Upload
david-ros
View
212
Download
0
Embed Size (px)
Citation preview
Computer Networks 46 (2004) 73–85
www.elsevier.com/locate/comnet
A mathematical model of the Paris Metro Pricing schemefor charging packet networks
David Ros a,*, Bruno Tuffin b
a GET/ENST Bretagne, Rue de la Chataigneraie, CS 17607, 35576 Cesson S�evign�e Cedex, Franceb IRISA/INRIA Rennes, Campus de Beaulieu, 35042 Rennes Cedex, France
Available online 9 April 2004
Abstract
Pricing has become one of the main challenges of the networking community and is receiving a great deal of interest
in the literature. In this paper, we analyze the so-called Paris Metro Pricing scheme which separates the network into
different and independent subnetworks, each behaving equivalently, except that they charge their customers at different
rates. In our model, each subnetwork is represented by a single bottleneck queue, and the ‘‘customers’’ (data packets)
choose their subnetwork taking into account not only the prices, but also the expected delay, which is supposed to have
an economic impact. We obtain some necessary and sufficient conditions for the stability of the system; we analyze the
problem of maximizing the network revenue and compare it with the case of a single network, and present a multi-
application extension of the model. Numerical results illustrating some key aspects of the system are provided
throughout the paper.
� 2004 Elsevier B.V. All rights reserved.
Keywords: Internet; Packet networks; Pricing; Queueing models
1. Introduction
Pricing in IP networks is becoming a mainconcern of Internet Service Providers (ISPs).
Classical pricing schemes, based on a subscription
fee and/or a (possibly unlimited) number of hours
of network access, depending on the country and
the type of usage, are becoming obsolete due to
three major reasons:
* Corresponding author. Tel.: +33-2-99-12-70-46; fax: +33-2-
99-12-70-30.
E-mail addresses: [email protected] (D. Ros),
[email protected] (B. Tuffin).
1389-1286/$ - see front matter � 2004 Elsevier B.V. All rights reserv
doi:10.1016/j.comnet.2004.03.019
1. Such simple schemes were an incentive to stim-
ulate utilization. Even if they probably helped
to develop the network and allowed ISPs togain market share, nowadays they may prove
sub-optimal. Indeed, demand in access net-
works may easily exceed the installed capacity,
leading to congestion.
2. Some users might wish to pay more than others
to avoid congestion and get an improved ser-
vice. Why not serve them first and (possibly) in-
crease the ISP revenue?3. Unlike the telephone network, the current
Internet––and, more so, the next-generation In-
ternet––has to deal with different kinds of appli-
cations with diverse quality of service (QoS)
ed.
74 D. Ros, B. Tuffin / Computer Networks 46 (2004) 73–85
requirements. Including these requirements in a
pricing scheme may, in some way, improve
users’ satisfaction.
A wide range of charging methods for packetnetworks has been devised and discussed in the
literature; see for instance the survey papers [1–6].
The main classes include resource reservation,
priorities between packets, auctions for bandwidth
or pricing based on transfer rates.
In this paper, we study another method, called
Paris Metro Pricing (PMP) and introduced by
Odlyzko in [7,8]. Under the PMP scheme, thenetwork is split into independent subnetworks;
more precisely, a fixed fraction of the capacity of
each link is totally allocated to each subnetwork.
The tariff will be different for each subnetwork, so
that (hopefully) congestion will be alleviated in the
most expensive ones. This method does not offer
any QoS guarantees, so that it is somehow weak
with respect to the abovementioned techniques;however, it is very attractive because of its imple-
mentation and management simplicity. In this
paper, we develop a mathematical model which
may be helpful in setting prices in a PMP-based
network. Particularly, our model will use the delay
introduced by the network as a cost.
Class-based service differentiation has already
been proposed in the literature. In [9], packets aremarked according to customer’s QoS require-
ments, but no pricing scheme is developed––the
model is only performance-based. In [10], a pricing
scheme is presented where time is divided into slots
and only a given number of packets can be served
in each slot. Users compete by choosing in which
priority class they send their packets (higher pri-
ority classes being more expensive). Note that theperformance metric in [10] is throughput, whereas
ours is delay. Besides, a continuous-time model
like the one we introduce in this paper is more
representative in many cases.
It has been argued by Gibbens et al. [11] that
network segmentation might not work under
competition (i.e., no Nash equilibrium exists);
nonetheless, their result applies to a model quitedifferent from ours: they use the traffic load as the
measure of congestion. However, delay seems to be
more relevant since it is a QoS parameter users
actually experience. Moreover, note that even if the
optimization problem in [11] looks similar to that
presented in this paper, Gibbens et al. deal with the
competition between two subsets of subnetworks.
Nevertheless, it is worth mentioning that we arrive
at the same conclusion as Gibbens et al.––i.e., theISP is better off by not splitting its network.
Note also that in [12], PMP has been adapted in
such a way that, instead of logically separating the
subnetworks, a round-robin service discipline is
used to share the bandwidth among them. This
‘‘improved’’ version of PMP, even though it de-
serves attention, is not considered here.
The layout of the paper is as follows. The PMPscheme is introduced in Section 2. In Section 3
the mathematical model is described. Necessary
and sufficient conditions for stability are then
provided, and an optimization problem (maxi-
mizing the ISP revenue) is introduced. Section 4
provides some numerical illustrations of the re-
sults of Section 3. In Section 5, we compare the
revenues of a PMP network with those of anequivalent, non-PMP network. Section 6 is de-
voted to the case of multiple applications with
different delay sensitivities. Finally, conclusions
are given in Section 7.
2. PMP: a brief description
The original PMP proposal in [8] consists of
partitioning a network into a few logically separate
networks (or classes), each having a fixed fraction
of the capacity of the entire network. Every sub-
network would route and handle packets accord-
ing to the current Internet protocols. There is no
formal guarantee of QoS, but by charging different
rates for different classes (served in the same way),it is supposed that the most expensive classes will
be less congested by way of self-regulation, hence
delivering a better QoS. The name given to this
model, Paris Metro Pricing, stems from the rules
of the Paris Metro about 20 years ago, where
trains were composed of cars of two classes,
offering exactly the same quality of seats. As
tickets prices were different, the cars for the mostexpensive class were less congested, leading to a
better perceived QoS.
D. Ros, B. Tuffin / Computer Networks 46 (2004) 73–85 75
As pointed out by Odlyzko, the advantage of
PMP is that, even if we do not have any strict QoS
guarantee using this scheme, but rather a statistical
guarantee, it would permit dispensing with com-
plex, non-scalable mechanisms (like, say, resource
reservation protocols) and keep the simpler andcheaper current model of the Internet, while
improving the QoS experience. Indeed, no signal-
ling protocols would have to be used, and the tariff
would be fixed, which is preferred by most users
[13].
From the preceding discussion, it is intuitively
clear that prices play in PMP an important role in
controlling congestion and, notably, in distribut-ing the load among subnetworks: prices are the
key to quality-of-service differentiation. The
problem of finding the ‘‘right’’ set of prices for a
given bandwidth partition may be informally sta-
ted as follows: for a PMP network to work in an
efficient manner, 1 the charge for using the most
expensive subnetwork should be high enough, so
that this subnetwork is lightly loaded, but not ‘‘toohigh’’––otherwise, the subnetwork would remain
empty, because the high quality offered would not
compensate for the (very) high monetary cost in-
curred by users.
In the next section, we present a mathematical
model of PMP that allows us to analyze its sta-
bility properties with respect to prices, as well as to
solve the revenue maximization problem.
3. Mathematical model
For the sake of clarity, we begin our study of
PMP by presenting a simplified model in which all
packets are generated by the same kind of appli-
cation and all users have the same valuation ofQoS (represented in our model by the mean packet
delay). Later, in Section 6, we will describe a more
realistic model corresponding to a multi-applica-
tion scenario where a different QoS valuation may
be associated to each application.
1 Efficiency may be defined, for instance, in terms of
optimizing the network operator’s revenue and providing an
adequate usage of each subnetwork’s resources.
3.1. Model presentation
Assume that there are I classes (i.e., subnet-
works) in a PMP network, and that the class-i per-packet price at the entrance of the network is pi(16 i6 I). We assume that there is a potential total
arrival rate ~k of packets at the network, corre-
sponding to the arrival rate when prices are set
to 0.
A utility measure is associated to each packet. It
is assumed to be a random variable U that follows
the same distribution for every packet, and that it
is independent of other packets’ utility.Also, a total cost function
pi þ cdi
is associated to a class i, where di is the mean delay
for a packet in the network and c is a constant
converting delay in money.
A packet enters (i.e., chooses) network i ifi ¼ argminjðpj þ cdjÞ and U P pi þ cdi, that is, it
chooses the less expensive subnetwork in terms of
total cost (which is a linear combination of delay
and price). If U < minj ðpj þ cdjÞ, the packet does
not enter at all, meaning that the network is too
expensive for it. Traffic elasticity is then a central
assumption of our model since a larger delay can
be accepted by a user, provided that the price isdecreased proportionally.
The actual total arrival rate is then
k ¼ ~kP ðU P minj
pj þ cdjÞ:
We denote by �F the complementary cumulative
distribution function of U , i.e., �F ðxÞ ¼ P ðU P xÞ:Finally, let ki (16 i6 I) denote the actual arrival
rate at subnetwork i, so that k ¼PI
i¼1 ki.
3.2. Stability: existence and uniqueness
Using this model, in equilibrium the distribu-
tion of packets among classes has to be stable,
meaning that the total cost pj þ cdj is the same forall classes j. Indeed, if for a given class j the valuepj þ cdj were smaller than the total cost of the
other classes, then new packets entering the net-
work would choose class j until its total cost
reaches that of other classes. This corresponds to a
76 D. Ros, B. Tuffin / Computer Networks 46 (2004) 73–85
Wardrop equilibrium [14]. Let us call ptot the valuepj þ cdj (identical for all j). As we want all sub-
networks to be used, we have the following set of
I þ 1 equations with I þ 1 unknown variables ki
(16 i6 I) and ptot:PIi¼1 ki ¼ ~kP ðU P ptotÞ
pi þ cdi ¼ ptot; for 16 i6 I ;
�ð1Þ
where di is a function of ki.
The remaining of this section is separated into
two parts. In the first part, the di represent themean waiting times (i.e., service excluded). In the
second part, results are obtained as a corollary
when the di represent the response times (i.e., ser-
vice included). What users and applications care
about is total delay, so the latter case seems more
realistic; nonetheless, we will begin by treating the
waiting time case for the sake of clarity.
In order to determine ptot and ki 8i, we have thefollowing theorem.
Theorem 1. Assume that di ¼ fiðkiÞ 2 Rþ is astrictly increasing and continuous function of thearrival rate, representing the mean waiting time of aclass-i user. Without loss of generality, we supposethat p1 > p2 > � � � > pI . Assume also that the dis-tribution of U is absolutely continuous and strictlyincreasing. Then the solution of Eq. (1) exists and isunique if and only if
p1 6 �F �1 1
~k
XIi¼1
f �1i
p1 � pic
� � !: ð2Þ
Remark 1. Note that if condition (2) is not veri-
fied, then it means that the highest price p1 is too
high to obtain an equilibrium between classes.
Consequently, this class would be ignored (i.e., the
subnetwork would remain empty) and the com-
putation would have to be carried out again withthe I � 1 remaining classes.
Thus, condition (2) has to be taken as an
assumption to get the equilibrium for I classes.
Proof of Theorem 1. Denote by �F the comple-
mentary cumulative distribution function of ran-
dom variable U . The system (1) can be re-written
as
ptot ¼ �F �1 ð1=~kÞPI
i¼1 ki
� ki ¼ f �1
i ððptot � piÞ=cÞ; for 16 i6 I :
(
In particular we can get
ptot ¼ �F �1 1
~k
XIi¼1
f �1i
ptot � pic
� � !: ð3Þ
If we are able to prove that there exists a unique
ptot satisfying this equation, then the existence and
uniqueness of the ki will be straightforward toshow.
It is important to note that, by definition,
ptot P pi 8i (i.e., ptot P p1) because di P 0.
Note that the left-hand side of (3) is a contin-
uous and strictly increasing function of ptot. On the
other hand, since f �1i is strictly increasing 8i and
�F �1 is strictly decreasing (and both are continu-
ous), then the right-hand side of (3), that is,�F �1 ð1=~kÞ
PIi¼1 f
�1i ððptot � piÞ=cÞ
� is a continuous
and strictly decreasing function of ptot. Thus the
solution of (3), if it exists, is unique.
Hence, the existence of a solution––i.e., whether
the left-hand side and the right-hand side of (3)
cross each other––depends on border values of ptot.Let us first see what happens when ptot ! þ1. Let
EðSiÞ denote the mean service time for class i. Thedelay goes to infinity when the mean arrival rateapproaches the mean service rate 1=EðSiÞ, hence:limdi!1 f �1
i ðdiÞ ¼ 1=EðSiÞ. So the left-hand side of
(3) tends to infinity and the right-hand side tends
to �F �1�ð1=~kÞ
PIi¼1 1=EðSiÞ
�when ptot tends to
infinity. Therefore, the existence of a solution de-
pends on whether, at the minimal value of ptot (i.e.,p1), the right-hand side of (3) is greater than or
equal to p1: this condition is expressed by (2). Insuch a case, Eq. (3) has a unique solution. h
Corollary 1. Assume now that di ¼ f ðRÞi ðkiÞ repre-
sents the system’s response time. Then di 2½EðSiÞ;1Þ, with EðSiÞ denoting the mean servicetime.
In this case, and without loss of generality, theclasses are supposed to be ordered by their minimumtotal cost p1 þ cEðS1ÞP p2 þ cEðS2ÞP � � � P pI þcEðSIÞ.
Then the solution of (1) exists and is unique if andonly if
2 The standard numerical and optimization packages in-
cluded in, say, Matlab or Mathematica can be used to solve this
problem.3 Qualitatively similar results (not shown for space reasons)
were obtained by taking �F uniformly distributed––see [16].
D. Ros, B. Tuffin / Computer Networks 46 (2004) 73–85 77
p1 þ cEðS1Þ
6 �F �1 1
~k
XIi¼1
ðf ðRÞi Þ�1 p1 � pi
c
� þ EðS1Þ
�!; ð4Þ
or equivalently, using the waiting times,
p1 þ cEðS1Þ
6 �F �1 1
~k
XIi¼1
f �1i
p1 � pic
� þ EðS1Þ � EðSiÞ
�!:
Proof. Since the mean waiting time cannot be
negative, we have to use the fact that ptot Pmaxi ðpi þ cEðSiÞÞ ¼ p1 þ cEðS1Þ. By replacing p1 inEq. (2) using f ðRÞ
i instead of fi, we get Eq. (4).
Besides, using the fact that f ðRÞi is given by
f ðRÞi ¼ fi þ EðSiÞ, we can easily obtain the equation
with the waiting times. h
3.3. Optimization problem
The idea, from a network provider’s perspec-
tive, is to find the pi maximizing the revenue of the
network, i.e.,
R ¼ maxpi ;8i
XIi¼1
kipi;
subject to pi P 0 8i.Based on the existence and uniqueness condi-
tion of the arrival rates and total cost (defined in
terms of pi), and using the previous notations, the
problem can be reformulated as
maxpi;8i
XIi¼1
f �1i
ptot � pic
� �pi;
subject to
ptot P pi 8i;
ptot ¼ �F �1 1
~k
XIi¼1
f �1i
ptot � pic
� � !;
pi P 0 8ior its equivalent form, if we rather talk about the
response time instead of the waiting time.
As this problem looks analytically intractable in
the general case, in the next section we are going to
present numerical results illustrating the stability
domain, as well as the impact of prices and band-
width partition on delay, total cost and revenue.
Note that, in some sense, our approach can berelated to that of Honig and Stieglitz [15] (albeit
applied to a different model), where revenue, prices
and performance were studied.
4. Numerical results
In this section we will present some numericalresults obtained with the above model. 2 More
detailed results can be found in [16]. In what fol-
lows we will use the response time of the system as
a measure of delay, that is, di ¼ f ðRÞðkiÞ. We will
also relax the restriction p1 þ cEðS1ÞP p2 þ cEðS2Þimposed in the previous section.
For the sake of simplicity, let us consider the
case in which the ISP partitions its network inI ¼ 2 subnetworks. Indeed, even if the case I > 2
is, in principle, not harder to solve numerically
than the I ¼ 2 case, we will only consider the latter
in order to being able of graphically presenting the
results. Each subnetwork i is viewed as a single
bottleneck, modeled as a M/G/1 FIFO queue with
service rate li ¼ 1=EðSiÞ, so
f ðRÞi ðkiÞ ¼
1
liþ
1þ CV2� �
qi
2li 1� qið Þ ;
where qi ¼ ki=li and CV2 is the squared coefficient
of variation of the service law.
Unless stated otherwise, the following parame-
ters will be used throughout this section:
• Total capacity of the network: c ¼ 2.
• Potential total arrival rate: ~k ¼ 3.
• Utility distribution function �F : exponential 3
with mean �U ¼ 1.
0.5 1 1.5 2 2.5 3p1
0.5
1
1.5
2
2.5
3p2
p2 = p1γ γ
−1 µµ 2
+
2.5
3p2
p2 = p1γ γ
−1 µµ 2
+
(a)
78 D. Ros, B. Tuffin / Computer Networks 46 (2004) 73–85
• Cost per unit of delay: c ¼ 1.
• Service time distribution: exponential (i.e., each
queue is a M/M/1 queue and so CV2 ¼ 1). Note
that the M/D/1 queue gives similar results.
Note that, in the case of TCP flows, bandwidth
sharing among flows has been successfully mod-
eled as a M/G/1 processor sharing (PS) queue (see
for instance [17]). However, such a queueing sys-
tem has exactly the same response time as a M/M/
1/PS queue with identical mean service time, and
the latter has in turn the same response time as a
M/M/1/FIFO queue with the same characteristics.Therefore, our analysis of a M/G/1/FIFO queue
actually encompasses the M/G/1/PS case.
We assume, without loss of generality, that the
total bandwidth c of the network is shared among
subnetworks according to:
l1 ¼ ac; l2 ¼ ð1� aÞc with 0:56 a < 1:
ð5Þ
0.5 1 1.5 2 2.5 3p1
0.5
1
1.5
2
(b)
Fig. 1. Equilibrium region, �F exponential. (a) a ¼ 0:5, (b)
a ¼ 0:7.
4.1. Equilibrium region
It is interesting to look at the shape of the
equilibrium region, defined as the set E of all pairs
ðp1; p2Þ such that the system of equations (1) has aunique solution, that is: E ¼ E0 [ E1 [ E2, with E0
defined by the set of prices such that
p1 þ c=l1 ¼ p2 þ c=l2 that also verify the stability
condition and, for i 2 f1; 2g
Ei ¼ ðp1; p2Þ: p�i
�þ c
l�i
< pi þcli6 �F �1 1
~kðf ðRÞ
�i Þ�1 pi � p�i
c
��þ 1
li
�� ;
where �i denotes the element of f1; 2g which is not
i. Ei, for i 2 f1; 2g, can be regarded as the price
area such that the charge plus the cost of service,
i.e., the total cost minus the waiting cost, is more
expensive for class i (and such that the stability
condition is verified).
Fig. 1 shows the equilibrium region for twodistinct bandwidth allocations, in the case where
the utility is exponentially distributed. The equi-
librium region is shown in gray; the straight line
pointed to by an arrow corresponds to the ‘‘fron-
tier’’ E0 where the total costs (minus the waiting
cost) are the same for both classes.Note how the equilibrium region above the
frontier gets narrower as the parameter a in-
creases. Note also that, as prices get higher and
higher, the equilibrium region degenerates into the
straight line given by the frontier condition, no
matter the value of a. Since there is no upper
bound to the utility (meaning that some users are
willing to tolerate a very high cost), and the fixedcost pi þ c=li increases for both subnetworks,
queueing delay tends to zero (together with the
arrival rate), so equilibrium can only be reached if
D. Ros, B. Tuffin / Computer Networks 46 (2004) 73–85 79
the fixed cost is about the same in each subnet-
work; otherwise, traffic would switch entirely to
the less-expensive one.
4.2. Total delay
Delay in each subnetwork is shown in Fig. 2,
for two different bandwidth allocations. Observe
that, for subnetwork i, delay is always minimal
along the frontier of the equilibrium region Ei.
The tradeoff between prices and delay is fairly
evident in these figures. When one of the prices
(say, p1) is kept fixed, delay in the corresponding
subnetwork (i.e., d1) decreases as the other sub-network’s price (p2, in this case) decreases, while
delay d2 increases. This is due to the fact that a
00.5
11.5
22.5 0
0.5
1
1.5
2
2.5
0
0.5
1
1.5
2
p2
p1
d 1
0.51
1.52
2.5 0
0.5
11.5
22.5
0
0.25
0.5
0.75
1
p2
p1
d 1
(a)
((c)
Fig. 2. Delay in each subnetwork (shown as a dark-gray surface), �F ex
region; the set of prices E0 is shown as a solid line over the equilibrium
all figures. (a) a ¼ 0:5, subnetwork 1. (b) a ¼ 0:5, subnetwork 2. (c) a
lower p2 will tend to attract more customers to
subnetwork 2, increasing queueing delay in it––
and so decreasing queueing delay in subnetwork 1.
4.3. Equilibrium cost, revenue and optimality
Fig. 3 shows that the total cost ptot is strictly
increasing with pi. This, indeed, can easily be
proved analytically.
Fig. 4 shows the revenue of the network, R, as afunction of prices, for two different bandwidth
allocations. It is interesting to remark that there
appears to be a single maximum (i.e., there is a
single pair of prices which is optimal, in the sensethat revenue is maximized). As a increases, the
position of the maximum value of R moves closer
00.5
11.5
22.5
00.5
11.5
2
0
0.5
1
1.5
2
p2
p1
d 2
0.51
1.52
2.5
00.5
1
1.5
2
2.5
0
0.5
1
1.5
2
2.5
p2
p1
d 2
(b)
d)
ponential. The light-gray shadow corresponds to the equilibrium
region. Note that the orientation of the axes is not the same in
¼ 0:7, subnetwork 1. (d) a ¼ 0:7, subnetwork 2.
0
0.5
1
1.5
2
p1
0
0.51
1.52
p2
1
2
3
4
ptot
0
0.5
1
1.5
2
p1
2
0
1
2
3
p1
0
1
2
3
p2
1
2
3
4
ptot
1
2
3
p
1
(a)
(b)
Fig. 3. Equilibrium cost, �F exponential. (a) a ¼ 0:5, (b) a ¼ 0:7.
0
1
21 0
1
2
p2
0
0.1
0.2
0.3
R
0
1
2p1
0.51
1.5
2
2.51 0
1
2
p2
0
0.1
0.2
0.3
0.4
R
0.51
1.5
2p
(a)
(b)
Fig. 4. Revenue, �F exponential. (a) a ¼ 0:5, (b) a ¼ 0:7.
80 D. Ros, B. Tuffin / Computer Networks 46 (2004) 73–85
to the rear ‘‘edge’’ of the surface, which corre-
sponds to the upper edge of the equilibrium regionE2 in Fig. 1.
4.4. Sensitivity of the optimal revenue
4.4.1. Sensitivity to the bandwidth partition
Let us now examine how the optimal revenue, as
well as the corresponding prices, delay and total
cost, depend on the bandwidth allocation param-eter a (see Fig. 5).
First, note that the optimal revenue
Ropt ¼ maxp1;p2P 0 R is highly sensitive to the value
of the bandwidth allocation parameter a. Fig. 5(a)suggests that the ISP may find interesting, from an
economic point of view, to operate at the maxi-
mum-revenue allocation aopt. However, since Ropt
(and, hence, the revenue for any non-optimal set ofprices) rapidly decreases for values of a above aopt,
a too-unequal bandwidth allocation policy may
result in lower income. On the other hand, a
‘‘safe’’ bandwidth allocation (say, a ¼ 0:5) may
result in a maximum revenue as low as � 70% ofthe highest Ropt.
Note that, interestingly enough, the optimal
revenue is maximized when the price p2 of the
lowest-capacity subnetwork drops to zero. For
higher values of a, since p2 is at its lowest-possiblevalue, the ISP cannot compensate for the poorer
performance of subnetwork 2 by lowering its price,
so more and more traffic tends to flow throughsubnetwork 1 and the price p1 must be raised in
order to have a stable network. For a < aopt, a
falling p2 tends to attract more traffic to subnet-
work 2 in spite of the degradation in delay d2 whena increases, resulting in an increasing revenue.
4.4.2. Sensitivity to the users’ valuation of delay
The c parameter, which expresses the cost perunit of delay, can be regarded as how much users
value having a good quality of service (in terms of
delay): the higher the value of c, the higher the
impact of delay on the total cost pi þ cdi, and
the lower the probability that a given packet enters
the network.
0.5 0.6 0.7 0.8 0.90
0.1
0.2
0.3
0.4
0.5
α
Ropt
0.5 0.6 0.7 0.8 0.90
1
2
3
4
5
6
7
8
9
10
α
price
p1
p2
0.5 0.6 0.7 0.8 0.92
3
4
5
6
7
8
9
10
11
α
ptot
0.5 0.6 0.7 0.8 0.90
2
4
6
8
10
12
α
delay
d1
d2
(a)
(c)
(b)
(d)
Fig. 5. Optimal revenue and prices (and corresponding total cost and delay) as a function of a, with an exponentially distributed
utility. (a) Optimal revenue Ropt; (b) optimal prices pi; (c) total cost ptot; (d) delay di.
0.5 0.6 0.7 0.8 0.90
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
α
Ropt
γ = 0.25
γ = 0.5
γ = 1
γ = 2
Fig. 6. Effect of the c parameter on the optimal revenue, with
an exponentially distributed utility.
D. Ros, B. Tuffin / Computer Networks 46 (2004) 73–85 81
Fig. 6 illustrates the sensitivity of the optimal
revenue to the value of c, for a in the ½0:5; 0:95�range. For c ¼ 0:25, Ropt is fairly stable for a
wide range of bandwidth allocations: Roptð0:5Þ=RoptðaoptÞ � 0:87, with aopt � 0:91.
On the other hand, when c ¼ 2 (meaning that
users’ valuation of delay is eight times higher than
in the previous case) we have that Roptð0:5Þ=RoptðaoptÞ � 0:66, i.e., the ‘‘loss’’ incurred by
operating the network at a safe bandwidth allo-
cation is higher than in the c ¼ 0:25 case. In other
words, the model predicts that revenues are higherand fairly insensitive to bandwidth allocation
when users are more tolerant of delay (which is an
intuitively appealing result).
82 D. Ros, B. Tuffin / Computer Networks 46 (2004) 73–85
Notice also how the optimal allocation aopt is
dependent on c, going from aopt � 0:91 to 0.67. In
practical terms, this means that an ISP would have
to operate its network on the ‘‘safe’’ side (i.e., with
a close to 0.5) if the value of c cannot be accuratelyestimated.
5. Comparison with the case of a single network
An interesting question is how well a PMP
network performs, in terms of revenue, with re-
spect to a one-tiered (I ¼ 1), non-PMP network.
We will focus on the two-tiered (I ¼ 2) case ofPMP. We will suppose that both the one-tiered
network and the two-tiered PMP network have the
same link capacity c; in the latter case, this capacity
is shared among subnetworks according to Eq. (5).
This would correspond to the scenario in which an
ISP wants to evaluate the consequences of intro-
ducing PMP in its network by deploying the
appropriate mechanisms in routers (e.g., packetclassification and scheduling mechanisms), but
without increasing the installed capacity.
We observed in our preliminary numerical
experiments 4 that the revenue generated by a
single network (I ¼ 1) is greater than the revenue
generated by two subnetworks (I ¼ 2) when the
total network capacity––i.e., the service rate––is
unchanged, regardless of the partition parametera. Let us formally prove this result for the M/M/1
queue in the following theorem.
Theorem 2. Consider a M/M/1 queue (representingthe network) with service rate l and the responsetime as the delay cost. The revenue generated by thissingle queue is greater than the revenue generated bytwo separate queues with service rates al andð1� aÞl, respectively.
Proof. Consider first the case I ¼ 1 with a price pper packet. Since the response time d ¼ 1=ðl � kÞis also given by ðptot � pÞ=c, we can write
k ¼ l � c=ðptot � pÞ, which yields the stability
equation
4 For an example, see [16].
�F �1ðptotÞ ¼1
~kl
�� cptot � p
�:
In the case I ¼ 2 with prices p1 and p2, we have
�F �1ðptotÞ
¼ 1
~kal
�� cptot � p1
þ ð1� aÞl � cptot � p2
�
¼ 1
~kl
�� cptot � p1
� cptot � p2
�:
The proof of the theorem is in two steps: we first
show that, for a given ptot, the revenue generated is
greater in the case I ¼ 1; next, we show that, for allptot obtained in the case I ¼ 2, there exists a p in
the case I ¼ 1 giving the same ptot.Hence, for a given ptot, from the above equa-
tions in ptot we have that, necessarily,
1
ptot � p¼ 1
ptot � p1þ 1
ptot � p2;
that is,
p2 ¼ ptot �ðptot � p1Þðptot � pÞ
p � p1:
As p2 P 0 and p; p1; p2 6 ptot, we necessarily havethat pP p1. By symmetry, we also get that pP p2.Since, for a given ptot, the total arrival rate is the
same in both the I ¼ 1 and the I ¼ 2 cases, we
obtain that the revenue ðk1 þ k2Þp in the case I ¼ 1
is greater than k1p1 þ k2p2 in the case I ¼ 2.
Consider now a ptot obtained in the case I ¼ 2.Then a price
p ¼ ptot �1
1=ðptot � p1Þ þ 1=ðptot � p2Þ
gives the same value of ptot in the case I ¼ 1. The
only thing that remains to be proven is that such ap is non-negative. Since 1=ðptot � p1Þ þ 1=ðptot �p2Þ > 1=ptot, this is always the case. h
Remark 2. This result can be related to the well-
known problem of ‘‘splitting’’ a server in two.
Such separation introduces a supplementary mean
waiting time as one of the two servers may by idle
while the other has some customers waiting forservice, something that does not occur in the sin-
gle-server case (see, for instance, [18]).
D. Ros, B. Tuffin / Computer Networks 46 (2004) 73–85 83
Remark 3. The same result applies if we consider
the waiting time instead of the response time.
Moreover, in the former, the revenue generated in
the case I ¼ 2 tends to the revenue in the case
I ¼ 1 if a tends to 1. The reason why this does nothappen when considering the response time is that
the first class (which gets almost all the service rate
when a gets close to 1) has to match the total cost
of the second class, which tends to infinity with its
mean service time 1=ðð1� aÞlÞ.
6. A multi-application extension with per-applica-tion delay valuations
We will now consider the (more realistic) case in
which there are K types of applications/customers,
each having a different delay valuation. Let cðkÞ
denote the delay valuation for the class-k appli-
cation. Note that we will use the superscript k for
denoting application classes, which are differentfrom the PMP classes––i.e., subnetworks––noted
by the subscript i.Instead of a total potential arrival rate ~k and a
utility random variable U , we now have a collec-
tion of such inputs ð~kðkÞ;U ðkÞÞ for 16 k6K. De-
note by kðkÞ the total arrival rate of application-kpackets, kðkÞ
i the arrival of such packets in sub-
network i and, as before, ki the total arrival rate insubnetwork i. We have ki ¼
PKk¼1 kðkÞ
i and
kðkÞ ¼PI
i¼1 kðkÞi :
In this case, the set of Wardrop equilibrium
equations (1) becomes:
kðkÞ ¼PI
i¼1 kðkÞi ¼ ~kðkÞP ðU ðkÞ P pðkÞtotÞ
for 16 k6K;pi þ cðkÞdi ¼ pðkÞtot
for 16 i6 I and 16 k6K
8>>><>>>:
ð6Þ
with di ¼ fiðkiÞ:This means that the set (6) has to be verified for
all k, but with a different ptot for each k. Note
however that the equations cannot be solved
independently 8k since they are related through
the queueing delays di (16 i6 I).
Remark 4. In this case, the system behaves like a
game where each player k (i.e., each application)
tries to stabilize its traffic, that is, it tries to
equalize pi þ cðkÞdi for all i, meaning that pðkÞtot exists.
In this sense, a solution of (6), if it exists, is a Nash
equilibrium between all the classes.
Proving the existence and uniqueness of thesolution is more complicated in this case, since we
do not have a single ptot (see [16, Section 6.2.1] for
the single-valuation, multi-application case).
Therefore, finding out conditions for existence and
uniqueness is still an open question.
Nevertheless, based on the results in previous
sections we can still find a necessary, but not suf-
ficient, condition for equilibrium.
Proposition 1. Assume that p1 P p2 P � � � P pI . Anecessary condition for solving (6) is that
XKk¼1
~kðkÞ�F ðkÞðp1ÞPXIi¼1
f �1i
p1 � picðkÞ
� �: ð7Þ
Proof. We have, 8k,
pðkÞtot ¼ ð�F ðkÞÞ�1 kðkÞ
~kðkÞ
!: ð8Þ
where �F ðkÞðxÞ ¼ P ðU ðkÞ P xÞ. If (6) is verified, thenwe also have that, 8i,
pðkÞtot ¼ pi þ cðkÞdi: ð9Þ
We will follow now an argument similar to that
we used in the proof of Theorem 1. From (8), we
can see that kðkÞ decreases when pðkÞtot increases, so
k ¼PK
k¼1 kðkÞ ¼PK
k¼1~kðkÞ�F ðkÞðpðkÞtotÞ decreases when
pðkÞtot increases. From (9), we see that ki increases
when pðkÞtot increases, so k ¼PI
i¼1 ki ¼PI
i¼1 f�1i
ðpðkÞtot � piÞ=cðkÞ�
increases as well. We want to find
a necessary condition under which there exists pðkÞtot
such thatPK
k¼1 kðkÞ ¼PI
i¼1 ki. SincePI
i¼1 f�1i
ððp � piÞ=cðkÞÞ is a strictly increasing function of p,and
PKk¼1
~kðkÞ�F ðkÞðpÞ is a strictly decreasing func-
tion of p, the inequality given in the proposition
must hold at the smallest value p ¼ p1, since
pðkÞtot P p1 8k. h
84 D. Ros, B. Tuffin / Computer Networks 46 (2004) 73–85
The reason why this proof does not provide a
sufficient condition for the existence and unique-
ness of equilibrium is that, in the proof, we look at
the existence of pðkÞtot for a given k, and not at the
existence of the pðkÞtot all together.
7. Conclusions
In this paper, we have introduced a mathe-
matical model of the Paris Metro Pricing (PMP)
scheme for charging packet networks. This pricing
method looks convenient for Internet Service
Providers since it is fairly easy to implement and
deploy in an ISP network using current, off-the-
shelf technologies. Even if PMP does not provide
strict QoS guarantees, users would likely prefer itto more complex pricing schemes since the total
charge is linear in the volume of data, hence pre-
dictable.
Our model has allowed us to find some neces-
sary and sufficient conditions on the prices re-
quired to obtain an equilibrium. An extension to
the model, taking into account the multi-applica-
tion case, is presented. Numerical illustrations areprovided for the basic model. Other possible
extensions include: time-of-day pricing and con-
sidering other QoS metrics than delay only. In [16],
we discuss in detail a model in which quality, as
perceived by a user, is a mixture of delay and loss
probability.
The model has pointed out a possible drawback
of the PMP scheme, namely, that for a given net-work capacity revenues may be lower in a network
implementing PMP. This limitation, coming from
the adopted bandwidth-sharing policy among
subnetworks, could be alleviated by means of
more efficient scheduling mechanisms.
As directions for future research, we wish to
investigate more deeply the multi-application case
with a per-application delay valuation. Morespecifically, whereas in the case of a single delay
valuation we have obtained a sufficient condition
(Eq. (2)), we only have a necessary condition for
the multiple-valuation case. Besides, we have not
considered yet the case where condition (2) is not
satisfied for some applications, so that, for each
k, the expressions pi þ cðkÞdi ¼ pðkÞtot are only taken
into account for a subset of subnetworks i sincethe other subnetworks are considered too
expensive. Indeed, we have forced in (6) that
each application type send packets to all sub-
networks. All the possible combinations of those
cases have to be addressed. We are also inter-ested in trying to carry out the same kind of
analysis for the round-robin scheme of [12]. Fi-
nally, the case of strict QoS requirements could
be investigated.
Acknowledgements
The authors would like to thank Peter Reichl
and the anonymous reviewers for their construc-
tive remarks.
References
[1] L. DaSilva, Pricing of QoS-enabled networks: A survey,
IEEE Communications Surveys and Tutorials 3 (2) (2000).
[2] P. Dolan, Internet pricing. Is the end of the world wide
wait in view? Communications and Strategies 37 (2000) 15–
46.
[3] M. Falkner, M. Devetsikiotis, I. Lambadaris, An overview
of pricing concepts for broadband IP networks, IEEE
Communications Surveys and Tutorials 3 (2) (2000).
[4] T. Henderson, J. Crowcroft, S. Bhatti, Congestion pricing.
Paying your way in communication networks, IEEE
Internet Computing 5 (5) (2001) 85–89.
[5] B. Stiller, P. Reichl, S. Leinen, Pricing and cost recovery
for Internet services: Practical review, classification, and
application of relevant models, Netnomics 3 (2) (2001)
149–171.
[6] B. Tuffin, Charging the Internet without bandwidth
reservation: An overview and bibliography of mathemat-
ical approaches, Journal of Information Science and
Engineering 19 (5) (2003) 765–786.
[7] P. Fishburn, A. Odlyzko, Dynamic behavior of differential
pricing and quality of service options for the Internet, in:
Proceedings of ICE98, ACM, 1998, pp. 128–139.
[8] A. Odlyzko, Paris metro pricing for the Internet, in: ACM
Conference on Electronic Commerce (EC’99), 1999, pp.
140–147.
[9] K. Park, M. Sitharam, S. Chen, Quality of service
provision in noncooperative networks: Heterogeneous
preferences, multi-dimensional QoS vectors, and bursti-
ness, in: Proceedings of ICE98, ACM, 1998, pp. 111–
127.
[10] P. Marbach, Pricing differentiated services networks: Bur-
sty traffic, in: Proceedings of IEEE INFOCOM 2001, 2001.
D. Ros, B. Tuffin / Computer Networks 46 (2004) 73–85 85
[11] R. Gibbens, R. Mason, R. Steinberg, Internet service
classes under competition, IEEE Journal on Selected Areas
in Communications 18 (12) (2000) 2490–2498.
[12] P. Dube, V. Borkar, D. Manjunath, Differential join prices
for parallel queues: Social optimality, dynamic pricing
algorithms and application to Internet pricing, in: Pro-
ceedings of IEEE INFOCOM 02, 2002.
[13] A. Odlyzko, Internet pricing and the history of communi-
cations, Computer Networks 36 (5–6) (2001) 493–517.
[14] E. Altman, L. Wynter, Equilibrium, games, and pricing in
transportation and telecommunication networks, Tech.
Rep. 4632, INRIA, 2002.
[15] M. Honig, K. Steiglitz, Usage-based pricing of packet data
generated by a heterogeneous user population, in: Pro-
ceedings of IEEE INFOCOM 95, 1995, pp. 867–874.
[16] D. Ros, B. Tuffin, A mathematical model of the Paris
metro pricing scheme for charging packet networks,
Research report PI-1512, IRISA (March 2003). URL
http://www.irisa.fr/bibli/publi/pi/2003/1512/1512.html.
[17] S. Ben Fredj, T. Bonald, A. Prouti�ere, G. R�egni�e, J.
Roberts, Statistical bandwidth sharing: A study of conges-
tion at flow level, in: Proceedings of ACM SIGCOMM’01,
2001, pp. 111–122.
[18] K. Trivedi, Probability and Statistics with Reliability,
Queuing, and Computer Science Applications, second ed.,
Wiley, New York, 2002.
David Ros received his B.Sc. (withhonors) and M.Sc. degrees, both inElectronics Engineering, from theSim�on Bol�ıvar University, Caracas,Venezuela in 1987 and 1994, respec-tively, and his Ph.D. in ComputerScience from the Institut National deSciences Appliqu�ees (INSA), Rennes,France in 2000. He is currently work-ing as a Research Engineer at GET/ENST Bretagne, Rennes, in the Net-working and Multimedia Department.His active research interests includenetwork modelling and simulation, as
well as pricing and quality of service issues in IP networks.
Bruno Tuffin received his Ph.D. degreein applied mathematics from Rennes 1University, France in 1997. Since, hehas been with INRIA-Rennes, France.His research interests include develop-ing Monte Carlo and quasi-MonteCarlo simulation techniques for theperformance evaluation of computerand telecommunication systems, andmore recently developing Internet ac-tive measurement techniques and newpricing schemes. On this last topic, he isthe coordinator of the INRIA’s coop-erative research action PRIXNET (see
http://www.irisa.fr/armor/Armor-Ext/RA/prixnet/ARC.htm).