ISS97 ID PS675

ARCHITECTURAL CHOICES FOR PACKET SWITCHED

TELEPHONE NETWORKS

Mario Baldi and Davide Bergamasco

Politecnico di Torino
Dipartimento di Automatica e Informatica
Corso Duca degli Abruzzi 24 - 10129 Torino - ITALY
Tel. +39 11 564 7067 Fax +39 11 564 7099
e-mail: [ MBaldi | Bergamasco ]@polito.it
Eugenio Guarene
Centro Studi e Laboratori Telecomunicazioni (CSELT)
Via G. Reiss Romoli, 274 - 10148 Torino - ITALY
Tel. +39 11 228 5079 Fax +39 11 228 5069
e-mail: Eugenio.Guarene@cselt.stet.it

Areas of Interest:
A5. Multi-Media Communications Services
B2. Internet Architecture
B7. Broadband ISDN
C4. Routers, service switches, multimedia servers

ISS97 ID PS675

ARCHITECTURAL CHOICES FOR PACKET SWITCHED TELEPHONE NETWORKS

Abstract
Today, Internet telephony is a reality but this cannot be regarded as a true telephone service
because its quality is highly variable depending on network load. The end-to-end delay (i.e., the
time between when the speaker talks and when the listener hears) is a major quality index for
telephone services. This delay must be shorter than 100-200 ms in order to allow interaction
among users. We refer to this bound as the interaction bound.
This work has two main goals:
study the end-to-end delay of voice transmission over a packet switched network from an
analytical point of view, and
identify some possible architectural solutions for packet switched networks expressly designed to carry telephone traffic (we call them packet switched telephone networks).
We first identify the delay components and the main factors in determining their upper bound.
This analysis shows that the current Internet architecture is not suited for providing telephone
services with end-to-end delay shorter than the interaction bound.
Then we analyze the end-to-end delay more deeply to identify the key factors in the design of
a packet switched telephone network providing commercial quality telephone services, i.e., for
possibly replacing traditional telephone networks. We show that by limiting the number of hops
in the path between any pair of users and dimensioning link capacity accordingly, the end-to-end
delay can be kept below the interaction bound. Thus, packet switching could be exploited to provide a high quality telephone service even though the raw capacity needed is larger than with
circuit switching. This is justified by the possibility of carrying best effort traffic on the same
network and by the lower installation and management costs.
In this work we consider two enabling technologies for a packet switched telephone network:
the Asynchronous Transfer Mode (ATM) and the Internet Protocol (IP). Our analytical model
shows that ATM outperforms IP as was expected since the former was expressly designed to
support low bandwidth real-time traffic. Nevertheless, because of the large number of existing
applications based on IP and its internetworking capabilities, it must be taken into account as a
candidate technology for carrying both best effort and high bandwidth real-time multimedia traffic in packet switched telephone networks.
We propose integrated services IP/ATM models to exploit both IP and ATM in the implementation of packet switched networks for telephony. These models are derived from the integrated IP/ATM approaches currently being exploited for an effective operation of IP over ATM
networks. Integrated services IP/ATM models provide ATM based services for telephony and IP
(over ATM) based services for any other kind of traffic (namely, best effort and multimedia) in
order to get the best from the two technologies.

Areas of Interest:
A5. Multi-Media Communications Services
B2. Internet Architecture
B7. Broadband ISDN
C4. Routers, service switches, multimedia servers

ARCHITECTURAL CHOICES FOR PACKET SWITCHED TELEPHONE NETWORKS

ISS97 ID PS675

1 Introduction
A big deal of emphasis is being put on Internet telephony. Even though it is possible to transmit voice over the Internet, this cannot be regarded as a telephone service because quality is generally low. In fact, depending on the network load along the communication path, quality can
vary widely.
A major quality index in interactive voice services is the delay perceived by the users. In order
for a conversation not to be annoying, the end-to-end delay (i.e., the time between when the
speaker talks and when the listener hears) must be shorter than 100-200 ms. We call this maximum acceptable delay the interaction bound. If the voice service is used for interactive communication, the end-to-end delay is required to be always below the interaction bound.
This work has two main goals:
study the end-to-end delay of voice transmission over a packet switched network1 from an
analytical point of view, and
identify some possible architectural solutions for packet switched networks expressly designed to carry telephone traffic. We call such networks Packet switched Telephone Networks
(PTNs).
In Section 2 we identify the delay components and the main factors in determining the delay
bound. This first analysis shows that the current Internet architecture is not suited to the provision of telephone services with end-to-end delay shorter than the interaction bound.
In Section 3 we analyze the end-to-end delay more deeply and identify the key factors in the
design of a PTN conceived for providing commercial quality telephone services, i.e., for possibly
replacing traditional telephone networks. This is attractive because unused capacity can be exploited to carry data traffic, and network deployment and maintenance are easier since the network itself can be used for management purposes.
Section 4 shows some numerical results obtained by applying the analytical equations devised
in Section 3 to a PTN designed after the Telecom Italias telephone network (i.e., using the same
topology and number of supported calls). Various network configurations, which differ for enabling technology, link capacity, and speech encoding techniques, are compared, showing that
some of them guarantee the interaction bound. In this work we consider two enabling technologies for a PTN: the Asynchronous Transfer Mode (ATM) and the Internet Protocol (IP).
ATM was standardized by the ITU-T as the foundation of the Broadband - Integrated Services Digital Network (B-ISDN) [1] and was explicitly designed for the provision of real-time
services while taking advantage of statistical multiplexing for an efficient usage of network resources.
On the other side, the Internet protocol suite is currently the most widely deployed network
architecture. IP was designed as an internetworking protocol for carrying best effort traffic
among heterogeneous networks and lots of applications that use its services have been developed. The Internet community is actively working on the Integrated Services Internet [2], an
Internet architectural evolution aiming at providing real-time services over IP. This will be fundamental to the videoconferencing and telephony applications which are being deployed over the
Internet.
We claim that ATM and IP together can be successfully exploited for devising service inte1

In the context of this work packet switching is intended in a broad sense, encompassing also cell switching.

ARCHITECTURAL CHOICES FOR PACKET SWITCHED TELEPHONE NETWORKS

ISS97 ID PS675

gration over a PTN. IP is desirable for its broad diffusion and the large number of existing applications. Moreover, due to its internetworking capabilities, it allows systems to communicate
through heterogeneous technologies. When deployed together with mechanisms for guaranteeing
Quality of Service (QoS), it is also suited to real-time applications which require both high
bandwidth and low latency like videoconferencing. On the other side, ATM has been designed to
provide QoS guarantees, but it requires all the network to be implemented with the same technology and all the existing applications to be rewritten in order to take advantage of its capabilities. Nevertheless, ATM outperforms IP when dealing with telephony, i.e., low bandwidth realtime applications, as reported in Section 4.
In Section 5 various integrated IP/ATM architectural models for the exploitation of IP and
ATM in PTNs are described, while conclusions are drawn in Section 6.
2 Delay of Voice Transmission over Packet Switched Networks
The end-to-end delay in the transmission of voice over a packet switched network has three
components:
1. The processing delay Dproc is introduced when processing the audio signal. Digital transmission of voice requires the audio signal to be sampled 8000 times per second. Samples are
quantized, encoded, and transmitted to the receiver which plays them at a fixed pace. The
Pulse Code Modulation (PCM) encoding, which is used on traditional digital telephone networks, uses 8 bits to encode each sample, thus generating a 64 kb/s flow for each voice connection. The processing delay introduced by PCM is negligible. The voice signal can be encoded with different techniques in order to produce a bit stream at rates as low as 8 kb/s with
audio quality comparable to PCM [3]. Some of these techniques introduce a delay up to 100
ms [4], thus not being suited to interactive telephone services.
2. The network delay Dnet is given by the time to inject into and propagate through the network
the data stream. It has five components:
i. The voice encoder produces a bit stream at a rate of R bits per second. Before being transmitted through the network, bits are clustered in packets. In each packet P bits are stuffed
in the payload, this clustering introduces a packetization delay
P
(1)
D
=
pkt R
ii. The transmission delay
P
D = s
tr C
is introduced to send a packet of size Ps over a link having capacity C.
iii. Dpr is the delay due to signal propagation through the physical links connecting network
nodes.
iv. The node processing delay Dnp is introduced by a network node each time it has to forward
a packet.
v. The queuing delay is the time spent by packets in nodes buffers while contending for the
same output port. It is a relevant component in determining the end-to-end delay of a voice
connection. The queuing delay Qi experienced by a packet in the ith node on the route to the
destination, depends on the instantaneous status of the output buffer associated to the link

ARCHITECTURAL CHOICES FOR PACKET SWITCHED TELEPHONE NETWORKS

ISS97 ID PS675

on which the packet must be forwarded. As a consequence the queuing delay is not known
a priori; whenever a best effort service is provided (as currently is by the Internet) a bound
is given by
B P
(2)
s,
0Q i
i
C
i
where Bi is the dimension (in bits) of the buffer and Ci is the capacity of the output link2.
3. A good quality reconstruction of the voice signal at the receiver requires samples to be played
at regular intervals (e.g., every 125 s). Whenever the two previous delay components are
variable in time (i.e., the end-to-end delay presents a jitter), different samples experience different end-to-end delays. If they were played as soon as they are available to the receiving application, they would not be uniformly spaced (in time). Thus, the receiving application uses a
replay buffer to store the samples that experience a delay shorter than the maximum: it retrieves and plays the samples at a regular pace. The compensation delay, i.e., the time spent by
early samples in the replay buffer, should be between 0 and RB = Dproc + Dnet, where
Dproc and Dnet are the maximum3 variations of the processing (item 1 above) and network
(item 2) delays, respectively. Actually, since the delay experienced by samples when entering
the replay buffer is not known (see [5] for details), the compensation delay is bound by 2 RB.
Thus, the replay buffer also introduces the excess compensation delay Ec = [0, RB].
Due to the jitter compensation, the delay perceived by the user is
Dete = max(Dproc) + max(Dnet) + Ec
which can be rewritten as
Dete = Dproc + Dpkt + (N + 1)Dtr + Dpr + N Dnp + Qmax + Ec

(3)

where N is the number of network nodes on the path between sender and receiver and Qmax is
the maximum queuing delay given by
N
N B P
s
Q
= max Q = i
max
i
C
i =1
i =1
i

(4)

Equations (3) and (4) show that the end-to-end delay strongly depends on the number of network nodes on the path between sender and receiver. As it is shown by the numerical example in
Section 4, queuing delay is the main contribution to the end-to-end delay4.
The current architecture of the Internet has been designed to carry data traffic with a best effort service. In order to limit loss due to congestion in network nodes, buffers are large, thus
leading to a large Qmax. Moreover, the number of hops on the path between sender and receiver
2

The time Ps/Ci a packet spends in the buffer while being transmitted on the output link has been already taken into account as transmission
delay.
3
Usually, the compensation delay is dimensioned with respect to some percentile of the delay variation, instead of considering its maximum
value. As a result, the bound is smaller, but the samples experiencing delay larger than the chosen percentile are discarded. This yields unpredictable QoS because the distribution of the delay over the samples is not known a priori. In this work, the compensation delay is dimensioned
according to the maximum delay variation RB; nevertheless, the obtained results still hold (with proper adaptations) also when the bound is
probabilistic.
4
Actually, in Equation (3) the term Qmax accounts twice being RB = Qmax because the processing delay is constant, the queuing delay is the only
variable component of the network delay, and Qmax is the maximum variation of the queuing delay (see Equation (2)).

ARCHITECTURAL CHOICES FOR PACKET SWITCHED TELEPHONE NETWORKS

ISS97 ID PS675

can be very large (even some tens). Many applications for voice transmission over Internet heuristically choose the compensation delay and possibly adapt it to the actual delay experienced by
samples. As a result the quality of a voice call is unpredictable: when the nodes on the communication path are lightly loaded the quality is acceptable. Instead, during heavy traffic periods, either the received signal is not intelligible or the end-to-end delay is long (up to 1 second) [7].
Nevertheless, we believe that a voice transmission service with a commercial like guaranteed
quality can be provided over a packet switched network provided that (1) it exploits some
mechanisms for supplying QoS guarantees, and (2) its topology be designed so that the path of
any voice connection (i.e., phone call) encompasses few nodes. In the following section we identify the key parameters for dimensioning such a network.
3 Designing a Packet Switched Telephone Network
Equation (4) shows that queuing delay depends on the ratio between buffer size and link capacity. Thus, the end-to-end delay can be decreased by increasing link capacity while keeping
buffer dimension fixed. Nevertheless, the probability of losses due to congestion in network
nodes is high if buffers are small with respect to link capacity. The problem can be overcome by
creating separate queues for best effort and voice traffic, the latter being given higher priority
than the former. In this scenario, Equation (4) still holds, Bi being the size of the voice queue
alone.
The voice queue can be made small and losses avoided by introducing a call acceptance control mechanism, which limits the maximum number Mi of voice connections routed through a
link. This determines the maximum number of packets present in the voice queue to
M
M i
i I
i

(5)

where Ii is the number of input ports of the ith node. The voice queue is dimensioned so that it
can contain the above amount of packets thus leading to a maximum queuing delay. Actually,
due to the variation of delay experienced in upstream buffers, more packets than the number expressed by Equation (5) can be present in a queue. Nevertheless a maximum total queuing delay
Qmax, as given by Equation (6), is not exceeded. Details can be found in [5].

M
M i
N i I i
.
Q
=P
max
s
C
i =1
i

(6)

Equation (6) shows that given the capacity of a link, the less phone calls are allowed on it, the
smaller the maximum queuing delay of the connections traversing the link. The overall voice
traffic routed on a link accounts only for a fraction of the link capacity, leaving to best effort traffic the remaining capacity. In order to provide an indication of the fraction of link capacity dedicated to voice traffic, we define the voice allocation factor i on link i, so that
P
s
C = M
i i
i P/ R

(7)

ARCHITECTURAL CHOICES FOR PACKET SWITCHED TELEPHONE NETWORKS

ISS97 ID PS675

where P is the number of bits sent in each packet5 and R is the rate of the voice stream, i.e., P/R
is the time elapsed between two subsequent packets of the same voice connection. Since Ps is the
size of packets, (Ps R)/P is the (gross) bandwidth used by the connection, taking into account the
overhead due to packet headers. Assuming that all the phone calls on the network exploit the
same coding (same bit rate R) and the same packetization technique (same Ps and P), the second
term of Equation (7) is the bandwidth used by telephone traffic on link i.
We define the packetization efficiency as

pkt

P
P
s

(8)

and use it to rewrite Equation (7):

C = M
i i
i

R
pkt

The packetization efficiency shows the fraction of link capacity wasted due to packet overheads
(i.e., the header).
When dealing with variable size packets having fixed length header (e.g., in IP networks), the
packetization efficiency is particularly relevant to the end-to-end delay. Increasing pkt requires
packets to be enlarged, thus increasing some of the delay components. This can be made explicit
by writing the packet size Ps in terms of the number OH of overhead bits
OH
P =
s 1
pkt
and substituting it in Equation (6) thus obtaining

M
M i
N i I i
OH
Q
=

max 1
C
pkt i = 1
i

(9)

4 Numerical Results
In this section we consider the implementation of a PTN and provide some numerical results
obtained by applying to it the equations devised in the previous section. The network topology is
designed after the Telecom Italias telephone network exploiting both IP and ATM technology.
In a circuit switched telephone network, telephones are connected through the local loop to local exchange offices which both have concentration functionality and encode analog voice signals into 64 kb/s PCM flows. Local exchange offices are connected through digital links to local
offices that are circuit switching nodes. In metropolitan areas, local offices are connected together and with toll offices which build up a higher layer of switching offices connected to each
other, as shown in Figure 1.
5

Throughout this work we assume P to be constant during a voice connection.

ARCHITECTURAL CHOICES FOR PACKET SWITCHED TELEPHONE NETWORKS

Primary Trunk

ISS97 ID PS675

Toll
Toll
Office
Office

Backup Trunk

Local
Local
Office
Office

Local
Local
Exchange
Exchange

Local
Local
Exchange
Exchange

Toll
Toll
Office
Office

Local
Local
Office
Office

Local
Local
Office
Office
Local
Local
Exchange
Exchange

Toll
Toll
Office
Office

Long Distance
Trunk

Interoffice
Trunk

Figure 1: Excerpt from the Topology of a Circuit Switched Telephone Network.

The Telecom Italias telephone network serves about 28 million users connected to about
11,000 local exchange offices which are, in turn, connected to 660 local offices. Local offices are
fully meshed within metropolitan areas and each one is connected to at least two toll offices. The
higher hierarchical level counts 60 toll offices connected in a full mesh. Finally, 6 international
gateways connected to a number of toll offices provide international connectivity. The end-to-end
delay of any call on the considered circuit switched telephone network, unless satellite hops are
present along the communication path, is well below the interaction bound (usually shorter than
20 ms).
We analyze the delay perceived by users of the PTN obtained by substituting switching offices
with packet switching nodes (IP routers or ATM switches). Each local exchange office is connected to a packet switching node. We assume that local exchange offices are the initiators of the
call acceptance control procedures when users dial calls. Moreover, they are responsible for the
voice packetization process. The capacity of the link between two nodes is determined according
to the number Mi of telephone calls routed on the corresponding link of the real network, i.e.,
C = M n 64 kb / s ,
i
i

(10)

where 64 kb/s is the bandwidth used for each phone call and n is a factor used to oversize the link
capacity in the PTN. Given the meshing of the Telecom Italias network and the capacity of the
links between nodes as obtained from Equation (10), a packet switching node is required to have
between 80 and 150 SDH STM-1interfaces6.
When computing the queuing delay according to Equations (6) and (9) (for an ATM and IP
network, respectively), we introduce an approximation by considering that all the hops on the
path give the same contribution (i.e., we assume that iQi = NQ). This approximation is not a
significant one since queuing delay basically depends on Ci/Mi (see [5] for more details) which is
constant because link capacity is dimensioned according to Equation (10)7. The value used for Mi
is the maximum number of circuits on the link between two toll offices, i.e., Mi = 2,000.
6

For this calculation the Synchronous Digital Hierarchy (SDH) is assumed to be used to connect packet switching nodes. Due to the dimension
of the problem, 155 Mb/s STM-1 carriers are the candidate links.
7
Actually, a further approximation is introduced: the granularity of real SDH carriers is not taken into account when dimensioning the links
among nodes. When actually building the network, an integer number of SDH carriers is installed between each pair of nodes in order to provide a capacity greater than the value computed using Equation (10). This yields a smaller value for , i.e., actual delays shorter than those
presented in this section.

1
0.8
0.6
0.4
0.2
0

1
0.8
0.6
0.4
0.2
0
400
Delay (ms)

300
200
100

0
0.7

0.8

0.3
1
0.8
0.6
0.4
0.2
0

300
200

0.7

0.8
1
0.8
0.6
0.4
0.2
0

400
Delay (ms)

100

300
200
100

0
0.4
0.5
0.6
Packetization Efficiency
Rate 16 Kb/s, 4 nodes

0.7

0.8

0.3
1
0.8
0.6
0.4
0.2
0

300
200

0.4
0.5
0.6
Packetization Efficiency
Rate 16 Kb/s, 6 nodes

0.7

0.8
1
0.8
0.6
0.4
0.2
0

1
0.8
0.6
0.4
0.2
0

400
Delay (ms)

400

Voice Allocation Factor

0.3

Voice Allocation Factor

Delay (ms)

400

0.4
0.5
0.6
Packetization Efficiency
Rate 32 Kb/s, 6 nodes

Voice Allocation Factor

0.4
0.5
0.6
Packetization Efficiency
Rate 32 Kb/s, 4 nodes

Voice Allocation Factor

0.3

Delay (ms)

200
100

100

300
200
100

0
0.4
0.5
0.6
Packetization Efficiency
Rate 8 Kb/s, 4 nodes

0.7

0.8

0.3
1
0.8
0.6
0.4
0.2
0

300
200
100

0.4
0.5
0.6
Packetization Efficiency
Rate 8 Kb/s, 6 nodes

0.7

0.8

400
Delay (ms)

400

Voice Allocation Factor

0.3

Delay (ms)

300

Voice Allocation Factor

Delay (ms)

400

Rate 64 Kb/s, 6 nodes

Voice Allocation Factor

Rate 64 Kb/s, 4 nodes

ISS97 ID PS675

Voice Allocation Factor

ARCHITECTURAL CHOICES FOR PACKET SWITCHED TELEPHONE NETWORKS

300
200
100

0
0.3

0.4
0.5
0.6
Packetization Efficiency

0.7

IP
ATM
n=1
0.8 n = 2
n=4

0.3

0.4
0.5
0.6
Packetization Efficiency

0.7

0.8

Figure 2: End-to-end Delay of a Long Distance Call.

ARCHITECTURAL CHOICES FOR PACKET SWITCHED TELEPHONE NETWORKS

ISS97 ID PS675

Finally, to devise the results shown in this section, we consider encoding techniques which
provide bit streams at 8, 16, 32, and 64 kb/s. The encoding schemes currently deployed introduce
delays ranging from a hundred of s up to hundreds of ms. Since there are techniques operating
at the same bit rate which introduce significantly different delays, the numerical results plotted in
the figures of this section have been obtained considering Dproc = 0.
Figure 2 shows the delay (left y-axis) of a long distance call routed through 4 (left column)
and 6 (right column) IP routers (star sign) or ATM switches (square sign). The delay is calculated
through Equation (3) for a range of packetization efficiencies (x-axis). Three cases are considered
for the capacity of links: n = 1 (dotted line), n = 2 (dashed line), and n = 4 (continuous line). Four
encoding rates for the voice signal are considered (one in each row). On the right y-axis the voice
allocation factor is plotted for each configuration. As shown by Equation (8), given a value for
the packetization efficiency, the voice allocation factor does not depend on the packet characteristics (i.e., it has the same value for IP and ATM).
In the first two rows (voice encoding at 64 kb/s and 32 kb/s) the delay is not plotted some of
the configurations for all or a range of the values of packetization efficiency. For example, in the
upper left picture, for a link capacity double than in the circuit switched network (n = 2, dashed
line), the delay corresponding to a packetization efficiency smaller than 50 % is not plotted. This
happens because when pkt < 0.5, more than half of the link capacity is wasted to carry packet
headers and thus there is not enough capacity to carry the same number of phone calls as in the
original telephone network.
In IP networks the delay grows more than in ATM networks as packetization efficiency increases. A cell is assumed to be sent as soon as its payload has been filled. This happens when
AAL1 [6] encapsulation is used. Packetization efficiency is varied in ATM networks by filling
only part of the cell payload8. This slightly varies the packetization delay, but does not change the
queuing delay which, according to Equation (6), depends on cell dimension. Audio samples are
assumed to be carried over IP networks using the protocol architecture shown in Figure 3. The
Real-time Transport Protocol (RTP) [8] provides timing relationship between sender and receiver. The protocol headers result in a fixed per packet overhead. Thus, in IP networks packetization efficiency is increased by enlarging packets which translates in an increase of both packetization delay and queuing delay, as reflected by the steep slope of the curves.
RTP
RTP
UDP
UDP
IP
IP
PPP
SDH
SDH

Figure 3: Protocol stack used to carry voice over IP networks

When exploiting ATM technology, the end-to-end delay can be kept below the interaction
bound (it is shorter than 100 ms in all the configurations). When IP technology is exploited, the
interaction bound is respected only when operating at very low packetization efficiency (less than
50 %). As a consequence, more than half of the links capacity is wasted for transmitting packet
headers. The voice allocation factor curves show that when packetization efficiency is high, a
8

This feature is not currently supported by AAL1.

10

ARCHITECTURAL CHOICES FOR PACKET SWITCHED TELEPHONE NETWORKS

ISS97 ID PS675

large fraction of link capacity is still available to best effort traffic (1 - ). Decreasing packetization efficiency reduces the amount of best effort traffic that can be carried by the network, even
though the number of phone calls supported is the same. Header compression [9] can be exploited to increase packetization efficiency while keeping packets small. This introduces higher
storage and processing requirements on IP routers but decreases the average amount of overhead
traffic generated on links. The effect of header compression on the end-to-end delay, link utilization (), and buffer occupancy are out of the scope of this work (see [5] for more details).
Lastly, speech compression leads to unacceptable delay when exploiting IP technology and
packetization efficiency higher than 50%. Using ATM, speech compression slightly increases the
end-to-end delay9, but leaves more capacity in the network for the transmission of best effort
traffic (the voice allocation factor decreases).
In any configuration, increasing link capacity substantially improves the performance because
queuing delay (inversely proportional to link capacity) is a major component of the end-to-end
delay. The extra capacity is not wasted as it can be used by best effort traffic. In addition, overdimensioning links appears to be sensible since bandwidth is an ever cheaper resource in both
links and node ports (when using packet switching technology). If the network carries guaranteed
traffic other than telephony, more complex queuing policies (e.g., weighted fair queuing [10])
should be used in nodes and link capacity should be dimensioned differently (see [5] for details).
General results remain almost the same.
5 Integrated IP and ATM Architectures for Packet Switched Telephone Networks
In this section we describe some architectural choices for a PTN based on the integration of IP
and ATM.
5 .1

Motivations

In the previous section ATM has been shown to be the best technology for telephony because
it offers end-to-end delays lower than IP. Nevertheless, it is strongly desirable for a PTN to be
equipped with IP forwarding capabilities for the following reasons:
IP is the most widely deployed network protocol in both the Internet and many intranets, i.e.,
lots of applications based on the services provided by the TCP/IP protocol suite are daily deployed on many platforms. Many of them, for example electronic mail, news and the World
Wide Web, are nearly ubiquitous.
Being an internetworking protocol, IP enables communications among networks based on
different technologies. On the contrary, native ATM applications require ATM to be exploited
along the whole path between communicating entities.
IP networks have a consolidated and widely supported management framework based on the
Simple Network Management Protocol (SNMP) and the related definition of Management
Information Bases (MIBs) for a variety of network devices.
Moreover, no particular advantage is given by the exploitation of ATM for high bandwidth
real-time applications (like, for example, videoconferencing and video on demand), with respect
to IP with QoS support. In fact, due to the high burstiness of the generated traffic, the gain in
statistically multiplexing long burst of small cells over statistically multiplexing large packets is
9

The increase in the end-to-end delay is due to the packetization delay: the lower the bit rate at the exit of the voice encoder, the longer the time
needed to fill the packet payload.

11

ARCHITECTURAL CHOICES FOR PACKET SWITCHED TELEPHONE NETWORKS

ISS97 ID PS675

not significant, and queues in ATM switches grow almost as large as in IP routers.
Recently, various proposals for the operation of IP over ATM networks have been made. They
aim at providing IP forwarding services while taking advantage of the high speed and, in some
cases, the QoS guarantees offered by ATM. Some of these proposals exploit ATM as a data-link
layer technology on which IP packets are transferred among routers (e.g., the Classical IP
Model). Others, like Tag Switching by Cisco Systems Inc.[11], IP Switching by Ipsilon Networks
Inc. [12], and Cell Switch Router by Toshiba Corp. [13] are good candidate foundations for PTNs
since they integrate IP and ATM mechanisms.
These integrated IP-ATM approaches aim at maximizing the exploitation of ATM switching
and minimizing the use of IP forwarding. This translates in transporting IP traffic over ATM
Virtual Connections (VCs) whose endpoints are as close as possible (possibly coinciding) to IP
endpoints. The use of ATM signaling and routing protocols is extremely limited, if not avoided
at all, i.e., only ATM switching functionality and possibly QoS guarantees are exploited.
The basic ideas common to the various integrated IP-ATM approaches may be summarized as
follows:
Integrated IP-ATM routers are based on ATM switching fabrics so that they can forward ATM
cells coming from either IP endpoints (hosts or routers) equipped with ATM interfaces or
other integrated IP-ATM routers.
IP routers somehow identify traffic flows between IP endpoints and autonomously decide
whether these flows are best served by classical hop-by-hop IP forwarding (e.g., e-mail, DNS
queries, etc.) or ATM cell forwarding, i.e., flows must be carried on dedicated VCs (e.g.,
multimedia traffic, file transfers, etc.).
When needed, integrated IP-ATM routers along the communication path create a dedicated
VC on the fly. IP packets are segmented into ATM cells at the source VC endpoint, forwarded cell-by-cell, and reassembled only at the destination endpoint.
In order to setup the dedicated VCs on which IP flows are transported, integrated IP-ATM
routers communicate among themselves (and possibly with IP endpoints equipped with ATM
interfaces) through simple and efficient service protocols. They allow integrated IP-ATM routers
to exchange information about IP flows and the associated VCs. Integrated IP-ATM routers setup
and tear down ATM VCs controlling directly their own switching fabric. The various integrated
IP-ATM approaches essentially differ in the way these service protocols operate.
If a PTN is based on one of the integrated IP-ATM approaches discussed above, telephony has
a quality (in terms of end-to-end delay) close to the one obtained over an IP network. Over an IP
network, voice samples are put into IP packets and forwarded by routers, as shown in Figure 4a.
Over an integrated IP-ATM network, IP packets containing digital speech samples are segmented
into ATM cells which are forwarded by the ATM switching fabrics to their destination, as depicted in Figure 4b. The IP like performances stem from the fact that (1) cells are sent into the
network in bursts of the dimension of an IP packet, thus affecting buffers like IP packets, and (2)
at the receiver IP packets must be reassembled before the contained voice samples can be played
back. Thus, speech samples cannot be played as soon as ATM cells arrive, but only when the last
cell of each IP packet has arrived. As a result, the delay of the whole packet is that experienced
by the last ATM cell (See [5] for further details).

12

ARCHITECTURAL CHOICES FOR PACKET SWITCHED TELEPHONE NETWORKS

5 .2

ISS97 ID PS675

Integrated Services IP-ATM Models

In order to get better performance from a PTN, we propose to devise integrated services IPATM models derived from the integrated IP-ATM approaches described above. The basic idea
behind integrated services IP/ATM is that once an end-to-end VC is in place, digital speech samples are encapsulated directly into ATM cells (Figure 4c), i.e., they are transmitted like over
AAL1. At the receiver, voice samples are played back as soon as the ATM cell carrying them
arrives. Current integrated IP-ATM approaches use the Unspecified Bit Rate (UBR) class of
service, i.e., they support only best effort traffic. On the contrary, telephony requires guaranteed
QoS. Since most speech encoding schemes produce a constant rate stream, integrated services IPATM routers must provide the Constant Bit Rate (CBR) class of service.
IP Router

Voice

IP Router

IP

Voice

IP

a) IP: speech samples are forworded hop-by-hop into IP packets.

Integrated
IP/ATM Router

Integrated
IP/ATM Router

Local
Local
Exchange
Exchange

Local
Local
Exchange
Exchange

IP Router

Integrated
IP/ATM Router

ATM VC

Voi ATM

ce IP ATM

ce IP ATM

b) Integrated IP/ATM: speech samples are transmitted end-to-end into IP packets

segmented into ATM cells.
Integrated Services
IP/ATM Router

Integrated Services
IP/ATM Router

Voice ATM

Voice ATM

Integrated Services
IP/ATM Router

ATM VC

Voice ATM

Voice ATM

c) Integrated Services IP/ATM: speech samples are transmitted end-to-end into ATM
cells.

Local
Local
Exchange
Exchange

Local
Local
Exchange
Exchange

Voi ATM

Local
Local
Exchange
Exchange

Local
Local
Exchange
Exchange

IP Packet carring voice

Figure 4: Forwarding models

In order for a phone call to immediately get its dedicated VC, integrated services IP-ATM
routers should be able to set up the VC without having to previously identify a traffic flow. Some
alternative solutions can be envisioned:
1. A signaling IP packet is used to announce the beginning of a phone call; all routers along
the communication path should react to it by setting up a VC dedicated to the phone call.
When a phone call has terminated, the corresponding VC must be torn down. This can be accomplished by means of a time-out mechanism (already exploited in some of the integrated
IP/ATM approaches). This mechanism can be quite reactive since while a call is active, traffic
is generated regularly.
2. Telephony applications can be written to exploit User to Network Interface (UNI) signaling
[14]. Integrated services IP-ATM routers must implement UNI signaling and Network Node
Interface (e.g., PNNI [15]) routing10. ATM signaling allows telephony applications to require
the network for the needed QoS, while best effort and multimedia IP traffic is still handled as
in integrated IP/ATM approaches. Moreover, native ATM applications (independent of their
10

These are usually implemented in into ATM switches.

13

ARCHITECTURAL CHOICES FOR PACKET SWITCHED TELEPHONE NETWORKS

ISS97 ID PS675

QoS requirements) are fully supported by the PTN. The main drawback of this solution is that
integrated services IP-ATM routers must run a considerably more complex and large software
with respect to the previous model.
3. The amount of software running on integrated services IP-ATM routers can be reduced by
exploiting the Integrated Private Network Node Interface (I-PNNI) routing protocol [16] to
carry routing information for both IP and ATM. Routers still support both UNI signaling and
service protocols exploited in integrated IP/ATM approaches, but IP routing protocols are not
run any more.
6 Conclusions
Voice transmission over packet switched networks is here studied from the point of view of
the quality perceived by the user in terms of delay. The results show that the Internet cannot provide good quality, except in particular conditions. Nevertheless, if a packet switched network is
expressly designed for telephony by (1) limiting the number of hops in the path between any pair
of users and (2) dimensioning link capacity properly, the end-to-end delay is low enough to allow
for interaction. Thus, packet switching could be exploited to build a commercial like telephone
network even though the raw capacity needed is larger than with circuit switching. This is justified by the possibility of carrying best effort traffic on the same network and by the lower installation and management costs.
The numerical results also show that ATM outperforms IP as was expected since the former
was expressly designed to support low bit rate real-time traffic. Nevertheless, because of the
large number of existing applications based on IP, it must be taken into account as a candidate
technology for carrying both best effort and high bandwidth real-time traffic in packet switched
telephone networks.
Thus, we propose that both IP and ATM be employed in the implementation of packet
switched networks for telephony based on integrated services IP-ATM models. These models are
derived from the integrated IP-ATM approaches currently being exploited for an effective operation of IP over ATM networks. Integrated services IP-ATM models provide ATM based services
for telephony and IP (over ATM) based services for any other kind of traffic (namely, best effort
and high bandwidth real-time) in order to get the best from the two technologies.
Another work shall be done in order to evaluate more carefully the performance of the proposed integrated services models.
Acknowledgments
We thank Prof. Silvano Gai for his advice and Agostino Vailati for the useful information provided. We thank Luca Fantolino for his important contribution.
References
[1] ITU-T, Recommendation I.327, B-ISDN Functional Architecture, March 1993.
[2] R. Braden, D. Clark, and S. Shenker, Integrated Service in the Internet Architecture: an
Overview, RFC 1633, Internet Engineering Task Force, July 1994.
[3] J. Woodard, Speech Coding, http://www-mobile.ecs.soton.ac.uk/speech_codecs/
[4] T. Robinson, Speech Analysis, http://squid.eng.cam.ac.uk:80/~ajr/SA95/SA95.html

14

ARCHITECTURAL CHOICES FOR PACKET SWITCHED TELEPHONE NETWORKS

[5]

[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]

ISS97 ID PS675

M. Baldi and D. Bergamasco and S. Gai, Telephony over Packet Switched Networks,
Technical Report, Politecnico di Torino - Dipartimento di Automatica e Informatica, December 1996 (http://www.netgroup.polito.it/publications/tr-pstn.ps).
ITU-T, Recommendation I.363, B-ISDN ATM Adaptation Layer (AAL) Specification,
March 1993.
R. Gareiss, Voice over the Internet, Data Communications, September 1996, pages 93100.
H. Schulzrinne, S. Casner, R. Frederick, V. Jacobson, RTP: A Transport Protocol for
Real-Time Applications, RFC 1889, January 1996.
S. Casner, V. Jacobson, Compressing IP/UDP/RTP Headers for Low-Speed Serial Links,
Internet Draft, November 1996.
A. Demers and S. Keshav and S. Shenker, Analysis and Simulation of a Fair Queuing
Algorithm, ACM Computer Communication Review (SIGCOMM'89),1989, pages 3-12.
Y. Rekhter, B. Davie, D. Katz, E. Rosen, George Swallow, Tag Switching Architecture
Overview, Internet Draft, January 1997.
P. Newman, G. Minshall, T. Lyon, L. Huston, IP Switching and Gigabit Routers, IEEE
Communications Magazine, January 1997.
Y. Katsube, K. Nagami, H. Esaki, Router Architecture Extensions for ATM: overview,
Internet Draft, October 1996.
The ATM Forum, ATM User-Network Interface Specification - Version 3.1, September
1994.
D. Dykeman and M. Goguen, Private Network-Network Interface Specification Version
1.0, March 1996.
R. Callon, Integrated PNNI for Multi-Protocol Routing, ATM Forum 94-0789, September 1994.

15