arbitrarily short durations. Not all bits require the same quality
of service.

use cells. The packets must be broken into cells and then later reconstructed into packets.
We will use the term packet in the broad sense of a data parcel, unless specific reasons
require the terminology of a cell.
Connections. A connection is the logical concept of binding end-points to exchange data.
Connections may be point-to-point, or point-to-multipoint for multicasting, although not
all technologies support the latter. A connection may last from a few seconds (as in the
access of web pages) to years (as in the connection of a company’s network to the Internet
backbone). Depending on the technology in use, a connection may or may not be required.
The transfer of web page data as packets requires a connection to be made. In contrast,
there is no need to make a connection prior to sending the packets of a datagram service.
Clearly, the greater is a technology’s cost for setting up a connection the less well suited
it is to short-lived connections. Once a connection has been set up, the network may have
to allocate resources to handle the connection’s traffic in accordance with an associated
Service Level Agreement.
Flows. The information transported over a connection may be viewed as a continuous flow
of bits, bytes, cells or packets. An important attribute of a flow is its rate. This is the amount
of information that crosses a point in the network, averaged over some time period. The job
of a network is to handle continuous flows of data by allocating its resources appropriately.
For some applications, it may have to handle flows whose rates are fluctuating over time.
We call such flows ‘bursty’. When network resources are shared, instead of dedicated on a
per flow basis, the network may seek to avoid congestion by using flow control to adjust
the rates of the flows that enter the network.
Calls. These are the service requests that are made by applications and which require
connections to be set up by the network. They usually require immediate response from the
NETWORK CONTROL 43
network. When a customer places a call in the telephone network, a voice circuit connection
must be set up before any voice information can be sent. In the Internet, requests for web
pages are calls that require a connection set-up. Not all transport technologies possess
controls that provide immediate response to calls. Instead, connections may be scheduled
long in advance.

propogation time
connection
interarrival time
minutes
months, years
Figure 3.1 Network control takes place on many timescales. Cell discard decisions are made
every time a cell is received, whereas pricing policy takes place over months or years. Pricing
mechanisms (algorithms based on economic models) can be used for optimizing resource sharing at
all levels of network control.
44 NETWORK TECHNOLOGY
packet conforms to the traffic contract. If it does not, then the node takes an appropriate
policing action. It might discard the packet, or give it a lower quality service. In some cases,
if a packet is to be discarded, then a larger block of packets may also be discarded, since
losing one packet makes all information within its context obsolete. For instance, consider
Internet over ATM. An Internet packet consists of many cells. If a packet is transmitted
and even just one cell from the packet is lost, then the whole packet will be resent. Thus,
the network could discard all the cells in the packet, rather than waste effort in sending
those useless cells. This is called ‘selective cell discard’.
A crucial decision that a network node must take on a per packet basis is where to forward
an incoming packet. In a connectionless network, the decision is based on the destination of
the packet through the use of a routing table. Packets include network-specific information
in their header, such as source and destination addresses. In the simplest case of a router
or packet switch the routing table determines the node that should next handle the packet
simply from the packet’s destination.
In a connection-oriented network, the packets of a given connection flow through a path
that is pre-set for the connection. Each packet’s header contains a label identifying the
connection responsible for it. The routing function of the network defines the path. This is
called virtual circuit switching, or simply switching. More details are given in Section 3.1.4.
Forwarding in a connection-oriented network is simpler than in a connectionless one, since
there are usually fewer active connections than possible destinations. The network as a

2
/;:::;.l
n
; a
n
/, with labels a
i
; i D 1;:::;n. Labels are unique identifiers and
may be coded by integers. Such a label-switched path is programmed inside the network by
1. associating r
a
at node A with the pair .l
1
; a
1
/, and at node B with .l
n
; a
n
/;
2. adding to the switching table of each of the intermediate node i the local mapping
information .l
i 1
; a
i 1
/ ! .l
i
; a
i
/, i D 2;:::;n.

1
l
2
l
n−1
l
n
header
Figure 3.2 A label-switched path implementing a virtual circuit between nodes A and B.
changes the label to the new value a
i
, as dictated by the information in its switching table,
see Figure 3.2. At the end of the path, the packets of connection a arrive in sequence at node
B carrying label a
n
. The pair .l
n
; a
n
/ identifies the data as belonging to connection a.When
the connection is closed, the label-switched path is cleared by erasing the corresponding
entries in the switching tables. Thus, labels can be reused by other connections.
Because a label-switched path has the semantics of a circuit it is sometimes called a vir-
tual circuit. One can also construct ‘virtual trees’ by allowing many paths to share an initial
part and then diverge at some point. For example, binary branching can be programmed in
a switching table by setting .l
i
; a
i
/ ! [.l

may also need to reflect the network provider’s policy concerning bandwidth reservation and
admission priorities for certain call types.) It is not realistic to have complete information
about the state of the network at the time of each admission decision. This would require
excessive communication within the network and would be impossible for networks whose
geographic span means there are large propagation delays. A common approach is for the
network management to keep this information as accurately as possible and update it at
time intervals of appropriate length.
The call admission control mechanism might be simple and based only on traffic
contract parameters of the incoming call. Alternatively, it might be complex and use data
from on-line measurements (dynamic call admission control ). Clearly, more accurate CAC
allows for better loading of the links, less blocking of calls, and ultimately more profit
46 NETWORK TECHNOLOGY
for the network operator. To assess the capacity of the network as a transport service
‘production facility’, we need to know its topology, link capacities and call admission
control policy. Together, these constrain the set of possible services that the network can
support simultaneously. This is important for the economic modelling of a network that we
pursue in Chapter 4. We define for each contract and its resulting connection an effective
bandwidth. This is a simple scalar descriptor which associates with each contract a resource
consumption weight that depends on static parameters of the contract. Calls that are easier
to handle by the network, i.e. easier to multiplex, have smaller effective bandwidths. A
simple call admission rule is to ensure that the sum of the effective bandwidths of the
connections that use a link are no more than the link’s bandwidth.
In networks like the Internet, which provide only best-effort services, there is, in
principle, no need for call admission control. However, if a service provider wishes to
offer better service than his competitors, then he might do this by buying enough capacity
to accommodate his customers’ traffic, even at times of peak load. But this would usually
be too expensive. An alternative method is to control access to the network. For instance, he
can reduce the number of available modems in the modem pool. Or he can increase prices.
Prices can be increased at times of overload, or vary with the time of day. Customers who
are willing to pay a premium gain admission and so prices can act as a flexible sort of call

network switches
source
destination
X
traffic contract
Figure 3.3 In a connection-oriented network each newly arriving call invokes a number of
network controls. Call routing finds a path from the source to destination that fulfils the user’s
requirements for bandwidth and QoS. Call admission control is applied at each switch to determine
whether there are enough resources to accept the call on the output link. Connection set-up uses
signalling mechanisms to determine the path of the connection, by routing and CAC; it updates
switching tables for the new virtual circuit and reserves resources. Above, X marks a possible route
that is rejected by routing control. Flow control regulates the flow in the virtual circuit once it is
established.
individual packets. When a call is admitted, the network uses its signalling mechanism
to set the appropriate information and reserve the resources that the call needs at each
network node along the path. This signalling mechanism, together with the ability to reserve
resources for an individual call on a virtual circuit, is a powerful tool for supporting
different QoS levels within the same network. It can also be used to convey price
information.
During the signalling phase, call admission control functions are invoked at every node
along the connection’s path. The call is blocked either if the entry node decides that there
are insufficient resources inside the network, or if the entry node decides that there may be
enough resources and computes a best candidate path, but then some node along that path
responds negatively to the signalling request because it detects a lack of resources. A similar
operation takes place in the telephone network. There are many possibilities after such a
refusal: the call may be blocked, another path may be tried, or some modification may be
made to the first path to try to avoid the links at which there were insufficient resources.
Blocking a call deprives the network from extra revenue and causes unpredictable delays
to the application that places the call. Call blocking probability is a quality of service
parameter that may be negotiated at the service interface. Routing decisions have direct

prices to optimize the overall performance of the network. Observe that such an approach
reduces the complexity of the network, but places more responsibility with the users. It
is consistent with the Internet’s philosophy of keeping network functions as simple as
possible. However, it may create dangerous instabilities if there are traffic fluctuations and
users make uncoordinated decisions. This may explain why network operators presently
prefer to retain control of routing functions.
3.1.7 Flow Control
Once a guaranteed service with dynamic contract parameters is admitted, it is subject to
network control signals. These change the values of the traffic contract parameters at the
service interface and dictate that the user should increase or decrease his use of network
resources. The service interface may be purely conceptual; in practice, these control signals
are received by the user applications. In principle the network can enforce its flow control
‘commands’ by policing the sources. However, in networks like the Internet, this is not
done, because of implementation costs and added network complexity.
In most cases of transport services with dynamic parameters (such as the transport service
provided by the TCP protocol in the Internet), the network control signals are congestion
indication signals. Flow control is the process with which the user increases or decreases his
transmission rate in response to these signals. The timescale on which flow control operates
is that of the time it takes the congestion indication signals to propagate through the network;
this is at most the round trip propagation time. Notice that the controls applied to guaranteed
services with purely static parameters are open-loop: once admitted, the resources that are
needed are reserved at the beginning of the call. The controls applied to guaranteed services
with purely dynamic parameters are closed-loop: control signals influence the input traffic
with no need for apriori resource reservation.
Flow control mechanisms are traditionally used to reduce congestion. Congestion can be
recognized as a network state in which resources are poorly utilized and there is unaccept-
able performance. For instance, when packets arrive faster at routers than the maximum
speed that these can handle, packet queues become large and significant proportions of pack-
ets overflow. This provides a good motivation to send congestion signals to the sources
before the situation becomes out of hand. Users see a severe degradation in the perfor-

signalling mechanisms at the network nodes. In Chapter 10 we investigate flow control
mechanisms that control congestion and achieve economic fairness.
The use of flow control as a mechanism for implementing fair bandwidth allocation relies
on users reacting to flow control signals correctly. If a flow control mechanism relies on the
user to adjust his traffic flow in response to congestion signals and does not police him then
there is the possibility he may cheat. A user might seek to increase his own performance
at the expense of other users. The situation is similar to that in the prisoners’ dilemma (see
Section 6.4.1). If just one user cheats he will gain. However, if all users cheat, then the
network will be highly congested and all users will lose. This could happen in the present
Internet. TCP is the default congestion response software. However, there exist ‘boosted’
versions of TCP that respond less to congestion signals. The only reason that most users
still run the standard version of TCP is that they are ignorant of the technological issues
and do not know how to perform the installation procedure.
Pricing can give users the incentive to respond to congestion signals correctly. Roughly
speaking, users who value bandwidth more have a greater willingness to pay the higher
rate of charge, which can be encoded in a higher rate of congestion signals that is sent
during congestion periods. Each user seeks what is for him the ‘best value for money’ in
terms of performance and network charge. He might do this using a bandwidth seeking
application. It should be possible to keep congestion under control, since a high enough
rate of congestion charging will make sources reduce their rates sufficiently.
50 NETWORK TECHNOLOGY
Sometimes flow control may be the responsibility of the user rather than the network. For
instance, if the network provides a purely best-effort service, it may be the responsibility
of the user to adjust his rate to reduce packet losses and delays.
3.1.8 Network Management
Network management concerns the operations used by the network to improve its
performance and to define explicit policy rules for security, handling special customers,
defining services, accounting, and so on. It also provides capabilities for monitoring the
traffic and the state of the network’s equipment. The philosophy of network management is
that it should operate on a slow timescale and provide network elements with the information

reconsiders his tariffs and this leads to further adjustment of user demand. The timescale
over which these adjustments take place is typically months or years. Moreover, regulation
may prevent a supplier from changing tariffs too frequently, or require that changes make no
customer worse off (the so-called ‘status-quo fairness’ test of Section 10.1). In comparison,
dynamic pricing mechanisms may operate on the timescale of a round trip propagation
time; the network posts prices that fluctuate with demand and resource availability. The
SERVICE TECHNOLOGIES 51
user’s software closely monitors the price and optimally adjusts the consumption of network
resources to reflect the user’s preferences.
Dynamic pricing has an implementation cost for both the network and the customers. A
practical approximation to it is time-of-day pricing, in which the network posts fixed prices
for different periods of the day, corresponding to the average dynamic prices over the given
periods. This type of pricing requires less complex network mechanisms. Customers like it
because it is predictable.
It is a misconception that it is hard for customers to understand and to react to dynamic
prices. One could envision mechanisms that allow customers to pay a flat fee (possibly zero)
and the network to adapt the amount of resources allocated at any given time so that each
customer receives the performance for which he pays. Or customers might dynamically
choose amongst a number of flat rate charging structures (say, gold, silver or bronze) and
then receive corresponding qualities of service. In this case prices are fixed but performance
fluctuates. Alternatively, a customer might ask for a fixed performance and have a third party
pay its fluctuating cost. This is what happens in the electricity market, in which generators
quote spot prices, but end-customers pay constant prices per KWh to their suppliers. A
customer might buy insurance against severe price fluctuations. All of these new value-
added communication service models can be implemented easily since they mainly involve
software running as a network application.
Suppose that a network service provider can implement mechanisms that reflect resource
scarcity and demand in prices, and that he communicates these to customers, who on the
basis of them take decisions. Ideally, we will find that as the provider and users of network
services freely interact, a ‘market-managed network’ emerges, that has desirable stability

At a higher layer, asynchronous technologies such as IP, ATM and Frame Relay, break
information streams into data packets (or cells) that are placed in the frames (or the
smaller sub-frames). Their goal is to perform statistical multiplexing, i.e. to efficiently
fill these frames with packets belonging to different information streams. At the lowest
layer, these framing services may operate over fibre by encoding information bits as light
pulses of a certain wavelength (the ‘½’). Other possible transmission media are microwave
and other wireless technologies. For example, a satellite link provides for synchronous
framing services over the microwave path that starts from the sending station and reflects
off the satellite to all receivers in the satellite’s footprint. In contrast to SONET, Gigabit
and 10 Gigabit Ethernet is an example of a framing service that is asynchronous and of
variable size. Indeed, an Ethernet frame is constructed for each IP packet and is transmitted
immediately at some maximum transmission rate if conditions permit. As we will see, since
Ethernet frames may not depart at regular intervals (due to contention resulting from the
customers using the same link), Ethernet services may not provide the equivalent of a fixed
size bit pipe. Guaranteed bandwidth can be provided by dedicating Ethernet fibre links to
single customer traffic. Finally, note that ATM is an asynchronous service that is used by
another asynchronous service, namely IP. The IP packets are broken into small ATM cells
which are then used to fill the lower-level synchronous frames.
Our discussion so far suggests that customers requiring connections with irregular
and bursty traffic patterns should prefer higher layer asynchronous transport services.
Asynchronous services then consume lower layer framing services (synchronous or
asynchronous), which usually connect the network’s internal nodes and the customers to the
network. Framing services consume segments of fibre or other transmission media. Observe
that a customer whose traffic is both great and regular enough efficiently to use large
synchronous containers, might directly buy synchronous services to support his connection.
Similarly, large customers with bursty traffic may buy asynchronous container services, e.g.
Ethernet services, that allow further multiplexing of the raw fibre capacity.
Figure 3.4 shows a classification of the various transport services that we present in the
next sections. For simplicity we assume that the physical transmission medium is fibre. In
fact, microwave and wireless are also possible media. This may complicate the picture some-

granularity, and the underlying controls for call set-up do not work in very fast timescales. Services
towards the top offer flexible pipes of arbitrarily small granularity and small to zero set-up cost that
can be established between any arbitrary pair of network edge points. Fibre is ‘technology neutral’
in the sense that the higher layer protocols dictate the details of information transmission.
Services towards the top of the diagram build flexible pipes of arbitrarily small
granularity. These are mainly TCP/IP and UDP/IP pipes, since the dynamic call set-up of
the ATM standard is not implemented in practice. (Note, also, that we have denoted ATM
and Frame Relay as guaranteed services, in the sense that they can provide bandwidth
guarantees by using an appropriate SLA. These service have more general features that
allow them to provide best-effort services as well.)
Connections using services at the top of the diagram have little or no set-up cost, and can
be established between arbitrary pairs of network edge points. This justifies the use of the IP
protocol technology for connecting user applications. In the present client-server Internet
model (and even more in future peer-to-peer communications models), connections are
extremely unpredictable in terms of duration and location of origin-destination end-points.
Hence the only negative side of IP is the absence of guarantees for the diameter of the
pipes of the connections. Such a defect can be corrected by extending the IP protocol, or
by performing flow isolation. This means building fixed size pipes (using any of the fixed
size pipe technology) between specific points of the network to carry the IP flows that
require differential treatment. This is the main idea in the implementation of Virtual Private
Networks described in detail in Section 3.4.1 using the MPLS technology.
We now turn to detailed descriptions of the basic connection technologies.
3.3.2 Optical Networks
Optical networks provide a full stack of connection services, starting from light path
services at the lowest layer and continuing with framing services, such as SONET and
Ethernet, up to ATM and IP services. We concentrate on the lower layer light path services
since the higher layers will be discussed in following sections.
Dense Wavelength Division Multiplexing (DWDM) is a technology that allows multiple
light beams of different colours (½s) to travel along the same fibre (currently 16 to 32 ½s,
with 64 and 80 ½ in the laboratories). A light path is a connection between two points in

of the light pulses. In the case of a light path provided over an all-optical network, where
there is optical to electrical signal conversion for switching and regeneration, the electro-
optical components may pose further restrictions on maximum bit rates that can be supported
over the light path.
A dark fibre service is one in which a customer is allocated the whole use of an optical
fibre, with no optical equipment attached. The customer can make free use of the fibre. For
example, he might supply SONET services to his customers by deploying SONET over
DWDM technology, hence using more than a single ½s.
There is today a lot of dark fibre installed around the world. Network operators claim that
their backbones have capacities of hundreds of Gigabits or Terabits per second. Since this
capacity is already in place and its cost is sunk, one might think that enormous capacity can
be offered at almost zero cost. However, most of the capacity is dark fibre. It is costly to add
lasers to light the fibre and provide the other necessary optical and electronic equipment.
This means there is a non-trivial variable cost to adding new services. This ‘hidden’ cost
may be one reason that applications such as video on demand are slow to come to market.
3.3.3 Ethernet
Ethernet is a popular technology for connecting computers. In its traditional version, it
provides a best-effort framing service for IP packets, one Ethernet frame per IP packet. The
framed IP packets are the Ethernet packets which can be transmitted only if no other node of
the Ethernet network is transmitting. The transmission speeds are from 10 Mbps to 10 Gbps
in multiples of ten (and since the price of a 10 Gbps Ethernet adaptor card is no more than
2:5 times the price of a 1 Gbps card, the price per bit drops by a factor of four). Ethernet
SERVICE TECHNOLOGIES 55
technologies that use switching can provide connection-oriented services that are either
best-effort or have guaranteed bandwidth. Ethernet can provide service of up to 54 Mbps
over wireless and over the twisted-pair copper wires that are readily available in buildings.
Twisted-pair wiring constrains the maximum distance between connected equipment to 200
meters. For this reason, Ethernet has been used mainly to connect computers that belong
to the same organization and which form a Local Area Network (LAN). It is by far the
most popular LAN technology, and more than 50 million Ethernet interface cards are sold

statistically multiplexed among the competing edge devices in a best-effort fashion; see
Figure 3.5. This may be a good idea if such a service is provided for data connections
that are bursty. Bursty data sources value the possibility of sending at high peak rates,
such as 10 Mbps, for short periods of time. Statistical arguments suggest that in high speed
links, statistical multiplexing can be extremely effective, managing to isolate each data
source from its competitors (i.e. for most of the time each device can essentially use the
network at its maximum capability). Proprietary Ethernet switching technologies allow for
manageable network resources, i.e. virtual circuits may be differentiated in terms of priority
and minimum bandwidth guarantees.
Connectivity providers using the Gigabit and 10 Gigabit Ethernet technology provide
services more quickly and in more flexible increments than competitors using the traditional
56 NETWORK TECHNOLOGY
N
B
E
C
F
D
A
1 Gbps
ISP
1
ISP
2
R
1
R
2
100 Mbps
Figure 3.5 The left of the figure shows a simple Ethernet network. N is an Ethernet switch, and

SONET service or four services of 622.08 Mbps, or two 622.08 Mbps and four 155.52 Mbps
services. In that sense, SONET and SDH can be seen as multiplexing technologies for
synchronous bit streams with rates being multiples of 155.52 Mbps.
An important quality of service provided by SONET and SDH networks is the ability
to recover in the event of fibre disruption or node failure. The nodes of SONET and
SDH networks are typically connected in a ring topology which provides redundancy by
keeping half of the capacity of the ring, the ‘protection bandwidth’, as spare. If the fibre
of the ring is cut in one place, SONET reconfigures the ring and uses the spare capacity