CHARGING SCHEMES FOR MULTISERVICE NETWORKS

David Songhurst and Frank Kelly

1. Introduction
Usage-sensitive charging schemes aim to ensure that connection charges properly reflect the load imposed on
the network. In nnultiservice networks the bandwidth and quality of service requirements of different connections
can impose greatly differing loads. With usage-sensitive charging customers have an incentive to prioritise their
traffic appropriately, and the network operator’s revenue should match the investment required to provide
capaclty to carry the traffic.
In a multiservicie environment the balance between simplicity and complexrty needs to be carefully drawn.
Charging and billing may form a large part of the total cost of a network, increasing the appeal of schemes
based solely on access and subscription charges. But the heterogeneity of traffic characteristics and quality-of-
service requirerrients increases the difficulty of network management and control; and, in networks where
intelligence is distributed to customers and their equipment, usage-sensitive charging schemes may, through the
incentives they provide, improve the stability and robustness of the entire system, comprising both network and
customer equipment.
The development of charging schemes raises hdamental questions about how to characterise network load, to
what extent should customers be involved in declaring their traffic load and priorities, and how significant are
usage-related costs relative to ather fixed network costs over various time-scales.
Several charging schemes have been proposed for the lntemet and for ATM networks. Approaches that have
been studied include iterative distributed pricing schemes, and schemes based on effective bandwidth. A ‘smart
market’ approach has also been proposed where packet prioritisation and pricing are based on user bids. These
schemes aim to be incentivetompatible - that is they provide the right incentive for customers to declare their
true priorities.
The European A.CTS project CA$hMAN (Charging and Accounting Schemes in Multiservice ATM Networks)
is studymg these: issues. CA$hMAN aims to identie good candidates for ATM c h a r p g schemes, develop the
necessary hardware and software facilities to implement them in user trials, and to use these trials to gain
experience of implementation issues and to acquire user feedback.
In section 2 of this paper we discuss the important issues attached to usage-sensitive char-g and summarise
the research on this subject.
In section 3 we review the concept of effective bandwidth and present a charging scheme suitable for delay-
sensitive connections, where statistical sharing is important over short time-scales, comparable to or less than
round-trip delay times across the network. We assume that such sources are policed, for example by leaky
bucket regulators of the sort used to define the peak or sustainable cell rate parameters of an ATM traffic
contract 1191. lihese policing parameters provide upper bounds on the behaviour of sources, but may not
characterize sources well: for example sources may only occasionally need to burst at rates close to the bounds.
We describe charging and connection acceptance mechanisms based on the traffic produced by a source, as well

Lyndewode Research
Lyndewode-Re!;earch@lyndewode.co. uk

0 1996 The Institution of Electrical Engineers.

411 Printed and published by the IEE,Savoy Place, London WCZR OBL, UK.
as any agreed policing parameters.
In section 4 we consider delay-tolerant traffic. Statistical sharing is somewhat simpler for such traffic, since
flows may be coordinated over time periods longer than round-trip delay times across the network, and since
delay itself is available to convey control information. On the other hand the aims of pricing may be more
complex, and may include providing feedback to users on network congestion and the revelation of user
valuations (as discussed in [l]). We describe a simple scheme based on measurements of time and volume that
can again achieve most of what could be expected of any scheme.
In section 5 we outline a unified pricing model suitable for a wide range of quality of service classes, described
in terms of static parameters fixed for the duration of a connection and dynamic measurements of time and
volume.
Finally in section 6 we describe how this unified pricing model is being implemented and tnalled in project
CA$hMAN.

2. Usage-sensitive charging
2.1 Services and charging in multiservice networks
ATM is being developed as the technology for broadband multiservice networks. The ATM Forum [20] has
proposed classes of network services, dependmg on whether the information flow is sensitive to delay and
-
information loss Constant Bit Rate, Variable Bit Rate (real-time and non-real-time), Available Bit Rate, and
Unspecified Bit Rate. ITU standards [19] define various transfer capabilities, including Deterministic Bit Rate,
Statistical Bit Rate and Available Bit Rate.
In contrast to this mix of guaranteed and best-effort services the Intemet currently provides all services on a
best-effort basis. The services are likewise undifferentiated by tariff and in general end users pay only for
access and not for usage. This flat-rate charging has undoubtedly been an important factor in stimulating traffic
growth and development of new applications, but it will no longer be tenable in an environment of rapid
commercial growth and increasing demand for high bandwidth services. In fact a number of new Intemet
service definitions are currently under study [2]. These provide for a range of qualrty of service specifications
wtth differing levels of guarantee on packet loss and delay, includmg Guaranteed, Controlled-Load, Predictive,
Controlled Delay and Committed Rate. This work is also addressing how these service definitions can be
mapped onto ATM services to assist interoperability between IP and ATM.
The efficient operation of multiservice networks requires clear hfferentiation of services and the use of charging
schemes that are usage-sensitive and reflect the service being offered. Usage-sensitive charging offers a number
of benefits:
0 It gives users an incentive to declare their priorities and constrain their traffic demands appropriately.
It allows the provision of a range of service classes with appropriate cost recovery.
0 Overall charges reflect network costs, thus giving network operators the right incentive to upgrade capacrty.
0 It can moderate undue incentives for resale of capacity to third parties.
In practice telecommunication charges have generally contained several components - flat-rate charges such as
service access fees and subscriptions, usagedependent charges, and other dependencies (for example, distance,
time-of-day, quality of service). In multiservice networks the usagedependent component can become much
more complex, making the whole tariff structure unacceptably complicated. There is a need to consider
carefully the requirements of all parties involved, including customers, network operators, service providers,
and suppliers of terminal equipment and application software.
CA$hMAN surveys have found that customers tend to like flat-rate charging because it is simple and their
charges are predictable. Usage-sensitive chargmg may be acceptable to the extent that it is understandable and
controllable, but customers are less likely to accept congestion-sensitive charges which they cannot control.

412
However usagesensitive chargmg schemes enable the network operator to operate the network efficientlywith a
range of diffenntiated services. Hence the challenge is to develop charging schemes that adequately reflect
resource usage while being sufficiently simple for user acceptability and costetlkctive implementation.

2.2 Proposed charging schemes

A number of researchers have proposed distributed iterative schemes for setting prices m order to -se
resource allocation [3,4, 51. These schemes use repeated optimisation calculations in order to update capacity
prices, and they generally as- that users respond to changing prices according to some benefit function
which determines how their b a n h d t h requirement increases or decreases depending on price.
MacKie-Mason and Varian [6] discuss the concept of the “smart market”. The intention is that the usage price
should be zero when the network is unwngested, but when there is congestion the price should reflect the
incremental social cost determined by the marginal delay cost to other users. At the equilibrium price the user’s
willingness to pay for additional data packets equals the marginal increase m delay cost generated by those
packets. This is implemented by requiring all users to set bids on their packets. At cangested routers, packets
are prioritised b,asedon these bids. Users are not charged the price that they bid but rather the bid of the highest
prionty packet not admitted to the network.
The approaches: mentioned above are suited to besteffort services. For guaranteed services an additional
problem is to ccimbine pricing wah resource allocation in such a way as to enable statistical sharing. One way
to achieve this i.s to use effective bandwidth to characterise user traffic [7, 8, 91. A problem here is that the
ATM traffic corltract information may n a provide a good estimate of efEctive bandwidth. Courcoubetis et a1
[9] propose the use of on-line measurements to estimate effective bandwidth. Another approach, described in the
following sections, uses charges based on bounding tangents to the effective bandwidth function, with the
possibility of gaining increased information through the user’s choice of tariff.

3. Charging; based on effective bandwidths

3.1 Effective lbandwidths
In this section we review the concept of an effkctive bandwidth, beginning wrth a simple model of statistical
sharing at a tranmission, switching or other scarce resource.
Suppose that J slources share a smgle resource, and let Xj(0.r ) be the load produced by source1 in the interval
(0.0.Assume that (xi(O,t), t >O) has stationary increments, and that the loads produced by different sources are
independent. Can the resource cope wrth the superposition of the J sources? The answer to this question is, by
now, quite well .understood. For quite general models of sources and resources it is possible to associate an
effective bandwidth ullth each s a u c e sucb that, provided the sum of the effective bandwidths of the sources
using a resource is less than a certain level, then the resource can deliver a performance guarantee. The relevant
definition of the effective bandwidth is generally of the form

ai(,t) = (st)-! log E [exp(s xj(0,t))l (1)

for particular choices of s and t. There may be several constraints on effective bandwidths corresponding to
dif€erent physical or logical resources within a network, and to priority disciplines [lo, 111.
To illustrate the calculation of an effective bandwidth, consider the very simple case of an d o f f source which
produces workload at a constant rate h while in an ‘on’ state, and produces no workload while in an ‘off state.
Suppose the periods spent in ‘on’ and ‘off states are large and that the mean rate of the source is m. The
effective bandwiakh of such a source is

413
and this expassion provides a bound on the effective bandwidth (1) of any source wah peak rate h and mean
rate m. For k e d h the function (2) is increasing and concave in m, while for fixed m it is increasing and convex
in h. As s -+ 0 (corresponding to a very large capacxty C in relation to the peak h), the effective bancivvldth
approaches m,the mean rate of the source. However as s increases (corresponding to a larger peak h in relation
to the capacrty 0 the effective bandwidth increases to the peak rate h of the source.
More generally, suppose that source j is policed by parameters (rfi). bfi), k=I.2,...K), so that Xj(0, cannot
exceed r(k)T+b(k) for any T and for k=I.2,...,K. Then an upper bound on the effective bandwidth (1) is
provided by expression (2) with h replaced by the minimum over k of r(k) +b(k)/?.

3.2 Charging and connection acceptance control

Next we consider how an effective bandwidth might be used for charging and connection acceptance control.
One possible charging mechanism might assess the effective bandwidth of a connection, using an empirical
average to replace the expectation (I), and then charge according to the assessment. Apart from the d~fficukyof
interpreting this tariff to users, there is a conceptual flaw, which can be illustrated as follows. Suppose a user
requests a connection policed by a high peak rate, but then happens to transmit very little traffic over the
connection.Then an a posteriori estimate of quantrty (1) will be near zero, even though the a priori exp&cm
may be much larger, as assessed by either the user or the network. Since tariffing and connection acceptance
control are primarily cuncemed wrth expectations o f j h r e quality of service, the distinction matters. Put
another way, the guarantee of a low cell loss rate for a connection may be valuable to a user, and may be costly
for the network to provide, even though the user transmits little or no traffic over the connection.
An altemative charging mechanism might calculate the largest possible effective bandwidth subject to the
agreed policing parameters, and charge accordmgly. For example, the expression ( 2 ) might be evaluated w& h
and m set equal to a peak and sustainable cell rate respectively. This tariff is certainly easier to explain, but
severely penalizes users whose mean traffic may be unpredictable and not easily characterized by policing
parameters.
An altemative approach is to regard the effective bandwidth as a function of both static parameters (such as the
parameters of leaky bucket regulators) and dynamic parameters (such as duration and volume); to police the
static parameters and measure the dynamic parameters; to bound the effective bandwidth by a linear function of
the measured parameters, with coefficientsthat depend on the static parameters; and to use such linear functions
as the basis for slmple charging and connection acceptance mechanisms.
We illustrate the approach for the case of a source subject to a policed peak rate and a measured mean rate,
with bound (2) on the effective bandwidth. Let
afi.m) + bfi.m)M (3)
be the tangent to the function B ( h M at the point M=m, for any gven value of h, and suppose that the charge
for a connection wrth policed peak rate h is given by the expression a(h.m) T + b0z.m) V where T is the
duration of the call, V is the volume of the call (e.g. in cells), and m is freely chosen by the user in the range
[O,h] at the time of call set-up. Then Kelly [8] shows that the user minimizes his expected charge by choosing m
to be his a priori expectation of the mean rate of the cannection, and with this choice of m the expected charge
is exactly expression (2). Courcoubetis et al. [IO] extend this approach to the case where h may label a vector
of policing Parameters and m a vector of measured parameters.
Observe that the choice of m simply labels the choice of a linear function (3), and that the presentation of tadF
choices for a given peak rate h may be entirely couched in terms of pairs (u(h,m),b(h,m)),giving a charge per
unit time and a charge per unit volume respectively, with no mention of the word “mean”. It is not essentiai to
provide the user w& a continuum of tariff choices: the relevant functions may be well approximated by a small
number of tangents, especially if the capacxty C is large in relation to the peak rate h.
Although the above charging scheme does not require the use of any particular connection acceptance control
mechanism, lscussron of a simple mechanism will help illustrate some of the potential links between the

414
charging and clonnection acceptance. Suppose that a resource has accepted connections 1J7..J7 and write (oj,
bj) for the coefficients (u(h,m),b(h,m))describing a tangent (3) to the bound (2) on the efktive banhdth
function of connectionj . Suppose thatthe resource measures the arriving workload X,[t] from mection j over
a period of leq@ r, and let Mj = A'j[t]/t.Define the effective loud on the resource to be
J
c (cl.+ bpi). (4)
j=1
Then a connection acceptance control may be defined as folIows. A new request for a cuxmedion should be
accepted or rejected according as the most recently calculated ef5xt.m load lies below or above a threshold
value, with the proviso that if a request is rejected then later requests are also rejected d either a short
interval has e1,apsed or an existing umnection has terminated. For further discussion of sucb controls, see
Gibbens et al. [I 121.
Both the chargmg mechanism and the connection acceptance control described above use bounding tangents to
the effechve bandwidth function. If the same tangents are used for both purposes then the eilkave load has a
natural interpn&tion as an aggregate charge at the resource over a recent short period. But there is no necesse
for identical tangem to be used for charging and for connectim acceptance. lhus users choosing a small peak
rate might be clfked no further choice of tariff, so that for charging purposes the effective bandwidth function
is bounded by a single tangent. In umtrast, the resource might choose its tangent to the efkxive bandwidth
function at the point where the mean rate is the long-term observed average for traffic with that peak rate. Or, if
distinct ef5ctrve bandwidth functions are used for charging and connection acceptance ccmtrol, then the
resource mi& still choose its tangent according to user's declaration of expected mean rate.
Distinct eflktive bandwidth functions might be appropriate for charging and connection acceptance control,
since the two ,areas have quite different timescales and requirements for precision. Conneuion acceptance
control must use accurately calculated effective bandwidths, based on the buffer sizes, port speeds and other
features of cunrent hardware to make decisions on connections as they request connection, othenvise quality of
service guarantees on loss rates may be compromised. While charges need to be precisely defined, they
influence users' behaviour and software application design over much longer time-scales, where features of
hardware may evolve. Thus tariff design might include consideration of the possible ef%&ve bandwidth
functions appropriate to future hardware and network scale.
A fuller discuasion of connection acceptance umtrol would consider multiple constraints on effective
bandwidth, network routing, and the shadow prices associated with dfferent physical and logcal resources (cf.
[13]). The analysis and implementation of dynamic routing schemes often use Lagrange muhpliers or shadow
prices for each of the intemai resources of the network, but only certain aggregates of this network detail might
usefully influence charges to users. For example, competition between network providers mi& appear to users
as a choice benveen routes, and averaged congestion measures might motivate a predictable time-of-day element
to charges for dlelay-sensitive connections. Section 4 will consider other ways of providing dynamic feedback on
congestion, for users able to respond to such feedback.

4. Delay tolerant traffic

Section 3 concerned delay-sensitive traffic, and we have seen that charging schemes based on measurements of
time and volume can encourage the coordination of users and the network, and allow statistical sharing over
short time-scales. In this section we consider traffic less sensitive to delay, arising from applications that have
been termed elaistic (Shenker [14]). Statistical sharing is somewhat simpler for such traffic, smce flows may be
coordinated over time periods longer than round-trip delay times across the network, and smce delay itself is
available to corivey umtrol information. On the other hand the aims of pricing may be more complex, and may
include prowding feedback to users on network congestion and the revelation of user valuations. We shall
describe a s q l e scheme based on measurements of time and volume that can again achieve almost all of what
could be expected of any scheme.
 We shall describe the-scheme in terms of the available bit rate (ABR) service class of ATM standards. This
service class assumes that a connection responds to rate control messages from the network, but that a user may
choose a minimum cell rate (MCR) below which the connection will not be asked to fdl. The essence of the
scheme is that traffic up to the MCR is charged at one rate, while traffic above the MCR is charged at a lower
rate. If a resource within the network has spare capaclty beyond that required for high-prionty and MCR
traffic, then it may be shared amongst ABR connections: to remove any incentive for a connection to split into
several connections, we assume that the network’s allocation policy is to share spare capaclty amongst
connections in proportion to their MCRs [21]. Thus the choice of MCR by a user buys a share of spare
capacity, as well as providing a minimum cell rate.
More precisely, we suppose that there is a charge of a times the chosen MCR per unit time, and additionally a
charge of b per unit volume, where b may be zero. Thus traffic above the chosen MCR is charged at a lower
rate, possibly a substantially lower rate. To illustrate the properties of this scheme, consider a typical ABR
application, such as a file transfer of a given size. By choice of MCR a user can obtain an upper bound on the
time taken to transfer the file, although the user would expect a much Edster transfer if the network were lightly
loaded. Note the important feature that both the time taken to transfer the file and the total charge to transfer the
file will be larger when the network is congested, since the higher charge a applies to a larger volume of the
transfer. Users may of course complain that they are charged more for a slower service, but this is the key
characteristic of any incentive-compatible scheme designed to ease congestion. At times of congestion the user
can speed up file transfers by increasing the chosen MCR for connections: each user is able to act according to
its own trade-off between delay and cost.
Note that users and network can achieve a coordinated response to congestion without the need for the charges a
and b to depend upon the level of congestion or even upon fhctors such as the time of day. The key point is that
- fbr traffic that is not highly delay sensitive, both price and delay are available as coordination signals (cf.
Shenker et al. [15]). The scheme described here allows delay to‘carry feedback on network congestion to users;
the charges a and b tum the delay signal into a price signal Wrth many of the attractive properties of the “smart
market” of MacKie-Mason and Varian [l]. in particular, the price signal encourages the revelation of user
preferences.

.-5. A unified charging model

The essence of the pricing model developed is as follows. The cost of a connection is given by the expresskm
a@) T + b(x) Y + c(x) (7)
where T is the duration of the connection (measured in seconds, hours or months), Y is the volume of the
connection (measured in megabits or gigaceils), and x describes tariff choices allowed to the user by the network
.’
at the time of connection acceptance. The tariffchoice x includes the service class (for example variable bit rate,
rt or nrt, or available bit rate), the traffic contract parameters (such’asthe peak cell rate or minimum cell rate),
’
and further choices which nught allow a user to lower the “per unit time” rate a(x) at the cost of raising the “per
unit volume” rate b(x),as described in sestion 3. The charge per wnnection is c(x).
The above scheme thus makes a distinction between static and dynamic parameters of a connection. The static
parameters contained in x may be many and varied, but are fixed for the duratmn of the connection:the dynamic
parameters l’and V are measured in real-time. The cost of a connection to the customer are based on both static
and dynamic parameters, but in a very restricted manner: through linear functions of the dynamic pameters,
with coefficients that depend on the static parameters.
For expositional convenience we have called a*b and c charges, and expression (7)the cost, but these are really
measured in unrts of resource usage. The conversion of such units into nionetary units and a final bill is likely to
involve many other factors, for example customer discounts, marketing promotions and subscription charges,
whose &scussion seems not so inextricably l d e d wah statistical sharing. We have suggested that for deiay-
sensitive traffic the conversion might depend upon the time of day, but that for deiay-tolerant traffic statistical
sharing is possible vvltlaout a dependence upon tinie of day, much less upon dynamic estiniates of network load.

416
6. Project CAShMAN
CA$hMAN (Charging and Accounting Schemes in Multiservice ATM Networks [16]) is part of the European
ACTS progranme. It is a three year project aimed at developing, implementing, and assessing chargmg
schemes for A T M networks. It will achieve this by developing an appropriate range of pricing models,
implementing these efficiently in hardware and software, and makmg use of National Host testbed fhcilities (in
the Netherlands, Norway, and Switzerland) to trial these schemes and acquire user feedback. CA$hMAN has
participants fiom acadermc research, from communications hardware and software development companies, and
fromnetwork operators.
Within the first year of CA$hMAN the model described above in section 3 has been implemented, together with
a ‘graduated’ charging mechanism based on a more complex efktive bandwidth model. Charging units have
been developed for CA$hMAN, based on existing policing hardware, that can collect the necessary time, T,and
volume, V,measurements.
Key issues to bse explored in the CA$hMAN experiments include the following [17, 1 81:
Calculation of charges
Where should charges be calculated? In CA$hMAN the hardware measurements of time and volume are passed
to a Manager application which calculates call charges using the selected tariff parameters. Charge information
is passed back to the user in real time using a specially developed interface. For full-scale lmplementation in
commercial nelsvorks we have to consider whether the call charge calculation should be centralised in the
network, what information is transferred between networks, and whether the customer is given immediate
feedback on charge.
Customer response to tanfls
A central issue in the CA$hMAN experiments concems the customer response to the tariff structure:
How complex a tariff will the customer understand or accept?
Is the customer willing or able to make some estimate of traffic characteristics in order to choose from a
range of 0fft:red tariffs?
To what extent is the customer willing to shape his traffic and limit his quality-of-service requirement in
response to usage-sensitive tariffs?
A range of experiments has been designed to evaluate these questions. In the first year of the project these
experiments have been restricted to experimental users of the ATM pilot networks. Early results froin these
expenments show that users are strongly influenced by usage-sensitive charges, and constrain their traffic
demands appropriately.
Connection acceptance
The user’s choice of tariff should convey information on s t a t i s t i d characteristics of the traffic which can be of
assistance to connection acceptance control. The performance of such connection acceptance mechanisms
depends on the quality of this information.
Intelligence to support the user
While the initial CA$hMAN experiments are focussing on the user’s response to tariff structure, it is likely that
in firture the intelligence required to make tariff choices will be provided at least partly in software. How should
this be done, and should such intelligence be provided in the user’s application or terminal equipment, or
elsewhere in the: network?
Acknowledgements
This work was partly supported by the EU ACTS project AC-039 Project CAShMAN). This Project 1s a
collective effort, with many individuals making valuable contributions. We would like to thank all of them, and,

417
for comments influencing this paper, we are particularly grateful to Costas Courcoubetis, Emar Edvardsen,
Nicky van Foreest, Rajko Porobic, Marion Raffili, George Stamoulis and Richard Weber.
References
[l] J.K. MacKie-Mason, L. Murphy, J. Murphy, “The role of responsive pricing in the Intemet”, in Internet
Economics, L. W. McKnight and J. P. Bailey (editors), MIT Press, 1996.
[2] Intemet-Drafts, reports of work in progress by the Intemet Engineering Task Force.
[3] J. Murphy, L. Murphy, E.C. Posner, “Distributed pricing’for embedded ATM networks”, ITC14 (1994)
1053 - 1063, J. Labetoulle and J.W. Roberts (editors), Elsevier.
[4] S.H.Low, P.P. Varaiya, “A new approach to service provisioning in ATM networks”, IEEBACM
Transactions on Networks, Vol I, No 5,547 - 553, October 1993.
153 G. de Veciana, R. Baidick, “Pricing multi-service networks”, Technical Report, University of Texas at
Austin, July 1994.
[6] J.K. MacKie-Mason, H. Varian, “Pricing the Internet”, in Public Access to the Zntemet, B. Kahin and J.
Keller (ahtors), MIT Press, 269 - 3 14, 1995m
[7] H. Jiang, S. Jordan, ‘The role of price in the ~0nneCtionestablishment process”, to appear in European
Trans. Te1”munications and Related Technologies.
[SI F.P. Kelly, “Tariffs and effective bandwidths in multiservice networks”, ITCI4 (1994) 401 - 410, J.
Labetoulle and J.W. Roberts (edrtors), Elsevier.
[93 C. Courcoubetis, G. Fouskas, R. Weber, “On the performance of an effective bandwidths formula”, ITC14
(1994), 201 - 212, J. Labetoulle and J.W. Roberts (editors), Elsevier.
[lo] C. Courcoubetis, F.P. Kelly, R. Weber, “Pricing resource usage in communication networks” In
preparation.
[ l l ] F.P. Kelly, “Charging and accounting for bursty connections”. In Internet Economics, L.W. McKnight
and J.P. Bailey (editors). MIT Press, 1996.
[ 121 R. J. Gibbens, F. P. Kelly and P. B. Key, “A decision-theoretic approach to call a h s s i o n control in ATM
networks”, IEEE J. Selected Areas Communication 13, 1 101-1 1 14, 1995.
[ 131 F. P. Kelly, “Routing in circuit-switched networks: optimization, shadow prices and decentralization”.
Ahances in Applied Probability 20, 112-144, 1988.
[14] S . Shenker, “Service models and pricing policies for an integrated services Intemet”. In Public Access to
the Inremet, B. Kahin and J. Keller (editors). MIT Press. 315-337, 1995.
[15] S. Shenker, D.Clark, D.Estrin, and S. Herzog, “Pricing in computer networks: reshaping the research
agenda.”, 1996.
[ 161 CA$hMAN home page at http://www.isofk.intranet.gr/cashman/
[17] D.J. Songhurst, “Charging schemes for multiservice networks”, in Proc. 13th UK Teletrafic Svmposium,
IEE, 1996.
[18] D.J. Songhurst (editor), “Experiment design for Round l”, The CA$hMAN Consortium, ACTS project
AC-039, 1996.
[19] ITU Recommendation 1371:Traffic control and congestion control in B-ISBN. (1995)
[20] ATM Forum.ATM Forum traffic management specification version 4.0. (1996).
[21] F. P. Kelly, “Charging and rate control for elastic traffic”, European Transactions on Telecommunications,
1997.

4/8