Cost/Benet Tradeo of Quality of Service

Mechanisms in Integrated Services Networks

Junseok Hwang

Martin B. H. Weiss

([email protected])
([email protected])
Telecommunications Program
Telecommunications Program
University of Pittsburgh
University of Pittsburgh
Pittsburgh PA 15260
Pittsburgh PA 15260
August 24, 1999
Abstract
Internet Protocol (IP) and Asynchronous Transfer Mode (ATM) are two switching technologies that
are being used (or proposed) for large-scale integrated services networks. Integrated services networks
are designed to support real time services, such as telephony as well as data services. Both technologies
have the potential to radically change the economics and operations of telephone services. In addition,
engineers have developed more ecient approaches than simple over-engineering to supporting the quality
requirements of real time services (such as voice).
In this paper, we will perform cost-benet (tradeo) study of guaranteeing the voice quality of packet
telephony of various possible QoS-support approaches (IP/over-engineering, IP/prioritization, IP/RSVP,
ATM-CBR, ATM-VBR). We will focus this application on integrated services networks. As with our
previous work [19], we will use a simulation model to engineer a large scale network and will then analyze
the switching and trunking costs. We will compare the costs of the dierent approaches needed to meet
a prespecied QoS requirement.
Keywords: IP Telephony, QoS, VoATM, IntServ, DiServ, RSVP, Engineering Economic Model.

1 Introduction
Many telecommunications carriers (RBOCS, CLECS, IXCs, and ISPs) are currently engineering their networks to support integrated services. To do so, they must consider the various quality requirements of
dierent applications, especially voice, which many consider to be an essential application, and which has
traditionally been carried over circuit switched networks. Recent developments have supported the commercial emergence of packet voice (packet telephony) with the advantages of statistical multiplexing gain and
lower costs. The technologies for supporting packet voice include Internet Protocol (IP), Frame Relay (FR),
and ATM (Asynchronous Transfer Mode).
IP telephony is a family of real-time voice applications over IP networks. Voice over IP (VoIP) is implemented using ITU Recommendation H.323 which denes the RTP/UDP/IP protocol stack for packetizing
1

voice into IP packets. Even though today's best-eort IP networks do not provide guaranteed service quality, telephony applications are rapidly launching into the Internet in various forms in all aspects of the
telecommunications market [13], [16].
One of the challenging issues for the successful implementation of IP telephony system is the delivery
of PSTN comparable QoS (Quality of Service). In this paper, QoS refers to a set of performance measures
associated with basic telephony services. We will focus on delay and delay variation; other possible factors
include packet loss and reliability.
Today, many consumers report poorer service quality for Internet telephony due to network delay as well
as limitations surrounding the PC [4]. A recent study on the measurement of network delay for various type
of the Internet connections by Maxemchuk et. al. [12] suggests that the current Internet is better suited for
local bypass than its current role as a replacement for long distance or international connections.
IP is not the only packet switching approach to support telephony. Voice over Frame Relay (VoFR) is
already in place especially for VPN enterprise applications and ATM was originally designed to meet the
diverse QoS requirements of the integrated service environment. Unlike current the IP network, Frame Relay
and ATM are connection-oriented protocols, so that the migration to support PSTN comparable telephony
is more straight-forward (in terms of QoS).
There are various QoS-Support options in these switching technologies, especially in IP and ATM. Recent
developments in IP supporting QoS (IntServ, and DiServ), and the newly standardized AAL2 (rt-VBR)
for telephony applications makes the comparison among these technologies more interesting.
A previous study by Weiss and Hwang [18] found that a pure IP telephony network provides lower switching and trunking cost than the PSTN using an over-engineering approach for voice-only applications. More
recent ndings on the extended study on this by Hwang [8] using RSVP (Resource ReServation Protocol)
suggests that the guaranteed service approach through the reservation may not be as ecient as circuit
switching (PCM) if only guaranteed trac (IP telephony) is carried by the network. This is generally consistent with Baldi et. al.[2]. For integrated service networks (not pure Itel), however, QoS approaches seem
to be advantageous according to the initial ndings from [8, 19].
In this paper, we will extend our previous work to include general QoS-mechanisms of various transfer
technologies, especially IP and ATM. The introductory part of this paper provides some background of
QoS problems for telephony applications, review some of the related and previous studies and discuss the
relevance of relating the cost analysis to the various QoS-mechanisms for the integrated service networks.
The second part will:


provide a brief technical overview of the two packet data networks for telephony applications and the
QoS-mechanisms associated with them,

discuss the motivations of QoS-support for various packet telephony systems, and

look at the quality and performance requirements for PSTN-comparable quality based packet telephony
services.

Various performance parameters for packet telephony applications will be reviewed. The third part will
describe the simulation model we implemented for the IP and ATM networks with various QoS-support
mechanisms. The fourth part will compare the IP with and ATM (Asynchronous Transfer Mode) based on
the simulation results found. Finally, we will develop cost models associated with the QoS mechanisms.
2

2 Technical Overview
In this section, we will review some of the relevant technical details for the comparison of two dierent
switching architecture for the telephony application.

2.1 IP Telephony System Architecture and Operation

Figure 1 shows the protocol stack typically used for voice communication and signaling in Internet telephony.
When a G.729A codec is used, all voice trac are vocoded from 64 Kbps PCM voice into an 8 Kbps
compressed signal. Like most of the low bit rate vocoders, G.729A utilizes silence suppression to further
decrease the required data rate. The compressed voice signal is then packetized using an RTP/UDP/IP
protocol stack. RTP typically runs over UDP to utilize its multiplexing and checksum services. RTP also
species the payload type to support multiple data and compression types.
Using a G.729A codec, voice is packetized into 10-byte voice packets every 10 msec and is buered to make
a 10 to 30-byte payload. The payload is then encapsulated with a 40-byte RTP/UDP/IP header. Combined
with the RTP/UDP/IP overhead, the suppressed codec output drives the actual average throughput to
around 14 Kbps. The RTP sequence numbers and time stamps are used to reassemble digitized voice
packets into a real time voice channel. This approach, however, does not provide any QoS guarantees. A
packetized voice packet is delivered to its destination through various layer-2 protocols to the access server
and multiplexers, and layer-3 switches to nally route the packet to its destination.
A signaling protocol for Internet telephony is essential to achieve a PSTN-comparable QoS performance.
There are several possible approaches to providing PSTN compatible signaling capability over the Internet
telephony network. The implementation (and, hence cost) varies with the dierent signaling requirements
for Internet telephony calls. In addition, the locus of the signaling function impacts the choice of the call
controlling protocol to be used. Three main call control signaling protocols for Internet telephony have been
suggested: H.323 Based Signaling [9], SIP (Session Initiation Protocol) [7], and MGCP (Media Gateway
Control Protocol) [1]. These protocols dier signicantly in terms of control, design, and functionality.
Dierent classes of internet telephony, connectivity options, WAN transport and main applications result in
biases toward dierent signalling protocols. Although quite interesting and important, a detailed discussion
of these is outside the scope of this paper.

2.2 VoATM System Architecture and Operation

The system architecture of Voice over ATM (VoATM) resembles VoIP except for the switching method.
Unlike IP, ATM is connection oriented cell switching and uses signalling dened in UNI 4.0.
To illustrate the operation of VoATM, let us rst assume that:


a native ATM terminal exists,

all the COs in the service area are mesh-connected with pre-congured PVCs (Permanent Virtual
Circuits), and

each subscriber establishes individual end-to-end SVCs (Switched Virtual Circuits) through those
PVCs.
3

Itel Call Signaling

H.225 and H.245

G729A Packet Voice

Data (10bytes)
(Packet Voice)
RTP Header
>= 12 Bytes

Data (10bytes)
(Packet Voice)

UDP
8 Bytes

RTP Header
>= 12 Bytes

Data (10bytes)
(Packet Voice)

UDP
8 Bytes

RTP Header
>= 12 Bytes

Data (10bytes)
(Packet Voice)

RTP/
RTCP

UDP

TCP

IP Header
20 Bytes

IPv4, IPv6

PPP

802.x
LLC/MAC

HDLC/
PPP

PPP

SONET

Ethernet
T. R.

ADSL

V.34

Figure 1: Example of RTP Protocol Stack (10 Byte Packetization) in the IP Telephony System
When a call is originated on the ATM network using a particular virtual circuit, the functions for AAL1 or
AAL2 and associated processing are performed at the network edge. QoS negotiation and Call Admission
Control (CAC) are performed at the edge ATM switches of each COs and all ATM switches in the network
performs trac shaping based on the negotiated QoS parameters both in the individual SVC level and in
the aggregated PVC levels among CO switches. Figures 2 and 3 show the CBR and rt-VBR implementation
of VOATM system, respectively.

Compressed 8 Kbps Voice on AAL1

1 Voice Channel / 1 VC

10 msec
G729A
(10bytes)

ATM Header
5 Bytes

SN Field
4 bits

SNP Fields
4 bits

G729A
(10bytes)

G729A
(10bytes)

G729A
(10bytes)

Padding

Figure 2: CBR Implementation of VoATM

2.3 Performance Requirements for Packet Telephony

The performance needs of various applications dier greatly. In particular, telephony applications require
relatively strict performance limits; in this section, we will discuss those requirements especially for packet
telephony network.
The problem of controlling delay and its variation on packet networks is one of the most important
4

Compressed 8 Kbps Voice on AAL2 with Multiplexing

Voice 1

G729A
(10bytes)

ATM Header
5 Bytes

Voice 2

G729A
(10bytes)

Start Field
1 Bytes

AAL2

G729A
(10bytes)

AAL2 Header
3 Bytes

Voice 3

G729A
(10bytes)

AAL2 Payload

G729A
(10bytes)

AAL2 Header
3 Bytes

G729A
(10bytes)

AAL2 Payload

Figure 3: rt-VBR Implementation of VoATM

performance issues of packet-based telephony. To achieve toll-quality service, the end-to-end delay should be
below a certain threshold. For PSTN-like quality, the ITU recommends an end-to-end delay of less than 150
milliseconds. Since delay is stochastic in the packet switched networks, this requirement may be interpreted
as the 99th percentile delay. In addition, the delay variation should be short enough to enable the dejitter
buer of the receiving phone to equalize the delay and disassemble the voice packet. The jitter requirement
may vary depending on the selection of the CODEC and voice packet encapsulation approach.
While packet loss is also an important issue, the human perception on the speech is more sensitive to
delay than loss. A carefully-sized receiving buer can recover the usable late voice packets on an Internet
telephony terminal, and a good buer management protocol is essential to avoid excessive unfair packet loss
due to the dierent sizes of packets. According to [6], the overall packet loss ratio of less than 1% would be
able to achieve the toll speech quality of MOS (Mean Opinion Score) around 4.0. We chose not to study this
problem in the present paper.
Call set-up delay, also called access delay or post dialing delay, is another important performance parameter for packet telephony. This delay is the value of elapsed time between a access request and successful
access and comprises the sum of call request time, selection time, and post selection time [6]. For the PSTN,
this call setup performance has been dened to be about 1-3 seconds and assumes SS7 signaling [5], [6].
Currently many Internet telephony applications use H.323 signaling on TCP, and the call-setup delay time
is not guaranteed.
The call blocking rate is one of the primary performance measures of circuit switched PSTN networks.
The main reasons for call blocking are the network resource congestion (switch blocking and trunk blocking),
and network failure. In general trac engineering practice, the average blocking probability of telephony
networks is less than 1%. Adequate performance estimate procedures and methods for QoS-support CAC
(Call Admission Control) should be implemented to support a blocking rate of less than 1%.
Dial-tone delay is another important performance measure which is closely dependent on the resource
allocation on the access switches or routers in the packet telephony network. This is the time interval
between the receipt of a request for dial tone at the local switching oce, or packet telephony access servers
as a result of o-hook, and the beginning of dial tone transmission to the user [6]. The performance level
5

required for dial-tone delay in ITU-T are less than 0.1 percent and less than 0.5 percent of call experiencing
more than 3 seconds dial tone delay user normal and high loads, respectively [10].

2.4 The QoS-Support Approaches of IP Telephony

QoS (Quality of Service) in Internet telephony can include voice delay and jitter, packet voice loss, voice
clipping, call-setup delay, echo, and loudness distortion. Network delay, which is the primary focus of
this paper, is dependent on the stochastic behavior of the trac with dierent sizes of the packet length
and dierent application requirements. Thus, there should be some approaches which allocate the network
resources to compensate for this stochastic behavior. In this section, we will deal with various approaches
which would mitigate or bound the such stochastic behavior of the network delay.
The Internet is a \best-eort" network, so QoS cannot be guaranteed. As real time applications, such as
Internet telephony, are introduced to the public Internet, QoS issues become more critical because of their
real time requirements. To control quality in IP telephony, properly designed IP switches, required QoS
signaling protocol support, classication of services, and appropriate dimensioning and trac engineering of
the network are required. Generally, the strategies available for achieving this involve over-engineering the
networks, using IP over ATM, using QoS signaling (such as RTCP and RSVP), or class of service dierentiation. We classied the possible approaches for providing adequate QoS in the IP telephony architecture
as follows;

Over-engineering The simple way of achieving required QoS is providing enough network resources to

avoid the high network loads that result in delays and delay jitter. This approach is widely used
through the LAN and private network environment to provide IP voice service over packet networks.
For a carrier level Internet telephony network, other QoS techniques which will be discussed below
would provide better network resource usage.
Resource Reservation: RSVP One way of QoS signaling is reservation of bandwidth on the Internet
for time critical services; this is implemented in the RSVP protocol by IETF. RSVP is a reservation
based signaling protocol designed enable the allocation of resources to support the QoS requirements
applications, such as bandwidth and delay. Since RSVP does not provide QoS-dependent routing,
other approaches, such as dierentiation of service, must be implemented independently so that the
variable delay component produced by router processing can be minimized.
Class of Services(Priority) Dierentiation or Classication of Services is another approach to provide
better quality of service and better utilization of the network resources. An Itel call can be assigned
to the high priority using the ToS (Type of service) eld of current IPv4 packet. Some of IP switch
vendors (eg., Cisco) use this approach to enable their switch to dierentiate the VOIP packet. We
also investigate the eectiveness of this approach in the following section. This approach is embedded
in IPv6, where four priority bits are introduced to support real time trac requirement through
prioritization. Unlike RSVP signaling, this dierentiation approach provides a mechanism for network
nodes to use dierent routing, packet scheduling, and queueing for dierent of type of services. Fast
and intelligent IP switch routers have the functionality of having dierent routing, scheduling, and
queuing techniques such as WFQ (Weighted Fair Queuing).

2.5 The QoS-Support Approaches in VoATM

ATM is a connection-oriented cell (xed size packet) switching and transfer technologies to support various
application services with dierent QoS-requirements. The xed short size cell allows network cell delay to
6

be predictable and controlled. ATM provides stronger QoS-support mechanism than IP and Frame Relay,
and supports various options for implementing telephony applications.
CBR (constant bit rate) and rt-VBR (real-time variable bit rate) classes are the ATM service classes to
be used for VoATM. CBR is currently the most common approach, using AAL1 to provide CES (Circuit
Emulation Service). ATM Permanent Virtual Circuits (PVCs) act like trunk lines in the PSTN network.
Switched Virtual Circuits (SVCs) are also possible if the network supports end-to-end ATM services. In CBR,
Peak Cell Rate (PCR) is used as a trac parameter and max CTD (Cell Transfer Delay), CDV (Cell Delay
Variation), and CLR (Cell Loss Ratio) are specied as the associated QoS parameters. Accepted CBR calls
which transmit the cells at or below the negotiated PCR will achieve the committed QoS performance from
the network. Since a CBR connection needs to exchange timing information between source and destination,
only a single user of the AAL can be supported on a single ATM connection.
rt-VBR trac is another ATM option to send voice applications when the source rate is expected to
be variable and bursty (using AAL2). In addition to the QoS parameters and PCR parameters specied
in the CBR, the additional trac parameters such as SCR (Sustainable Cell Rate) and MBS (Maximum
Block Size) are specied. More ecient bandwidth allocation can be achieved by AAL2 due to variable
rates and the support of silence suppression. rt-VBR also enables multiple user channels on a single ATM
VC connection. Variable payload size is allowed within cells and across cells which improves the protocol
eciency compared with other \Voice over X" protocols.

3 Simulation Model Description

We have developed the simulation model to compare the performance and cost of those various QoSmechanisms for the integrated service applications. In this section, we describe the parameters of the
simulation models.

3.1 Topological and General Technology Assumptions

To develop a simulation model, we assumed the service area with 1 million population which is equivalent
to the U.S. State of Rhode Island. To implement the realistic model close to the practical real network, we
have made following assumption:


Five large COs with core edge switches (ATM or layer 3 switches) are forming part of the SONET
Ring in the Service Area.

Additional edge switches are attached to the core switches with DS3 or Sonet interfaces, as illustrated
in Figure 4.

An average population density of 2.2 person per household uniformly distributed over the COs, and
each household has one telephone lines with 0.1 Erlang call density in the busy hours.

All the switches in the service areas are equipped its associated QoS-support functionalities.

We assumed the local loop interface with ADSL for both of the network topology.

All voice trac would be compressed from 64 Kbps PCM voice to 8 Kbps compressed voice using
G729A vocoder at the access network.
7

3.2 IP Network Model

The Itel simulation model assumed in our analysis is based on RTP/UDP/IP standardized on ITU H.323.
In this protocol, sequence numbers and time stamps are used to re-assemble the real time voice trac.
The simulation is concentrated on the IP layer 3 switching and trunking. The assumptions made for Itel
simulation model are specied below and in Table A-2.


Silence suppression will be enabled in each codec, with 60% of a session being silent each way.

On the suppressed codec output, RTP, UDP and IP overhead will make actual average throughput
around 9.6 Kbps (assuming 20 byte packetization).

Voice is packetized into a 10 byte voice packet every 10 msec and buered to make a 20 byte payload
from compression codec and encapsulated with the 40 byte RTP/UDP/IP header.

Voice is modeled as an on-o process, with an average 350 msec active state (exponentially distributed)
and a 650 msec exponentially distributed silence state.

Itel calls are modeled as connectionless UDP/IP sessions with exponentially distributed session lengths
of 240 sec.

For Prioritization QoS-support, the highest priority is given to the telephony application.

For RSVP, we calculated the required reserved bandwidth for each trunk for the telephony only scenario,
then calculated the worst case delay. Then we add the integrated service trac and tuned the simulation
to achieve the same quality of voice with RSVP.
Note that the parameters we chose for RSVP are very conservative (summarized in Table A-3. The
bandwidth required by RSVP is very sensitive to the choice of parameters, so that less conservative
parameters would result in reduced bandwidth requirements.

Using the recent trac data measured by [3], [15], and [11] on the Internet backbone OC-3 trunks, we
modeled the integrated service trac as a cross section of the Internet backbone trac, and computed
the intensity of this trac relative to Itel call demand.

3.3 ATM Network Model

In our VoATM simulation, we considered CBR (AAL1) and rt-VBR (AAL2) service categories in support
of telephony applications. For these services we assumed following conditions.


We assumed 20 bytes (20 msec of speech) of payload of AALs for both type of the services.

CBR does not support voice suppression and multiplexing of multiple AAL1s into a ATM cell.

rt-VBR supports voice suppression and multiplexing of multiple AAL2s into a ATM cell.

CBR and rt-VBR are given to higher priority than nrt-VBR, ABR and UBR.

The equivalent load of integrated service trac used in the IP model is translated into ATM trac.

Table A-4 summarizes the simulation parameters used in ATM simulation model.
8

4 QoS Mechanism Simulation Analysis

In this section, we present the main results of the simulation for the various types of QoS-mechanisms for
the ISN scenarios. The objectives of this simulation were:
1. to compute the required switching and trunking capacities while maintaining adequate QoS, and
2. to compute the 99th percentile delay for voice packets when integrated services trac is added or when
the number hops is increased.
All simulations were performed using Comnet III.
We assumed the base ISN trac scenario to be one where the the trac load of voice and data are
identical in terms of byte throughput (9091 erlangs of voice trac and 128 Mbps of data trac). The
networks were engineered to carry this load based on maintaining acceptable voice delays. To \stress" this
model, we added additional data (i.e., integrated service) trac until the QoS for voice exceed the delay
tolerance (without making any additional capacity \investments"). Figure 4 represents the simulation model
for the QoS sensitivity of various QoS-mechanisms over various ISN trac load.

4.1 QoS Sensitivity on Incremental Data Trac

Figure 5 illustrates the results of the simulations for each QoS-mechanism. The abscissa represents relative
throughput of the integrated data trac. The ISN baseline scenario represents the case when the throughput
of IP telephony and data trac are equivalent (at value of 1 of the abscissa). The ordinate represents 99th
percentile variable delay of packets or cells, which includes queuing delay and serialization delay on the core
edge switches. At the relative integrated service load level of \0", the networks are pure voice networks and
the performance of the approaches are indistinguishable, largely due to the low network utilizations at this
level of trac intensity.
The results show that ATM/VBR and ATM/CBR are statistically indistinguishable from each other and
provide the best performance of all the QoS technologies we considered. With ATM/CBR scenario, the
network would not be able to utilize more trac after relative ISN load scale of 2 due to the absence of
silence suppression and multiplexing implemented in AAL1 (for VBR). The best-eort and prioritization
schemes of IP worked until the point of ISN scale 1 and 1.5 respectively. For the prioritization, simple
non-preemptive prioritization is implemented.
Since ATM was designed with integrated services networks in mind, it is not surprising that they performed the best. That ATM/VBR was able to carry such large integrated services trac loads implies
that carriers can add a lot of non-voice trac without having to make signicant network upgrades, which
would result in a cost advantage to those carriers. For IP-based networks to carry similar loads at the same
performance levels would require additional investments in switching and trunking capacity.

4.2 QoS Sensitivity on Multi-Hop-Networks

Delay in packet networks is known to be dependent on the number of hops that a packet traverses. Thus,
we performed more simulation experiments to add a bit more realism to this study, as most packets traverse
more than one hop.
9

9091 Erlangs
Voice
IP Router
ATM Switch

OC3

9091 Erlangs

OC3
OC3

VOIP /
VOATM

Voice

Data
9091 Erlangs
Voice

OC3
OC3

Data

HTTP IP Router
FTP ATM Switch
DNS/UDP...
nrt-VBR
ABR
UBR

IP Router
ATM Switch
Data

OC3

OC3

OC3
OC3

9091 Erlangs

9091 Erlangs

Voice

Voice

OC3
IP Router
ATM Switch

Data

IP Router
ATM Switch

Data

Figure 4: Simulation Model for QoS Sensitivity

99th Percentile Variable Delay of Voice in msec.

60

50

40
CBR
rt-VBR
IP-RSVP
IP-Prioritization
IP-Overengineering

30

20

10

0
-1

1
2
3
4
Relative Integrated Service Data Traffic Load

Figure 5: 99th Percentile Variable Delay Trends for Various Trac Load
Data

Data
Voice

128 Mbps
Average
Throughput

Voice

128 Mbps
Average
Throughput

9091 Erlangs

9091 Erlangs
Routers/
Switches
Routers/
Switches

Routers/
Switches
Data

Routers/
Switches

Voice

128 Mbps
Average
Throughput

Routers/
Switches

9091 Erlangs
Data
Voice
9091 Erlangs

Data

128 Mbps
Average
Throughput

Voice

Multi-Hop Base Line ISN

Scenario Simulation Model

128 Mbps
Average
Throughput

9091 Erlangs

Figure 6: Simulation for Multihop Sensitivity

10

As it is shown in Figure 6, we consider 5 backbone core switch nodes which are spaced at 100 km evenly.
We assumed ISN baseline trac load to be oered to each of the core switches and to be traversed various
number of hops in the model. By monitoring the QoS measures of voice trac which traversed various hops,
we can nd that the equivalent utilization of the trunks over various QoS mechanisms provides acceptable
QoS. The only exception is the IP/over-engineering case where utilization is around 40%.
ISN-BaseLine Scenario : 70%--80% Engineering Utilization

99th Percentile Variable Delay of Voice in msec.

150

100
CBR-3*OC3
rt-VBR-2*OC3
IP-RSVP- OC12 + 2*OC3
IP-Prioritization-OC12
IP-Overengineering-2*OC12
IP-BestEffort-OC12
50

1.5

2
2.5
3
Number of Backbone Hop Traversed

3.5

Figure 7: multi-hop sensitivity of ISN-Base line scenario

From the results of this simulation summarized in Figure 7, the CBR is found to be the least sensitive
to the number of hops traversed (as expected). IP prioritization was roughly equivalent to rt-VBR after
multiple hops are traversed. One of the major ndings from this simulation is that each QoS-mechanism
shows dierent bandwidth eciency, which aects the cost of the network (discussed below).

5 Cost Analysis
Validating the technical comparison of the various QoS technologies was not the goal of this research {
understanding the implications on switching and trunking cost was1. With designs in hand, we consulted
with vendors to review the \reasonableness" of the design and to estimate the cost of the switches or routers.
The cost of the transmission links is based on leased line costs from AT&T with bulk price discount rate of
50% and 90% (to assess the sensitivity of the costs on trunking).
We considered ve backbone switch nodes which are spaced in 100 km evenly. The cost of the switching
technology is composed of two parts; initial investment costs and yearly recurring costs. Most of the switching
equipment in a CO will be considered the initial capital investment costs and transmission links (OC-3, etc.)
will be considered the recurring costs. We assumed the product life time of 3 years for switching equipment
and 10 % of MARR (Minimum Attractive Rate of Return) to calculate the NPV (Net Present Value) of
trunking costs. The calculated cost of each QoS mechanism is illustrated in Figure 8.
Figure 8 illustrates the costs of the various implementations discussed in this paper. These costs are
clearly in uenced very heavily by trunking costs, which were calculated to be a 50% discount o of AT&T's
1

We assume that the access networks are identical for all technologies, so we do not include them in our analysis.

11

NPV in
$(Thousands)
30000

Total Cost
Transport Cost
Switching Cost

20000

28888
27300

22738
21440
10000

13660

13005

14733

11670
9115
7780

1335

1335

rt-VBR
ATM

1588

1298

1073

CBR
ATM

Prioritization
IP

RSVP
IP

Overengineering
IP

Figure 8: NPV Comparison of Switching and Trunking Costs among Various QoS-Mechanisms
retail rates. Table 1 compares the 50% and 90% discount rates. Even at the lower transmission costs, the
same end results hold:


ATM/VBR is the lowest cost. This is even stronger given the ability of this technology to carry much
more integrated service trac at this level of capacity without signicantly degrading the packet delay.

Especially at the lower trunking cost, the costs of both ATM technologies and IP/Priority appear to
be roughly equivalent, as do the costs of IP/RSVP and IP/Over-engineering.

Technology

ATM/VBR

Cost

Transport Cost
Total Cost
ATM/CBR
Transport Cost
Total Cost
IP/Priority
Transport Cost
Total Cost
IP/RSVP
Transport Cost
Total Cost
IP/Over-engin. Transport Cost
Total Cost

50% Discount 90% Discount

$7.78M
$9.12M
$11.67M
$13M
$13.66M
$14.73
$21.44M
$22.74M
$27.3M
$28.89

$1.56M
$2.91M
$2.33M
$3.67M
$2.73M
$3.80M
$4.29M
$5.59M
$5.46M
$7.05M

Table 1: Sensitivity of NPV costs to Trunking Costs

12

6 Summary: Cost and Benets

The purpose of this paper has been to explore the implications on transmission and switching cost of various
QoS support technologies for a relatively large network. We have shown that the transmission and switching
costs of even simple QoS technologies (such as prioritization in IP) can be substantial as compared with
over-engineering. We previously reported that, using the same basic simulation topology, that IP/Overengineering resulted in trunking and switching costs that were about 50% lower than an equivalent circuit
switched network [18]. This paper claims that this cost advantage can be improved further by adding QoS
technologies.
The benet side of this scenario is more compelling. QoS support technologies allow carriers to support
more integrated service (i.e., data) trac when they deploy QoS support technologies in their networks
without upgrading the transmission and switching capacity. Circuit switched carriers would have to build a
separate infrastructure, incurring more capital and trunking costs, to provide equivalent integrated services
trac (see [19] for a further analysis).
From an engineering perspective, the technical results are not surprising, as that is precisely what QoS
support technologies are supposed to do. What was surprising to these researchers was that cost of ATM
switching technologies was comparable to the IP technologies. This was not the case a few years ago2.
From a policy perspective, it illustrates that, depending on the technology, the incremental costs of data
trac can be quite low. This may have substantial bearing on interconnection negotiations, as those have
been framed based on these incremental costs. Many of these costs would be \forward-looking," as they
represent are next generation technologies for telecommunications service providers.
There were a number of signicant costs that we did not consider explicitly in this paper. We assert that
these omissions are signicant only to the extent that the dierences in costs between these technologies are
signicant; in many cases, we claim that they are roughly similar, as we are comparing packet technologies
with each other and are not comparing them to circuit switching


One was the cost of gateways and converters. This is typically a large percentage of the capital
cost of a \Voice over X" carrier; since we were comparing only these types of carriers, and were not
comparing these costs to circuit switching, the signicance of this cost is conned only to dierence in
gateways/converters3.

We also did not explicitly account for call processing systems and support. We assert that the only
signicant bearing of this omission applies when the cost dierences are signicant between the various
packet technologies. Signalling for \Voice over X" is an area that is currently evolving very rapidly. It
is dicult to get reliable cost estimates for signalling and call processing components.

Finally, and perhaps most signicantly, we did not include operations costs. We believe that this is
an area where the dierences between the technologies are potentially large. It is often the case that
multi-QoS capability requires more engineering and management eort than single-QoS capability does.

2 This does not imply that operating costs of ATM are comparable to IP operating costs. We have no information on the
question of operating costs.
3 For devices using vocoders, this cost dierence is likely to be small, since they would all use Digital Signal Processors
(DSPs). This dierence could be signicant when comparing ATM/CBR to either ATM/rtVBR or IP, since CBR normally
packs PCM samples into ATM cells, avoiding the need for vocoding.

13

We did not analyze the implication of interconnection among carriers. Maintaining QoS across dierent
administrative domains in packet networks is quite dicult. Various solutions to this, such as assigning
reservations costs (see, generally, Peha et.al. [14, 17] seem to be appropriate to circumvent opportunistic
behavior on the part of competitors4 .
As current telecommunications carriers including begin deploying packet-based voice technologies in their
networks, they are likely to be increasingly interested in various QoS mechanisms. The set that we have
studied here are all \pure" implementations. Current engineering discussion is on \hybrid" approaches that
are able to adapt to large scale networks. Such hybrids include carrying IP trac over ATM channels to
improve quality, and using dierent technologies in the edge networks than in the backbone networks to
improve scalability. Examples of the latter include RSVP-DiServ-RSVP, RSVP-IP/ATM-RSVP, IP/ATMDiServ-IP/ATM, and DiServ-IP/ATM-DiServ, where DiServ is essentially the IP/Priority approach
discussed in this paper, and IP/ATM is IP over ATM.
These hybrid techniques are likely to occur in various forms when networks are interconnected but under
diering administrative domains. Technically, engineers have to solve several problems for such interconnected networks, including:
 How to deal with potential QoS losses as packets traverse networks using diering QoS technologies;


How to develop techniques that scale to large networks, high speed networks, networks with large
numbers of users, networks with high connection setup/teardown rates, etc ; and

How diverse networks can perform basic (eg., call setup/teardown) functions and more sophisticated
functions (eg., QoS signalling).
These are dicult problems that have potentially signicant cost implications, none of which we explored in
this paper.


References
[1] Arango, M., Dugan, A., Elliott, I., Huitema, C., and Pickett, S. Media gateway control
protocol (MGCP). Internet Draft, Internet Engineering Task Force, February 1999. Work in progress.
[2] Baldi, M., Bergamasco, D., and Risso, F. On the eciency of packet telephony. In In Proceedings
of the 7th International Conference on Telecommunications Systems (Nashville, TN., March 1999).
[3] Claffy, K. C., Miller, G., and Thompson, K. The nature of the beast: recent trac measurements
from an internet backbone. http://www.cacida.org/Papers/Inet98/index98, April 1998. Proceedings of
INET'98.
[4] Clark, D. D. A taxaonomy of internet telephony applications. Internet Telephony Consortium (1997),
51.
[5] Duffy, F. P., and Mercer, R. A. A study of network performance and customer behavior during
direct-distance-dialing call attemps in the U.S.A. Bell System Technical Journal 57 (January 1978),
1{33.
4 For example, a carrier could reserve large pieces of capacity on their competitors network that they have no intention of
using. This would eectively raise the cost of the competitor if adequate performance is to be maintained.

14

[6] Gruber, J. G., and Le, N. H. Performance requirements for integrated Voice/Data networks. IEEE
Journal on Selected Areas in Communications SAC-1, 6 (December 1983), 981{1005.
[7] Handely, M., Schuzrinne, H., and Schooler, E. SIP: Session initiation protocol. Internet Draft,
July 1997.
[8] Hwang, J. Guaranteeing the Delay-Bound of Internet Telephony in the Best-Eort IP Networks: A
Tutorial. Technical Report, University of Pittsburgh, March 1999. Work in progress.
[9] International Telecommunication Union. Visual telephone systems and equipment for local area
networks which provide a non-guaranteed quality of service. Recommendation H.323, Telecommunication Standardization Sector of ITU, Geneva, Switzerland, May 1996.
[10] International Telecommunication Union (ITU)-T. Specications of requirements of a switching
system. Recommendation GAS 6, ch. VI, Telecommunication Standardization Sector of ITU, Geneva,
Switzerland, 1981.
[11] Internet statistics and metrics analysis:
Engineering data and analysis.
http://www.caida.org/ISMA/isma9809/report.html, September 1998. Workshop Report of ISMA'98.
[12] Maxemchuk, N. F., and Lo, S. Measurement and interpretation of voice trac on the Internet.
In Conference Record of the International Conference on Communications (ICC) (Montreal, Canada,
June 1997).
[13] Mier, E. Voice-over-IP: Getting started. Business Communications Review 28, 5 (May 1998), 18.
[14] Tewari, S., and Peha, J. Competition among telecommunications carriers that oer multiple services.
Presented at the Telecommunications Policy Resarch Conference, October 1995.
[15] Thompson, K., Miller, G. J., and Wilder, R. Wide-area internet trac patterns and characteristics. http://www.vbns.net/presentations/papers/MCItrac.ps, December 1997. An abridged version
appears in IEEE Networks, November/December 1997.
[16] Tracey, L. Voice over ip: Turning up the volume. Telecommunications 32, 3 (March 1998), 28.
[17] Wang, Q., Peha, J., and Sirbu, M. Dynamic pricing of integrated services networks. Presented at
the Telecommunications Policy Resarch Conference, October 1995.
[18] Weiss, M. B., and Hwang, J. Internet telephony or circuit switched telephony: Which is cheaper?
http://www2.sis.pitt.edu/ mweiss/papers/itel.pdf, September 1998. Presented at the 26th Telecommunications Policy Research Conference, Alexandria VA.
[19] Weiss, M. B. H., and Hwang, J. Internet vs. Circuit Switched Telephony: Cost and QoS of Large
Scale Integrated Service Networks. Technical Report, University of Pittsburgh, November 1998. Work
in progress.

15

Appendix A: Detailed Modelling Parameters

Attributes
Trac Parameters:

CBR

rt-VBR

PCR, CDVT
SCT, CDVT, MBS

Specied
NA

Specied
Specied

Peak-to-Peak CDV
Max CTD
CLR

Specied
Specied
Specied

Specied
Specied
Specied

QoS Parameters:
QoS Requirements:

Low CDV
Moderate CDV
Moderate Loss Moderate Loss

Table A-1: ATM Layer Service Category Attributes for Telephony Application

Itel Network Parameters

Values

Packet Switch
IP Switch (>20 Gbps and 2 MPPS)
Fraction of outgoing call
0.1
Originated Trac per line
0.1 Erlangs
Packet Payload Size
20 bytes
Protocol Overhead
40 bytes (RTP/UDP/IP)
Packet Voice Burst Distribution burst 350 msec, silence 650 msec
Table A-2: Assumptions for Itel Simulation Model

16

RSVP Parameters

Values

Tspec

Peak IP datagram rate P

Token Bucket Rate r
Token Bucket Size b
Maximum Voice Packet Size M
Minimum Policed Unit m

24 kbps
24 kbps
60 bytes
60 bytes
60 bytes

Average Number of Hops H

Total C
Total D

5
300 bytes
20 msec

DTerminal
DReassembly
Dpropagation
Vqueuing

45 msec
50 msec
20 msec
35 msec

Adspec

Delay Budgets

Rspec

Required Reserved Service Rates R 82.3 Kbps

Table A-3: RSVP Parameters: 20 Byte Packetization with DS0 Line

VOATM Network Parameters

General
Cell Switch
ALL Payload Size

Values
ATM Switches (>20 Gbps)
20 bytes (G.729a)

Rate Control Parameters for Voice at the Core Switches

PCR (CBR)
PCR (VBR)
SCR (rt-VBR)

910 Kcells/sec
453 Kcells/sec
184 Kcells/sec

PCR (nrt-VBR)
SCR (nrt-VBR)
PCR (ABR)
MCR (ABR)
PCR (UBR)

354 Kcells/sec
236 Kcells/sec
59 Kcells/sec
10 Kcells/sec
120 Kcells/sec

Rate Control Parameters for non-Voice at the Core Switches

Table A-4: VOATM Network and Trac Control Parameters for Baseline ISN Network

17

Integrated Trac Parameters

Values
HTTP/TCP: From Server to Client
Percentage of Packets
Percentage of Bytes
Percentage of Flow
Average Packet Length
Major Packet Length
Packets per Flow
Average Flow Duration
Average Size per Flow

38
70
35
791 bytes
1500, 40, 552
14-18 packets
10-15 seconds
11 KBytes

Percentage of Packets
Percentage of Bytes
Percentage of Flow
Average Packet Length
Major Packet Length
Packets per Flow
Average Flow Duration
Average Size per Flow

38
8
35
83 bytes
40
14-16 packets
10-15 seconds
1 KBytes

Percentage of Packets
Percentage of Bytes
Percentage of Flow
Average Packet Length
Major Packet Length

9
17
5
600 bytes
40, 1500

Percentage of Packets
Percentage of Bytes
Percentage of Flow
Average Packet Length
Major Packet Length
Packets per Flow
Average Flow Duration
Average Size per Flow

5
2
15
165 bytes
40
2-3 packets
15 seconds
500 Bytes

Percentage of Packets
Percentage of Bytes
Percentage of Flow
Average Packet Length
Major Packet Length
Packets per Flow
Average Flow Duration
Average Size per Flow

10
3
10
401 bytes
40, 1500
50 packets
20-30 seconds
21 KBytes

HTTP/TCP: From Client to Server

FTP and Other TCP

DNS/UDP

RTP/UDP

Table A-5: Summary of Integrated Trac Parameters

18