SpeedGate: A Smart Data Pricing Testbed Based on

Speed Tiers
Yih-Farn Chen , Rittwik Jana
AT&T Labs-Research, 180, Park Avenue, Florham Park, NJ, 07932, USA

Abstract The explosive growth of cellular traffic and its Unfortunately, it is difficult for a service provider to pre-
highly dynamic nature often make it increasingly expensive dict the impact of pricing on the network congestion level,
or even infeasible for a cellular service provider to provision revenues, profits, and/or social welfare without conducting a
enough cellular resources to support the peak traffic demands. large-scale trial, which may require significant engineering
Some service providers have started exploring various eco- efforts in the wireless packet core, especially in the pol-
nomic incentives, including smart data pricing, to manage net- icy/QoS components. Figure 1 shows a simplified view of
work congestion. We present SpeedGate, a smart mobile data the logical policy charging and enforcement architecture of
pricing testbed that allows a service provider to experiment a cellular system.1 The 3GPP PCRF (Policy Charging and
with different dynamic pricing strategies. SpeedGate maintains Rules Function)[12] is a new logical element in the 3GPP
persistent VPN connections to smartphones as users roam packet core that makes policy/QoS decisions based on various
between different wireless networks (3G, 4G/LTE, WiFi). inputs and pushes the policies that are to be enforced for a
The maximum available bandwidth per user session can be subscribers data session down to the PCEF (Policy and Charg-
adjusted according to various data pricing strategies. We report ing Enforcement Function) component. PCRF also serves as
preliminary results on two trials with a total of 29 users for the interface with external systems for any required subscriber
assessing their willingness to pay (WTP) for various speed information to allow for the necessary policy decisions to
tiers. Preliminary observations suggest the challenges of QoS be made. The GGSN provides external connectivity between
guarantees through speed tiers in the field, the limited dynamic user equipment (UE) and external packet data networks, while
range of WTP values from individual users for different speed PCEF sits on the data path to provide gating control, i.e., the
tiers, and potential opportunities for auction-based dynamic blocking or allowing of packets based on the QoS policies.
pricing. While the PCRF/PCEF architecture provides some of the key
building blocks for smart data pricing and allows volume-
I. I NTRODUCTION based charging, time-based charging, event-based charging,
etc, it requires significant engineering efforts to modify the
As mobile data traffic has been growing explosively over the wireless core to support various dynamic pricing trials. Our
past few years and is expected to grow 18-fold between 2011 testbed aims to provide a capability of running pricing trials
and 2016 [3], how to provide high-quality mobile Internet ser- with a large number of participants without accessing or
vices to satisfy the ever-increasing traffic demands is becoming modifying the PCRF/PCEF components in the core network.
an urgent issue faced by todays wireless service providers. There has been a lot of research in pricing models for
The nature of this problem is different from traditional re- the Internet. They can be grouped in mainly two categories,
source allocation problems in which resources are under the namely static and dynamic. Static plans charge users according
control of a single service provider. The smartphone users to predetermined rates. They are easier to understand for the
have the choice to access a large array of third-party services end user; however, these plans do not reflect users willingness
with varying bandwidth demands at any time and almost from to pay (WTP) or network conditions in any way [5], [6], [10].
anywhere. How to incentivize data consumers to adjust their Dynamic pricing, on the other hand, tries to change the user
usage behaviors to help manage the network congestion is a prices according to the network conditions [9], [14], [8], [15].
significant challenge that service providers need to address. We recognize the importance of such dynamic pricing plans
Congestion caused by the onslaught of popular, data-intensive in the context of mobility. There is little evidence of a pricing
smartphone apps can lead to a poor user experience, including testbed that has been developed to try out different pricing
slow access to content and dropped call. One answer has strategies. Sen et al. [8] recently proposed such an architecture
been to add more cellular resources, at considerable cost, (TUBE), which is mainly designed for time-delayed pricing
and to complement cellular capacity with other technologies, (TDP).
such as WiFi and Femtocells. While these solutions have In this paper, we describe an alternative smart data pricing
successfully reduced network congestion and offloaded traffic (SDP) testbed called SpeedGate that has several unique fea-
from the cellular network, there is still work to be done. As one tures: (a) carrier-independence: It allows mobile users from
step toward keeping up with the exponential data growth and
1 UTRAN stands for UMTS Terrestrial Radio Access Network, SGSN
solving the congestion problem, both researchers and service
stands for Serving GPRS Support Node, GGSN stands for Gateway GPRS
providers are exploring economic incentives such as smart data Support Node, and HLR/HSS stands for Home Location Register/Home
pricing. Subscriber Server.
 2

Fig. 1. A Simplified Architecture Overview of the 3GPP PCRF and PCEF Fig. 2. SpeedGate: A Smart Data Pricing Testbed with Speed Tier Proxies
functions

conducting a trial based on Generalized Second-Price Auction

any carrier and anywhere to bring their own smartphones or
(GSP)[2]. We highlight the salient features that are necessary
tablets to participate in the pricing trials. (b) persistent con-
to conduct such a trial.
nection: The testbed maintains a persistent VPN connection
Figure 2 shows the architecture of the SpeedGate testbed,
even when the subject hops among different wireless networks
which can be configured to simulate multiple speed tiers
(3G, 4G/LTE, WiFi). (c) dynamic speed tier assignments:
(we have used speeds ranging from 128kbps to 32Mbps
The testbed can redirect smartphone users to different ports
during our various trials). It allows us to dynamically switch
on proxy servers in real time and adjust their maximum
any smartphone user to a different speed tier based on the
QoS level (allowed bit-rates) dynamically based on the smart
dynamic pricing algorithms. A smartphone owner needs to
data pricing algorithms. We have used SpeedGate to quantify
first download the Cisco AnyConnect VPN client[1], obtain
users utilities, what really affects users decisions and how
credentials from the trial operators, and then establish the
malleable users are towards different pricing strategies, all
VPN connection to the testbed. The Cisco VPN appliance[4]
without significant engineering efforts in the carriers wireless
can maintain a VPN connection persistently even when a
packet core. The rest of the paper is organized as follows:
smartphone switches among LTE, 4G, 3G, or various WiFi
Section II describes the hardware and software components
networks. The Microsoft Active Directory/Certificate Author-
of our SDP testbed SpeedGate. Section III describes two
ity server issues a certificate for each smartphone during the
trials that we ran on our SDP testbed to measure users
initial connection. The certificate instructs the VPN appliance
willingness to pay (WTP) for a given quality of service. The
which proxy server and port number to redirect the traffic of all
first trial uses the testbed to provide different speed tiers to
future VPN connections from this smartphone to. It also allows
cellular data users. The second trial uses the same testbed
each smartphone to authenticate itself to the VPN appliance as
(specifically bandwidth proxies) in a WiFi environment. The
it reestablishes the VPN connection. Dummynet traffic shaper
second trial also provided a survey-based economic technique
(through ipfirewall)[13] is installed on each proxy server to
for estimating the valuation of user experience. This type of
throttle the TCP connection at different speeds for different
technique has been used in experimental economics, known as
ports. Currently, each proxy server can support up to 64
contingent valuation [7]. Section IV describes how SpeedGate
smartphone VPN connections and therefore, the testbed can
can be extended with an auction engine to support dynamic
only support up to 256 concurrent users. We have plans to
pricing based on Generalized Second Price (GSP) Auction.
move the whole infrastructure to the cloud to scale the testbed
Section V concludes with a summary and future work.
to handle a larger trial. The syslog server is set up to collect
logs on the connection times (but not the details of the traffic)
II. T ESTBED
of each VPN user.
Our smart data pricing (SDP) testbed is designed to satisfy Note that our approach is different from Paris Metro
one key requirement in conducting dynamic pricing trials: Pricing[11], where the network is partitioned into several
the testbed must be able to dynamically adjust the QoS logically separate channels with the only difference being the
level (expressed as the maximum bit rate) of each individual prices paid for using them. Channels with higher prices would
trial user. This section describes the hardware and software attract less traffic and thereby provide better service. Paris
modules that have been developed to meet this requirement Metro Pricing uses price as the only primary mechanism to
and to facilitate running trials using various smart data pricing manage traffic and there is no QoS guarantee. In addition,
algorithms. To date we have used this infrastructure to estimate the scheme may sacrifice some of the utilization efficiency
customers willingness to pay (WTP) for given speed tiers. of the network. Our platform is intended for dynamic pricing
In the future, we will leverage the same infrastructure for schemes such as GSP[2]. In the GSP scheme, a user is allowed
 3

TABLE I
to bid at a price of his or her choice and then assigned a speed WTP PARAMETERS AND INITIAL TRIAL RESULTS
tier dynamically based on the current resource constraints and User Treatment A WTP A$ Treatment B WTP B $
bids from other users. The throughput is close to the speed stu10 p1, 3010, tier3 0.00 p1, 3002, tier1 0.00
tier as long as the underlying wireless speed is above the stu11 p1, 3011, tier3 20.00 p1, 3003, tier1 20.00
speed tier. The testbed can dynamically adjust the number stu12 p1, 3012, tier3 p1, 3004, tier1
stu13 p1, 3013, tier4 20.00 p1, 3005, tier2 25.00
of ports available for different speed tiers to either optimize stu14 p1, 3014, tier4 p1, 3006, tier2
the network utilization, the revenues, the total social welfare, stu15 p1, 3015, tier4 40.00 p1, 3007, tier2 70.00
or to follow the PFauc (Proportional Fairness with Auction) stu16 p1, 3016, tier4 p1, 3008, tier2
stu17 p1, 3005, tier2 10.00 p1, 3001, tier1
algorithm proposed in the GSP paper[2]. stu18 p1, 3006, tier2 10.00 p1, 3002, tier1 10.00
The SDP testbed can be used for various dynamic pricing stu19 p1, 3007, tier2 25.00 p1, 3003, tier1 25.00
trials without altering the core network of a carriers wireless stu20 p1, 3008, tier2 10.00 p1, 3004, tier1 10.00
stu21 p1, 3001, tier1 5.00 p1, 3005, tier2 3.00
infrastructure. In addition, trial participants can use any carrier stu26 p1, 3002, tier1 0.00 p1, 3010, tier3
(and any of a large number of smartphones supported by the stu27 p1, 3003, tier1 15.00 p1, 3011, tier3 15.00
VPN appliance). In the next section, we first describe two stu28 p1, 3004, tier1 15.00 p1, 3012, tier3 5.00
stu29 p1, 3005, tier2 5.00 p1, 3013, tier4 15.00
specific trials we conducted on the platform to measure the stu30 p1, 3006, tier2 0.00 p1, 3014, tier4
willingness to pay (WTP) of users for various speed tiers,
followed by a discussion on how the testbed may be used
to conduct a trial on GSP for Congestion Pricing[2] with the
addition of an auction manager. in time after 24 hours and were therefore not subject to the
subsequent treatment. This explains why some users WTP
III. A SSESSING A USER S W ILLINGNESS T O PAY (WTP): values are missing. Notions like p1, 3010, tier3 in the second
T RIALS AND E VALUATION R ESULTS and fourth columns state that the subjects VPN connection
We now describe two trials ran on our SDP testbed for was re-directed to proxy server 1 on the SDP testbed, port
assessing a users willingness to pay for a given quality of 3010, which uses speed tier 3. There are a couple of interesting
service. The quality of service can entail different speed tiers, observations that we make.
delay-bandwidth product, etc.
A large number of users kept their WTP amount similar
First trial: Measuring WTP through two 24-hour treat-
in both treatments: Stu11, 18, 20 and 27 did not adjust
ments in the field
their WTP for different treatments. This could have been
The trial entailed carefully recruiting students from multi-
explained by a variety of factors. We did not impose a set
disciplinary faculties. We observed the users WTP for differ-
of tasks to be completed (i.e. users were free to use their
ent allocated speed tiers. For the purpose of anonymity, we
phones as they wish). A user who typically checks emails
will not report on the particular speeds used, instead, we will
a couple of times a day will not be affected drastically by
declare the different allocated speeds as tiers 1 to 4. Higher
a speed tier change. Retrieving emails does not require a
tier is allocated a higher speed. Each user will experience two
large amount of bandwidth within strict timelines.
speed tiers in two separate treatments (higher to lower speed
On the other hand, a user who consumes a lot of video
tier, or lower to higher speed tier). The users are required to
data will have very different expectation. Interruption-
bring their own smartphones and download the Cisco VPN
free viewing is critical for such a user. Stu29 for example,
AnyConnect client first before the treatment starts.
increased his WTP significantly going from a lower speed
Using the testbed described in section II, we performed a
to a higher speed.
preliminary trial with 17 students.
stu10 switched to campus WiFi completely (as he stated
After initial screening and provisioning of users into the in his response) and opted to pay for $0 for either speed
system, users are first subjected to treatment A for 24 tier. This is probably true for stu26 as well.
hours. The 24-hour treatment period was necessary to What surprised us the most was when a subject opted to
ensure that there is sufficient time for the subject to give a lower price for a higher speed tier. This happened
exercise the applications that he/she uses on a daily basis. in the cases of stu13, stu15, stu21, and stu28. As we
They are then requested to report their WTP for the speed stated previously, while the SDP testbed strives to provide
tier they experienced in that treatment after receiving an the maximum throughput specified by the speed tier,
SMS message sent to their phones. The subjects are also due to the varying wireless coverage the subject may
told the current typical wireless pricing plan as a basis experience, there is really no guarantee that they will
for comparison. indeed receive that speed. In general, it is risky for any
Upon successful receipt of their responses, the users are wireless carrier to guarantee a particular speed for the
migrated to a new speed treatment B for an additional 24 same reason - and the user may find it annoying when
hours. he or she does not get the stated speed.
Users are once again requested to state their WTP for The most interesting observation during this preliminary
treatment B after 24 hours. trial is that most users two WTP values are pretty close
Table III shows the treatments applied to the 17 users and to each other. The average difference between the first
various responses from the users. Some users did not reply and second WTP values (among all valid replies) over
 4

Fig. 3. WTP mean and stdev of the three periods among 12 trial participants Fig. 4. Application rating on three apps: Stock, Pandora, and YouTube, based
on speed tiers 1 and 2

the corresponding higher WTP is only 21.47%, while the

speed difference ranges from 4 to 8 times between the on three apps with different levels of bandwidth require-
two tiers. ments: Stock, a simple app with simple text/graphics,
Pandora, a streaming music app, and YouTube, a video
The preliminary trial is not conclusive due to the limited
app. The users gave significantly lower ratings on
number of samples, but the last point suggests that most
YouTube (at the resolution of 360p) when using speed
subjects seem to have a limited dynamic range of valuation for
tier 1. The impact is less pronounced on Pandora, which
the wireless services regardless of the speed tiers offered - and
buffers enough of the audio stream initially to provide
this points to opportunities for auction-based pricing, where a
a smooth listening experience even at a low speed tier.
smartphone user would name a fixed price (based on their
There was virtually no difference in the rating for the
budget) and have their speed tier be ranked accordingly. We
stock app.
are currently expanding the trial to up to 1000 users in order
to get enough statistical significance and to answer various Additional Lessons: We have also run into several unex-
pricing-related questions for service providers. pected problems during the construction of the testbed and we
Second trial: Measuring WTP through application- would like to briefly summarize our experience here:
rating using three treatments in the Lab VPN Performance: It turns out that several older-
In the second setup, we invited 12 participants to come to generation smartphones could not achieve any throughput
a lab setup and conducted a 45-minute session (with short beyond 2Mbps (even with WiFi) when they use the
interviews) where the subject would go through three speed- VPN connection. This complicates our trial as we had
tier treatments (tier2, tier1, tier2), where tier2 is 16 times the to carefully map those phones only to speed tiers lower
speed of tier1. Due to the lab setup, the actual speed the user than 2Mbps.
experienced on the Wi-Fi network is close to the maximum Android Proxy-Redirection Issues: While we are able
throughput allowed by the speed tier. During each treatment, to establish VPN connections to Android smartphones,
we also asked each subject to perform several tasks using the Cisco VPN appliance could not redirect them to spe-
selected applications (Stock, Pandora, YouTube, Weather, and cific proxy servers and ports that the certificates specify.
Google Map) and rate the experience under the throttled speed. We expect this issue to be resolved by Cisco and/or
The reason for the third treatment with the same speed tier Android engineers in the future. In the mean time, we
used during the first treatment was to see if the user would are exploring a cloud-based solution to bypass this issue.
appreciate the higher speed tier more after experiencing the
lower speed tier. We have the following observations after the IV. S PEED G ATE FOR AUCTION -BASED DYNAMIC P RICING
trial: SpeedGate can be used to support auction-based dynamic
Unlike the first setup in the field, the stable speed due pricing algorithms through the addition of an auction manager.
to the lab environment and the 16 times difference in Figure 5 shows our design of the auction manager and its
speed between the two tiers has indeed created a major components. A bid for service-level adjustment is received
difference in the WTP values for the two speed tiers, as through the bid interface manager and sent to the auction
shown in Figure 3. The mean for WTP in period 1 (tier engine, which ranks each of the existing users connected to
2) was $21.08, while the mean for WTP for period 2 (tier the testbed. User bids provide an indication of value each user
1) was only $6.48. There was a slight increase of WTP places on receiving data services facilitated by the testbed. For
from period 1 to period 3 (same speed tier), which has a example, users may spend more money in return for a tier level
WTP mean of $21.88. of service that provides higher speed tier, thereby resulting
Figure 4 shows the average rating of the application ex- in faster data service performance. The auction engine works
perience by the 12 participants under the two speed tiers in conjunction with the resource manager to determine the
 5

number of available slots for different speed tiers, the user

profile store to get all existing bids, and the user bandwidth
monitor to guage the users wireless channel quality, before
determining the ranking of each user and allocating the speed
tier through the bandwidth allocation engine.
As an example, the current SDP testbed has 4 proxies,
each with the capability to support up to 64 VPN connec-
tions, resulting in a total of 256 VPN connections. The total
bandwidth is constrained by the 64Mbps connection from the
testbed to the Internet backbone. The core resource manager
decides how to allocate the actual number of active VPN
connections for each speed tier. For example, the testbed may
allocate 16 16Mbps slots, 32 4Mbps slots, 80 1Mbps slots,
and 128 256kbps slots based on the bandwidth constraints,
typical resource utilization, the expected revenue, etc., among
other factors. Note that not all active VPN connections would
consume the maximum bit-rates allocated to them (and some
Fig. 6. Smartphone UI for Smart Data Pricing
may remain idle for a long time); therefore, the total bandwidth
allocation may not exceed 64Mbps most of the times. The
The bandwidth tier indicator in the status bar informs a
auction engine takes input from the bandwidth monitor and the
user of a type of bandwidth service enabled on the wireless
user bids stored in the user profile store to rank all the current
device. For example, a first tier may represent a lowest level
active user connections, and determines if the speed tier of
of priority (such as 256kbps) for data communications, while
any existing VPN connection should be adjusted (based on an
a fourth tier may represent a highest level of priority (such
auction algorithm similar to PFauc[2] in nature). If the speed
as 8Mbps). The bandwidth tier indicator may flash, change
of a connection needs to be adjusted, it sends the request to the
color and/or the wireless device may beep and/or vibrate when
bandwidth allocation engine, which sends an LDAP request to
there is an opportunity to move to a different tier plan based
the certificate server (see Section II) to modify the proxy and
on the users preference. The user may invoke a bid via a
port setting associated with the users certificate and reset the
button press. An example app bid update may correspond
VPN connection, which in turn causes the speed tier of the
to a bidding application that allows the user to enter a bid
user to be adjusted. Note that our implementation is slightly
value amount in an effort to increase a tier status, decrease
different from the actual PFauc algorithm in that we adjust
a tier status and/or otherwise manage one or more bids. In
the speed tier at a macro level, while PFauc makes low-level
the event that the entered bid amount is sufficient to raise a
decisions every 2ms to determine the number of channelization
current tier status, the example bandwidth tier indicator in the
codes to be allocated to each user.
status bar may change its appearance.
In the future, we would like to explore how the SDP testbed
can be enhanced to support other well-known pricing schemes
for side-by-side comparisons on resource utilization, social
welfare, revenues/profits, etc.:
Paris Metro Pricing[11]: The SDP testbed can simulate
several logically separate channels by throttling at the
proxy level (say at 16 Mbps) instead of at the port level.
The proxy servers that host the channels with higher
prices would attract less traffic and hence likely to provide
better QoS since fewer users would share the 16 Mbps
bandwidth.
Time-Delayed Pricing (TDP)[8]: TDP addresses the con-
gestion problem by considering when a user consumes
Fig. 5. Auction manager of SpeedGate data, in addition to how much is used. The TUBE
While many components of the testbed have been built, we system was used to conduct a trial based on TDP and
are still at the design stage for the smartphone UI for GSP. it computes TDP prices so as to balance the cost of
In the illustrated example of Figure 6, the price indicator $ congestion during peak periods with that of offering lower
on the top status bar is a monetary value to represent the prices in less congested periods. The SDP testbed can
estimated cost of joining a higher and/or lower tier of service. simulate TDP by removing all throttling through proxies,
The core resource manager is tasked to measure the activity while adjusting prices charged to users by monitoring
levels. Depending on the degree of congestion and the current the VPN connection times through logs captured on
bids from other users, a corresponding price indicator may be the syslog server; however, many additional components
presented in the example status bar. from TUBE, such as their feedback-control mechanism
 6

and their smartphone GUI must be added as well to

facilitate a full-scale simulation.
V. C ONCLUSION
In this paper, we presented SpeedGate, a smart data pricing
testbed that allows a service provider to experiment with
different dynamic pricing strategies on mobile data. SpeedGate
maintains VPN connections to smartphones that can be ad-
justed to different speed tiers based on data pricing strategies.
The testbed can maintain persistent VPN connections as smart-
phone users switch among different wireless networks (3G,
4G/LTE, and WiFi networks). We report preliminary results on
two trials with a total of 29 users for assessing their willingness
to pay (WTP) for various speed tiers. Preliminary observations
point to the challenges of QoS guarantees through speed tiers,
the limited dynamic range of WTP values from each user for
different speed tiers, and potential opportunities for auction-
based dynamic pricing. We also describe how the testbed can
be extended with an auction engine to handle Generalized-
Second-Price (GSP) auctions for dynamic pricing.
VI. ACKNOWLEDGMENTS
The authors gratefully acknowledge the contributions of
Karthik Kannan and Harsha Raman in in managing the first
trial, and the administrative support from Steven Gao, Gang
Yao, Robert Hooper, and Fu Che during various stages of the
trials, and the support and guidance from Dana Tardelli on
this project.
R EFERENCES
[1] Cisco anyconnect secure mobility client for mobile platforms data sheet.
https://bitly.com/shorten/.
[2] Y. Chen, R. Jana, and K. Kannan. Using generalized second price auc-
tion for congestion pricing. In Global Telecommunications Conference
(GLOBECOM 2011), 2011 IEEE, pages 16. IEEE, 2011.
[3] Cisco visual networking index: Global mobile data traffic forecast
update, 2010-2015. Cisco White Paper, Feb. 2011. http://bit.ly/qjQmn.
[4] Cisco secure remote access cisco asa 5550 series ssl/ip sec vpn edition.
http://bit.ly/6emG5I.
[5] D. D. Clark. Internet cost allocation and pricing. Internet Economics,
1997.
[6] R. Cocchi, S. Shenker, D. Estrin, and L. Zhang. Pricing in computer
networks: Motivation, formulation and example. In IEEE/ACM Trans-
actions on Networking, 1993.
[7] Contingent valauation. http://en.wikipedia.org/wiki/Contingent
valuation.
[8] S. Ha, S. Sen, C. Joe-Wong, Y. Im, and M. Chiang. Tube: time-
dependent pricing for mobile data. In Proceedings of the ACM SIG-
COMM 2012 conference on Applications, technologies, architectures,
and protocols for computer communication, pages 247258. ACM, 2012.
[9] J. MacKie-Mason and H. Varian. Pricing the internet. In In Public
Access to the Internet - Prentice Hall, 1995.
[10] A. Odlyzko. Paris metro pricing for the internet. In Proc. of 1st ACM
Electronic Commerce, 1999.
[11] A. Odlyzko et al. Paris metro pricing for the internet. In Electronic
Commerce: Proceedings of the 1 st ACM conference on Electronic
commerce, volume 3, pages 140147, 1999.
[12] 3gpp ts 23.203 - policy and charging control architecture. http://www.
3gpp.org/ftp/Specs/html-info/23203.htm.
[13] L. Rizzo. Dummynet: a simple approach to the evaluation of net-
work protocols. ACM SIGCOMM Computer Communication Review,
27(1):3141, 1997.
[14] N. Semret, R. R.-F. Liao, A. T. Campbell, and A. Lazar. Pricing,
provisioning and peering: Dynamic markets for differentiated internet
services and implications for network interconnections. In IEEE Journal
on Selected Areas in Communications, 2000.
[15] S. Sen, C. Joe-Wong, S. Ha, and M. Chiang. Pricing data: A look at past
proposals, current plans and future trends. In arxiv.org/abs/1201.4197,
2012.