This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE/ACM TRANSACTIONS ON NETWORKING 1

A Pricing-Aware Resource Scheduling Framework

for LTE Networks
You-Chiun Wang, Senior Member, IEEE, and Tzung-Yu Tsai

Abstract— Long term evolution (LTE) is a standard widely On the other hand, the pricing policy plays an important
used in cellular networks today. Both resource scheduling and role in cellular networks, as it significantly affects operator
pricing are two critical issues. However, existing studies address revenue. Most operators classify users into different levels
them separately, making the goals of improving system perfor-
mance and increasing operator revenue conflicting. This paper based on the pricing categories. Higher-level users are charged
proposes a pricing-aware resource scheduling (PARS) framework with a higher rate but can enjoy more resource. According
to conquer this conflict. It classifies users into three levels and to [5], pricing policies are categorized into static and dynamic.
has scheduling and pricing modules, which are installed in a base In a static pricing policy, users pay a fixed rate no matter how
station and the core network of LTE, respectively. The scheduling their traffic loads increase. A dynamic pricing policy charges
module uses three-layer schedulers to assign resource to a flow by
considering its packet delay, traffic amount, channel condition, more money when the user’s load exceeds a threshold. In this
and user level. The pricing module uses price elasticity of demand way, it can help increase the operator’s revenue.
in economics to adaptively adjust the amount of money charged Unfortunately, the goals of improving system performance
to users. Through experiments by LTE-Sim, we show that PARS and increasing operator revenue may conflict, especially when
achieves a good balance between performance and revenue, and network resource is insufficient. Let us consider an exam-
provides quality of service for the flows with strict delay concerns.
ple with two-level users. Many high-level users encounter
Index Terms— Cellular network, long term evolution (LTE), bad channel quality, while most low-level users have good
pricing, quality of service (QoS), resource scheduling. channel quality. To improve performance, one would give
more PRBs to low-level users, thereby diminishing revenue.
I. I NTRODUCTION On the contrary, if we give more PRBs to high-level users,
the revenue increases but the performance degrades. However,
L ONG term evolution (LTE) has now been operated in
many countries to provide 4G service. Comparing with
past systems, LTE exploits some efficient techniques, including
the investigation of resource scheduling and pricing in LTE is
independent. Consequently, it motivates us to integrate LTE
orthogonal frequency division multiple access (OFDMA) [1], resource scheduling with a pricing policy, so as to achieve a
carrier aggregation [2], and heterogenous cells [3], to provide good balance between performance and revenue.
high-speed wireless access. Therefore, people can freely use This paper develops a pricing-aware resource schedul-
various broadband applications such as multimedia streaming ing (PARS) framework based on the above observation.
and video downloads on their mobile phones. Without loss of generality, we classify users into three levels:
In an LTE cell, the base station called eNodeB (also golden (high), silver (medium), and bronze (low). PARS con-
abbreviated to ‘eNB’) takes charge of scheduling spectral sists of both scheduling and pricing modules. The scheduling
resource to user equipments (UEs). With OFDMA, the down- module employs a three-layer scheduling strategy. It first
link resource is concretized by a 2D array of physical resource estimates the amount of resource used to support QoS for
blocks (PRBs) in time and frequency domains. Each PRB guaranteed-bit-rate (GBR) flows, then allocates PRBs to each
carries different number of data bits, depending on the channel flow, and finally checks if some PRBs can be reallocated
quality of a UE in respect of that PRB. In general, LTE to improve performance. The pricing module follows the
performance is decided by the way that the eNB allocates price elasticity of demand (PED) model [6], where a user’s
PRBs to UEs, which we call LTE resource scheduling, and demand is affected by the price. It then adaptively computes
many methods have been developed. They aim at improv- the amount of money charged to users, depending on their
ing system performance by, for example, increasing network resource consumption.
throughput, keeping fair transmission, or supporting quality of Our contributions are threefold. First, this paper indicates
service (QoS) [4]. that existing studies may face the dilemma of improving
performance or increasing revenue, as they solve the problems
Manuscript received February 4, 2016; revised August 22, 2016; of resource scheduling and pricing separately. Second, we
accepted November 14, 2016; approved by IEEE/ACM T RANSACTIONS ON propose the PARS framework to conquer the dilemma by both
N ETWORKING Editor S. Puthenpura. The work of Y.-C. Wang was supported
by the Ministry of Science and Technology, Taiwan, under Grant MOST 104- scheduling and pricing modules, which work from perspectives
2221-E-110-036-MY2 and Grant 104-2628-E-110-001-MY2. of engineering (i.e., resource allocation) and economics (i.e.,
The authors are with the Department of Computer Science and pricing with PED), respectively. Third, each module will refer
Engineering, National Sun Yat-sen University, Kaohsiung 80424, Taiwan
(e-mail: [email protected]; [email protected]). to the outcome of the other to make its decision, so the
Digital Object Identifier 10.1109/TNET.2016.2629501 results of scheduling and pricing in PARS will tightly couple.
1063-6692 © 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE/ACM TRANSACTIONS ON NETWORKING

TABLE I Offline charging provides statistics for event- and session-

S UMMARY OF C OMMON A CRONYMS based charging. Online charging helps P-GW terminate a
user’s service when certain conditions are met (e.g., when
the amount of traffic exceeds the limitation). Based on this
structure, PARS’s pricing module can be installed in PCRF.

B. Downlink Communication
LTE divides resource into non-overlapping PRBs, each
with 0.5ms duration and 180kHz bandwidth. PRBs are
non-sharable,1 so a PRB cannot be given to multiple UEs.
The eNB is responsible for allocating PRBs, in which PARS’s
scheduling module is installed. The minimum period to allo-
cate PRBs is called a transmission time interval (TTI = 1ms).
When the bandwidth of downlink channel is 1.4, 3, 5, 10,
15, or 20 MHz, the eNB can provide 6, 15, 25, 50, 75, or
100 PRBs in a TTI.
Through a modulation and coding scheme (MCS), each PRB
carries different number of bits. In general, a more complex
MCS allows the PRB to carry more data, but it requires
the UE to have better channel condition. To help the eNB
select the proper MCS, each UE has to report the channel
quality indicator (CQI), which reveals its channel quality in
every TTI. A UE can have multiple flows, where each flow
has a queue at the eNB to be its packet container. Packets are
stamped with arrival time once they are generated, and the
eNB sends a queue’s packets in a FIFO manner. The head-of-
line (HOL) packet delay of a flow is defined by the elapsing
time of the first packet in the queue after its arrival.
LTE uses QoS class identifier (QCI) to depict the QoS
Fig. 1. LTE structure, where we omit some components in the core network.
demand of a flow, which includes packet delay budget and
Extensive simulation results exhibit that PARS can increase the packet loss rate. The packet delay budget is the maxi-
operator’s revenue, improve spectral efficiency, support QoS mum tolerant time that each packet can be delayed between
for GBR flows, and ensure non-GBR transmissions. P-GW and its UE. When the delay of a packet exceeds the
This paper is organized as follows. Section II introduces budget, the packet is invalid. The packet loss rate limits the
LTE while Section III surveys related work. We propose the maximum probability that a packet is not received by its UE
PARS framework in Section IV and give some analyses in (e.g. due to interference or expiration). LTE categorizes flows
Section V. Section VI evaluates performance and Section VII into GBR and non-GBR ones. GBR flows mainly support
concludes the paper. We then summarize acronyms in Table I. real-time applications with strict delay constraints, such as
VoIP, live-streaming video, and on-line games. Non-GBR
flows are often used for other service with loose deadlines
II. LTE OVERVIEW (e.g., TCP-based service). Thus, GBR flows usually have
A. Network Structure smaller QCI values and packet delay budgets than non-GBR
flows.
LTE network consists of multiple cells and a core network,
as Fig. 1 shows. UEs are served by the eNB in a cell. The core III. R ELATED W ORK
network deals with the management job, and has three main
A. LTE Resource Scheduling
components: 1) Mobility management entity (MME) processes
signaling between each UE and the core network. 2) Serving LTE standards leave the resource scheduling problem
gateway (S-GW) routes data packets and acts as the mobility to implementers, so various solutions are developed.
anchor when a UE moves among cells. 3) Packet data network Capozzi et al. [4] survey some popular solutions below:
gateway (P-GW) connects to the external network. It also Max-CQI uses a greedy principle to allocate each PRB to the
performs policy enforcement and supports user charging. UE with the maximum channel rate ri . Proportional fair (PF)
Charging control is done by the cooperation of policy and considers the average channel rate riavg to support fairness,
charging rules function (PCRF) and P-GW [7]. PCRF is the and it iteratively picks the UE with the largest ri /riavg value
decision center to manage each flow in P-GW, and checks if to receive resource. Modified largest weighted delay first
the flow’s behavior follows its subscription profile. PCRF has 1 It occurs when the network uses SISO (single-input single-output) or
an application function to provide dynamic charging and QoS SU-MIMO (single-user multiple-input and multiple-output) for
data to check flows. LTE supports offline and online charging. communication.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

WANG AND TSAI: PRICING-AWARE RESOURCE SCHEDULING FRAMEWORK FOR LTE NETWORKS 3

(M-LWDF) adds a weight wi and HOL packet delay di to the 2G pricing becomes inapplicable [16]. Hence, various pricing
PF solution to reduce delay. Exponential proportional
 fair policies are developed in response to the technical change [17]:
introduces a term exp[(wi di − davg )/(1 + davg )] to the 1) Fixed price charging sets a constant rental fee for users,
PF solution, where davg is the average packet latency. Both so the operator need not record bandwidth consumption.
LOG-RULE and EXP-RULE refer to the spectral efficiency However, the operator cannot increase its revenue when net-
ψi of a UE, and select a flow that has the maximum value of work traffic grows, and some users may overuse the network.
(ψi log X) and (ψi exp Y ) to get each PRB, where X and Y 2) Metered charging asks users to pay for network connection
are terms defined in LOG-RULE and EXP-RULE, respectively. on a monthly basis and charge them for metered usage of
Some work adopts a multi-layer strategy to allocate PRBs. the service. However, the usage is measured by time, so it is
Luo et al. [8] develop a cross-layer framework to support video unfair for the users who leave sessions open without sending
delivery. They refer to the delay requirement, signal distortion, packets. 3) Packet charging computes the expense charged
and past rate of each video flow to decide its PRB allocation to a user based on the number of packets sent in a session.
and coding scheme. In [9], a two-layer scheduler is designed It provides accurate pricing but relies on a packet counting
to support multimedia service. One layer computes the amount method, which complicates the billing system. 4) Expected
of data that each flow has to send in a TTI to meet its delay capacity charging lets users pay for different amount of money
demand by the discrete-time linear control theory. Then, the by their expected bandwidth usage, so the price to each user is
other layer gives PRBs to each flow by using PF for fairness predictable. However, the operator has to continually monitor
concern. The work of [10] proposes a double-layer scheme the actual bandwidth spent by each user. 5) Edge pricing aims
to schedule LTE downlink resource. The first layer translates at the case when a user stays in two cells such that his packets
the scheduling problem to a bankruptcy game and then solves are relayed by two base stations. This policy simplifies the
it by the Shapley value. Based on the result, the next layer charging mechanism by making each base station consult local
allocates PRBs according to EXP-RULE. charging information, without exchanging their billing data.
A number of approaches reduce real-time packet dropping 6) Paris-Metro charging allows users to assign a preferred
by considering their deadlines. The work of [11] applies the class with an associated cost for their different traffic (e.g.,
earliest-deadline-first method to PF, so as to support fairness business mail is viewed more important than personal mail,
while ensure that the packets whose deadlines will expire soon so a high class is given to business mail). The policy provides
can be sent first. In [12], a virtual queue is used to predict flexility, but it also adds overhead to users for traffic-class
the incoming of future packets based on the existing packets decision.
in each queue. Then, [12] discards the packets that cannot Different pricing methods for 3G and 4G networks are also
satisfy their delay demands to avoid unnecessary transmission. proposed. The flat-rate pricing method [5] works like fixed
The study of [13] divides flows into urgent and non-urgent price charging, where the fee will not change no matter how
ones, where urgent flows are given with a high priority to network traffic grows. The fixed-PRB pricing method [18]
send their packets. Non-urgent flows, including non-real-time divides users into golden, silver, and bronze levels. It charges
flows and real-time flows whose packets have not expired a user by the level li and the number ni of PRBs used:
yet, are given with the same (low) priority for transmission.
Ci = Pf (li ) × ni , (1)
Wang and Hsieh [14] use max-CQI to compute the preliminary
PRB allocation, and tax non-urgent flows with reallocatable where Pf (·) is the fixed charge by levels (in units of PRB).
PRBs. Such PRBs are given to those flows in danger of packet The network load based pricing (NLP) method [19] considers
dropping. both network load L and QCI. When L increases, users are
Few studies combine resource scheduling with other factors. charged for more money, so the operator’s revenue increases
For example, [15] proposes a scheduling method with power accordingly. Specifically, each user is charged for
saving. Each UE is assigned with a priority (Fi (rg /riavg )2 + Ci = Pv (li ) × (ê − ê−αx ) × L, (2)
Qi ) × dˆi εi if it is a GBR UE, and (Fi + Qi ) × εi otherwise,
where Fi uses the PF concept, rg is the average throughput of where ê is the Euler’s number and α restricts the QCI value x
GBR UEs, Qi is ui ’s queue status, dˆi is a delay factor, and εi to [1..9]. In Eq. (2), Pv (·) is the variable charge and depends
is a DRX (discontinuous reception) indicator for power saving. on a load threshold δ. When L ≤ δ, Pv (li ) is set to a
Then, UEs can use their priorities to compete for PRBs. constant Pc , which means that each user is charged fairly if
To the best of our knowledge, none of existing work network load is light. Otherwise, we set Pv (G) > Pv (S) >
considers integrating resource scheduling with a pricing policy Pv (B), where G, S, and B denote golden, silver, and bronze
in LTE. This motivates us to develop the PARS framework levels, respectively. The subscriber class based pricing (SCP)
with both scheduling and pricing modules, so as to balance method [20] also considers three-level users. When L ≤ δ, it
between system performance and operator revenue. charges users by Eq. (1). When L > δ, SCP charges users by
⎧
⎪
⎨(Pf (G) + Pe ) × ni if li = G,
B. Pricing in Cellular Networks Ci = (2Pf (G) + Pe ) × ni if li = S, (3)
⎪
⎩
(2Pf (G) + Pf (S) + Pe ) × ni if li = B,
Past 2G networks use circuit switching for communication,
so operators can simply charge each call by its duration. After where Pe is the extra charging computed by κ/(nA − nG ).
2.5G, the technique changes to packet switching, and thus Here, κ is a pricing constant, nA is the number of total PRBs,
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE/ACM TRANSACTIONS ON NETWORKING

TABLE II
S UMMARY OF N OTATIONS U SED IN PARS

Fig. 2. System architecture of the PARS framework.

and nG is the number of PRBs reserved for golden users.

SCP is expected to greatly increase operator revenue when
the network becomes overloaded. However, such high pricing
in Eq. (3) also degrades users’ willingness to use the service, scheduler, opportunistic resource reallocation, finally checks
thereby hurting system performance. That is why we propose whether it is possible to exchange the usage of some PRBs by
the PARS framework to integrate resource scheduling with different levels of UEs, so as to improve system performance.
pricing. PARS also adopts PED to avoid overcharging users, It also involves extra charge to some users who acquire
so as to improve performance while keeping high revenue. additional PRBs by this scheduler, and such information will
feed back to the pricing module for calculation.
IV. T HE PARS F RAMEWORK The pricing module provides variable charges for different
levels of users. The charging strategy is the core of this
This section proposes the PARS framework. We first give
module. It takes the PED model into consideration, where the
the assumption and architecture of our framework. Then,
user demand for service will depend on its price and flow type.
we present both scheduling and pricing modules in PARS,
Then, the charging strategy adaptively adjusts the fee charged
followed by the design rationale. Afterward, we discuss how
to each flow according to the information of flow type and
to extend the PARS framework to the multi-cell environment.
PRB allocation from the scheduling module. Apparently, the
amount of money charged to a user will be the sum of fees on
A. Assumption and Architecture all flows that his UE uses. Next, we present the detailed design
We classify users into levels of golden (G), silver (S), of both scheduling and pricing modules. Table II summarizes
and bronze (B), with priorities of G > S > B. The eNB the notations used in the PARS framework.
assigns PRBs to UEs by referring to these priorities. PCRF
then measures the amount of resource spent by each UE and B. Scheduling Module
charges its user accordingly, where golden, silver, and bronze
users are charged with high, medium, and low unit prices, 1) GBR Estimation: The 1st-layer scheduler considers only
respectively. Some previous studies restrict the type of flows GBR flows, as shown in Fig. 2. We adopt FLS to evaluate the
that a UE can use (e.g., [20] does not allow bronze UEs to amount of data that should be transmitted for a GBR flow
have GBR flows). In this paper, we assume that each UE can to satisfy its delay requirement (called transmission need).
transmit any type of flows for flexibility and practicability. Specifically, let qi,j (k) and gi,j (k) be the queue length of
Fig. 2 gives the system architecture of our PARS framework, a GBR flow fi,j and its transmission need in the kth frame,
which consists of scheduling and pricing modules installed respectively, where 0 ≤ qi,j (k) ≤ q max , gi,j (k) ≥ 0, and q max
in an eNB and PCRF, respectively. The scheduling module denotes the maximum size of a queue. Then, the variation in
contains three layers of schedulers. The 1st-layer scheduler, queue length can be described by the following equation:
GBR estimation, measures how much resource that each GBR qi,j (k + 1) − qi,j (k) = φi,j (k) − gi,j (k), (4)
flow requires to satisfy its delay constraint. We borrow the idea
from frame layer scheduling (FLS) [9] to do the measurement. where φi,j (k) is the amount of newly generated data to fi,j ’s
The 2nd-layer scheduler, PRB allocation, then decides the queue in the kth frame, and φi,j (k) ≥ 0. To calculate the
number of PRBs given to each flow based on the GBR transmission need of fi,j , FLS defines a control rule by
transmission need (from the lst-layer scheduler), user level gi,j (k) = hi,j (k) ∗ qi,j (k), (5)
(from the pricing module), and CQI. It enhances the max-
CQI and M-LWDF methods [4] to cope with PRB allocation where ‘∗’ denotes the discrete-time convolution and
for GBR and non-GBR flows, respectively. The 3rd-layer hi,j (k) is a pulse-response function. According to [9],
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

WANG AND TSAI: PRICING-AWARE RESOURCE SCHEDULING FRAMEWORK FOR LTE NETWORKS 5

Consequently, the eNB first allocates PRBs to GBR flows to

meet rule 1. Since there may not be sufficient resource to
serve all GBR flows, we thus modify M-LWDF to select a
GBR flow to receive each PRB. In order to apply rule 2, we
define a price-based weight for each UE ui by
Cli
WiP = , (8)
CG + CS + CB
where CG , CS , and CB denote the average amount of money
charged to a golden, silver, and bronze user, respectively, and
Cli ∈ {CG , CS , CB }. Such information can be obtained from
Fig. 3. Calculate the transmission need of a GBR flow by FLS. the pricing module (discussed in Section IV-C). Then, we
select a flow fi,j to obtain the PRB as follows:


we can derive gi,j (k) by combining Eqs. (4) and (5) as follows: ri
fi,j = arg max wi,j di,j × avg × Wi , P
(9)
i,j ri

Mi,j
gi,j (k) = qi,j (k) + ĉi,j (m) × (qi,j (k − m + 1) where wi,j is the original weight defined by M-LWDF. Specif-
m=2 ically, wi,j = − log βi,j /σi,j , where βi,j is the maximum
− qi,j (k − m + 2) − gi,j (k − m + 1)). (6) probability of packet dropping (i.e., di,j > dmax i,j ), and σi,j
denotes the expected delay of fi,j . In Eq. (8), the price-based
In Eq. (6), Mi,j is the sampling interval for the flow. We can
weight WiP is limited to (0, 1], and a higher-level UE will
set Mi,j = dmax max
i,j − 1, where di,j is the maximum tolerant
have a larger WiP value.2 Thus, there is a higher possibility
delay of fi,j (in frames). Thus, once the eNB transmits at least
to pick its GBR flow to receive the PRB by Eq. (9). The eNB
gi,j (k) amount of data for a GBR flow during every frame k,
then iteratively uses Eq. (9) to allocate each PRB, until either
we can guarantee that the flow’s packets will never be dropped
1) all PRBs have been consumed or 2) the transmission need
due to expiration. On the other hand, ĉi,j (m) is a coefficient
gi,j (k) of every GBR flow is satisfied in the current TTI.
that satisfies two conditions [21]:
Afterward, if there still remain PRBs, the eNB distributes
0 ≤ ĉi,j (m) ≤ 1 ∀m ∈ Z+ 0,
them among non-GBR flows. We enhance max-CQI by intro-
ĉi,j (m) ≥ ĉi,j (m + 1), m ≥ 1 with ĉi,j (m) ∈ R. (7) ducing the price-based weight, so as to increase the overall
throughput of non-GBR flows while come up to rule 2:
The coefficient ĉi,j (m) is a real number between zero and one, 
and it monotonically decreases as the parameter m increases ui = arg max ri × WiP . (10)
i
(when m ≥ 1). One possible way to satisfy the conditions in
The eNB then iteratively uses Eq. (10) to allocate the remain-
Eq. (7) is to set ĉi,j (0) = 0, ĉi,j (1) = 1, and ĉi,j (m + 1) =
ing PRBs, until either 1) all PRBs have been allocated or 2)
ĉi,j (m)/2, for m = 1, 2, · · · , Mi,j − 1. In other words, we
the traffic demand of each non-GBR flow is satisfied.
have ĉi,j (2) = 1/2, ĉi,j (3) = 1/4, ĉi,j (4) = 1/8, and so on.
Next, we give two remarks for our PRB allocation in
This setting will be also used by our simulations in Section VI.
the 2nd-layer scheduler. Remark 1 discusses how to deter-
We give an example in Fig. 3 to demonstrate FLS, where
mine the current rate ri of each UE, which is used by
Mi,j = 10 frames. In the kth frame, an amount φi,j (k) =
both Eqs. (9) and (10). Then, Remark 2 addresses how to
1000 bits of data comes to fi,j ’s queue. Based on the above
estimate the number of PRBs actually required by a flow.
setting of coefficient ĉi,j (m), we can spread the enqueued data
Remark 1 (Determining the Current Rate ri of a UE):
over the kth, (k + 1)th, · · · , and (k + 9)th frames to 500,
Both max-CQI and M-LWDF methods are developed for
250, · · · , and 1 bits, respectively. Then, supposing that
general wireless networks [4], so the current rate ri of each
φi,j (k + 1) = 2000 bits and φi,j (k + 2) = 0 bit (i.e., no
UE can be easily determined if there is only one downlink
data generated in the (k + 2)th frame), we can calculate the
channel. However, LTE adopts OFDMA for downlink
amount of data spread over Mi,j observing frames accordingly.
communication, where PRBs may locate in different
Based on Eq. (6), we eventually obtain gi,j (k) = 500 bits,
subchannels that encounter frequency selective fading [22].
gi,j (k + 1) = 1250 bits, gi,j (k + 2) = 625 bits, and so on, as
In other words, ri may not be necessarily the same across all
shown in Fig. 3.
PRBs. Therefore, to make both Eqs. (9) and (10) function
2) PRB Allocation: From Fig. 2, the 1st-layer scheduler
well in the LTE environment, the rate ri can be defined by
passes the transmission need gi,j (k) of each GBR flow to
the data rate supported by the current PRB for a UE ui .
the 2nd-layer scheduler, and then the 2nd-layer scheduler
In fact, the LTE standard [23] defines a CQI table shown in
determines PRB allocation for both GBR and non-GBR flows.
Table III to determine the relationship between efficiency and
Specifically, it obeys two priority rules to allocate PRBs:
MCS (including the code rate) for each CQI value. Through
• [Priority rule 1] GBR flows are given precedence over the CQI table, we can calculate the average number of data
non-GBR ones, as they have stringent delay constraints. bits carried by each PRB when a UE has a certain CQI value
• [Priority rule 2] Golden UEs can acquire network
resource first, followed by silver and bronze UEs. 2 The case of WiP = 1 occurs when all UEs have the same user level.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE/ACM TRANSACTIONS ON NETWORKING

TABLE III respectively. We have ΨGBR ∩ ΨN GBR = ∅ and ΨGBR ∪

CQI TABLE D EFINED IN LTE ΨN GBR ⊆ Ψ, where Ψ denotes the set of available PRBs in
the current TTI. Also, we define two ratios ξGBR and ξN GBR
to respectively control the portion of PRBs to be reallocated in
ΨGBR and ΨN GBR . Below, we separate our discussion into
GBR and non-GBR cases.
Let us denote by pi,j the PRB allocated to a flow fi,j of
UE ui . For the GBR case, we sort all pi,j in ΨGBR by its UE’s
channel rate ri in an increasing order. Then, we create a subset
ΨRGBR ⊆ ΨGBR of candidate PRBs that can be reallocated.
Initially, we have ΨR GBR = ∅. Then, for each pi,j ∈ ΨGBR ,
we add it to ΨR GBR if the user level li ∈ {G, S} (i.e., golden
or silver UEs), until |ΨR GBR | reaches to |ΨGBR | × ξGBR ,
where ‘|·|’ denotes the number of elements in a set and ‘·’ is
the ceiling function. Then, we consider two reallocation rules:
• [Reallocation rule 1] li = G:
If the following condition satisfies
ri < max{ri | li ∈ {S, B} and fi ,j  is GBR}, (12)
which means that another GBR flow fi ,j  owned by a
lower-level UE actually has a higher channel rate to PRB
pi,j , we thus reallocate pi,j to fi ,j  to improve system
for that PRB. In this way, we can also determine its current performance. If Eq. (12) is violated, we remove pi,j from
rate ri (for the PRB) accordingly. ΨRGBR because there is no gain to reallocate the PRB.
Remark 2 (Calculating the Number of PRBs Used by a • [Reallocation rule 2] li = S:
Flow): As mentioned in Remark 1, a flow may have dif- If the following condition satisfies
ferent channel quality across its allocated PRBs. To find the
effective SINR (signal-to-interference-plus-noise ratio) γeff on ri < max{ri | li = B and fi ,j  is GBR}, (13)
these PRBs, we can employ the exponential effective SINR which implies that GBR flow fi ,j  owned by a bronze
mapping (EESM) approach [24] as follows: UE has a higher channel rate to pi,j than its original
1  −γk /ε
s flow fi,j , we reallocate pi,j to fi ,j  to increase network
γeff = EESM(Γ, ε) = −ε ln e , (11) throughput. When Eq. (13) is violated, we remove pi,j
s
k=1 from ΨR GBR as there is no need to reallocate the PRB.
where Γ is a vector [γ1 , γ2 , · · · , γs ] of the tone SINR value After the above examination, ΨR GBR will remain only the
for each subchannel, s is the number of subchannels, and PRBs that have been reallocated to other flows. The scheduling
ε is a tunable parameter (usually set to one). In the LTE module then passes ΨR GBR to the pricing module for extra
implementation, the eNB first uses one PRB to calculate γeff charge (discussed in Section IV-C).
by Eq. (11), and then checks if this PRB has sufficient capacity We deal with the non-GBR case following the above two
(by consulting Table III) to satisfy the traffic demand of reallocation rules. Then, the set ΨR N GBR is also passed to the
flow fi,j in the current TTI. If not, the eNB iteratively adds the pricing module to calculate the extra charge for those low-level
next PRB, recalculates γeff , and repeats the above check, until users who get additional resource by opportunistic resource
fi,j ’s demand becomes satisfied or there is no available PRB. reallocation. Remark 3 discusses the effect of both parameters
In this way, we can estimate the number of PRBs actually ξGBR and ξN GBR on PRB allocation.
used by each flow. Remark 3 (Impact of ξGBR and ξN GBR ): Both ξGBR and
3) Opportunistic Resource Reallocation: By introducing ξN GBR are the ratios of GBR and non-GBR PRBs that will
the price-based weight WiP to Eqs. (9) and (10), we allow be considered to be reallocated, respectively. Based on the
golden UEs to acquire PRBs first (and followed by silver and two reallocation rules in Eqs. (12) and (13), a PRB will be
bronze UEs). However, such PRB allocation may hurt system reallocated if we can find a lower-level UE that has better
performance, especially when high-level UEs encounter bad channel condition (i.e., larger CQI value) to that PRB. In
channel condition. In this case, their PRBs can only use simple other words, when both ξGBR and ξN GBR increase, the overall
MCS and carry quite few data bits (in other words, the spectral throughput could improve as more PRBs can be given to the
resource is wasted). To deal with the problem, the 3rd-layer UEs with the best channel condition. Nevertheless, higher-
scheduler adopts opportunistic resource reallocation, whose level UEs may be forced to give up more PRBs. Let us
idea is to allow a small portion of PRBs to be ‘reallocated’ to consider an extreme case where ξGBR = ξN GBR = 1
low-level UEs according to their channel rates. and bronze UEs have better channel condition than others.
Let ΨGBR and ΨN GBR be the sets of PRBs allocated In this case, the 3rd-layer scheduler reallocates each PRB to
to GBR and non-GBR flows by the 2nd-layer scheduler, the bronze UE that has the largest CQI value. Consequently,
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

WANG AND TSAI: PRICING-AWARE RESOURCE SCHEDULING FRAMEWORK FOR LTE NETWORKS 7

the result of PRB allocation will be the same with that of In Eq. (16), Pe is the extra charge incurred when fi,j uses
the max-CQI method. However, both golden and silver UEs PRBs in ΨRGBR or ΨN GBR . With the PED model, we define
R

will receive no PRB, which violates the principle of UE

Pe = γ − Ed , (19)
classification.
The design of opportunistic resource reallocation is to where γ ≥ max{Ed , ∀fi,j }. For example, we can set γ = 2.5,
improve network throughput under the prerequisite that UEs so Pe will be 1.2, 0.8, and 0.4 for VoIP, video, and non-
are given with resource based on their priorities (i.e., levels). GBR flows. Such setting is feasible due to two reasons. First,
Obviously, the values of ξGBR and ξN GBR should not be set VoIP service has the lowest price elasticity, so we can charge
too large. Therefore, we suggest setting ξGBR ≤ 0.1 and more money to increase operator revenue without significantly
ξN GBR ≤ 0.1 so that no more than 10% of PRBs will be degrading user demands. Second, since non-GBR service has
reallocated. In this way, the price-based weight WiP can have the highest price elasticity, we can reduce its extra charge to
dominating effect on PRB allocation. We will also set both encourage users to utilize such service. Notice that if flow fi,j
ξGBR and ξN GBR to 0.1 in our simulations. does not use any PRB in ΨR GBR ∪ ΨN GBR , its user need not
R

pay for such extra charge. Then, the amount of money charged
C. Pricing Module to a user ui will be the sum of charges to all its flows:

The pricing module refers to the user level and the PRB Ci = Ci,j . (20)
allocation from the scheduling module to charge each user, as ∀fi,j ∈ui
shown in Fig. 2. It also consults PED to model the reaction
of user demand to the change of price. Here, we adopt the In addition, the average amount of money charged to a
PED-related equation in [25], which is used to analyze the golden user can be derived by
⎧
effect of price P̃ on user demand D̃ in wireless networks: ⎪
⎨ {Ci | li = G}
ui
D̃ = λP̃ −Ed , (14) CG = if NG > 0 (21)
⎪ NG
⎩
where λ is a scaling constant to represent the demand 0 otherwise,
potential,3 and Ed is the coefficient of price elasticity. From where NG is the number of golden users. Similarly, we can
Eq. (14), we can derive Ed by compute CS and CB (i.e., the average amount of money
  Ed ln(D̃2 /D̃1 ) charged to a silver and bronze user, respectively) following
D̃2 /D̃1 = P̃1 /P̃2 ⇒ Ed = . (15) Eq. (21). As illustrated in Fig. 2, the parameters CG , CS ,
ln(P̃1 /P̃2 )
and CB will change depending on the number of PRBs allo-
In general, a larger Ed value implies that the user demand is cated to different levels of UEs (determined by the scheduling
relatively elastic. In other words, when the price increases, the module), and they are necessary to calculate the price-based
user demand will decrease more significantly, and vice versa. weight WiP in Eq. (8) used by the 2nd-layer scheduler. This
According to [26] and [27], VoIP and video applications have relationship exhibits that both scheduling and pricing modules
dominated the revenue of most telecommunications operators. can tightly couple with each other.
It implies that VoIP and video flows should have smaller Ed
values, as people usually use these applications. Therefore, D. Design Rationale
we set Ed to 1.3, 1.7, and 2.1 for VoIP, video, and non-GBR
flows, respectively, based on the suggestion in [25]. Most studies discussed in Section III independently cope
To compute the amount of money charged to a flow fi,j with the resource scheduling and pricing problems in LTE.
based on its consumption of network resource, we improve They aim at either improving system performance or increas-
Eq. (2) as follows: ing operator revenue. However, the two objectives may conflict
with each other if we do not take both of them into consider-
Ci,j = [Pv (li ) + Pe ] × (ê − ê−y ) × L, (16) ation. Specifically, when we simply allocate most resource to
where the variable charge is defined by the flows with better channel quality to improve performance,
⎧ the priority of high-level users would be omitted by the
⎪
⎪ Pc if L ≤ δ, scheduler. On the contrary, if we want to increase revenue
⎪
⎨P (G)
v if L > δ and li = G, by giving most resource to high-level users, performance may
Pv (li ) = (17)
⎪
⎪ P v (S) if L > δ and li = S, degrade when their UEs encounter worse channel condition.
⎪
⎩ To conquer this dilemma, our PARS framework proposes
Pv (B) if L > δ and li = B,
two tightly-coupled modules to handle PRB allocation and
y is flow fi,j ’s QCI,4 and the network load is defined by user charge, as shown in Fig. 2. The scheduling module relies
the number of used PRBs on the information of user level and charge from the pricing
L= . (18) module to calculate the price-based weight WiP , which plays
the number of available PRBs
a critical role in assigning PRBs for transmission. On the other
3λ can be set to equal to the value of D̃ when P̃ = 1. hand, the pricing module estimates user charge based on PRB
4 There were originally nine QCIs defined in LTE Release-8 standard [28]. allocation and two sets ΨR GBR and ΨN GBR outputted from
R
However, LTE Release-13 standard [7] adds four new QCIs, 65, 66, 69, and
70, for some special applications such as mission critical data. We thus restrict the scheduling module. In this way, the PARS framework can
y to range between 1 and 9 in Eq. (16) for backward compatibility. balance between system performance and operator revenue.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE/ACM TRANSACTIONS ON NETWORKING

In the scheduling module, there are three special designs to

support QoS for GRB flows:
• We employ FLS in the 1st-layer scheduler to estimate
the amount of data transmission required to meet the
delay constraint of each GBR flow. Thus, the eNB can
try its best to satisfy the transmission needs by the
2nd-layer scheduler. Notice that we modify FLS to com-
pute only the transmission amount used to avoid a GBR
flow dropping its packets. When some users request a
large amount of GBR traffic, non-GBR users can still
have network resource to use their service.
• For the priority rules in Section IV-B.2, we give rule 1
precedence over rule 2 to guarantee GBR transmissions of
low-level users. Let us consider an example where golden
users send a huge amount of non-GBR data. If we simply
satisfy their demands first, silver and bronze users will
never have a chance to receive network resource (i.e.,
starvation), even though they have delay-critical GBR
Fig. 4. Three types of LTE multi-cell networks. (a) homogeneous cells.
traffic. To solve this problem, the eNB first satisfies the (b) heterogeneous cells. (c) cells with relay stations.
‘necessary’ GBR demands (i.e., rule 1), and then gives
the remaining PRBs to non-GBR flows. When dealing
with the GBR and non-GBR cases, rule 2 can ensure that blank subframe, abbreviated to ‘ABS’) to transmit only
golden users have a higher priority to receive resource. low-power signals. Thus, pico-cell eNBs can send their
• GBR and non-GBR flows are scheduled in different ways. data without interference in ABS slots. To apply our
We use the enhanced M-LWDF method to schedule GBR scheduling module to such networks, the macro-cell eNB
flows, which takes care of the urgent degree (i.e., packet will allocate PRBs only in non-ABS slots. On the other
delay) of each flow. Non-GBR flows are scheduled by the hand, pico-cell eNBs can allocate PRBs in ABS slots.
enhanced max-CQI method for performance concern. • Cells with relay stations: Each UE can choose to receive
Furthermore, we propose the mechanism of opportunistic data directly from the eNB or via a relay station. In this
resource reallocation to adjust the scheduling result by replac- network, the eNB and all relay stations share the same
ing some ‘bad’ PRB allocation. In particular, we pick those spectral resource. Therefore, the eNB will partition PRBs
PRBs assigned to high-level UEs that encounter worse channel into three groups: 1) PRBs for the data from the eNB
condition, and check if each of such PRBs can be reallocated directly to UEs, 2) PRBs for the data from the eNB
to another UE with a lower user level but better channel to relay stations, and 3) PRBs for the data from relay
quality. In this case, these low-level users have to pay for extra stations to UEs. How to calculate the number of PRBs
fee (i.e., Pe in Eq. (16)) to receive the reallocating PRBs by the in each group can refer to our previous work [30].
pricing module. Moreover, our pricing module uses Eq. (19) After the partition, both the eNB and relay stations can
to adjust the extra fee according to different service types, separately use our scheduling module to allocate PRBs
which considers the price elasticity defined by the economic in the corresponding groups.
PED model. It thus helps increase operator revenue without On the other hand, since we install the pricing module in
significantly degrading user demands. PCRF (referring to the LTE structure in Fig. 1), there is no
need to modify the module when we switch from the single-
E. Extending to the Multi-Cell Environment cell environment to a multi-cell network. The major reason is
that each piece of data will be associated with the destination
Till now, the discussion of our scheduling module aims at
eNB, so it is easy to allow PCRF to calculate the amount of
a single-cell environment. However, it can be easily extended
data received by each UE in the network.
to a multi-cell environment. Below, we consider three types
of LTE multi-cell networks, as shown in Fig. 4:
• Homogeneous cells: This is the simplest case. Each eNB
V. T HEORETICAL A NALYSIS
independently manages the spectral resource in its cell In this section, we give analysis on performance and com-
without affecting other eNBs. Therefore, we can directly plexity of the PARS framework. For performance analysis, we
apply our scheduling module to each individual eNB. aim at whether the scheduling module can guarantee delay
• Heterogeneous cells: A large macro-cell may contain bound of GBR flows (i.e., QoS support). In particular, our
several small pico-cells. In this case, they may cause scheduling module uses FLS to calculate the GBR transmis-
signal interference with each other. To conquer this sion need in the 1st layer. The objective of FLS is to support
problem, LTE adopts the technique of enhanced inter- bounded-input, bounded-output (BIBO) stability [21], where
cell interference coordination (eICIC) [29]. In eICIC, the output of a system remains bounded in amplitude, provided
the macro-cell eNB will select some slots (called almost that the input is also bounded. In other words, the eNB will
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

WANG AND TSAI: PRICING-AWARE RESOURCE SCHEDULING FRAMEWORK FOR LTE NETWORKS 9

never seek to allocate an infinite bandwidth due to the reason Such charge occurs only when a user receives PRBs from the
that the system input (i.e., the incoming data rate) is bounded opportunistic resource reallocation method. In other words, if
in amplitude as any practical application cannot produce an there is no PRB reallocation, our pricing module will compute
infinite packet rate. Lemma 1 shows that FLS satisfies BIBO the same amount of money with NLP. Otherwise, it can
stability and also indicates its delay bound for GBR flows. further improve revenue comparing with NLP due to extra
Lemma 1: FLS is BIBO stable and ensures that the queuing charge Pe .
delay of any GBR flow fi,j is smaller than Mi,j + 1, where We then analyze the computational complexity. Lemma 2
Mi,j is the flow’s sampling interval. discusses the complexity to run the scheduling module on each
Proof: The proof can be found in [9, Th. 1]. eNB, while Lemma 3 shows the complexity to conduct the
With Lemma 1, the following theorem then proves that the pricing module by PCRF. Theorem 3 finally gives the overall
scheduling module in PARS can meet the delay requirement complexity of our PARS framework.
of GBR flows when the eNB has sufficient downlink resource. Lemma 2: Let NlGBR and NlNGBR be the number of GBR
Theorem 1: Given Rk PRBs supported by an eNB in the and non-GBR flows in the lth cell, respectively. Then, the
kth frame, where the minimum number of data bits carried worst-case complexity of the scheduling module for the lth
by each PRB is kept above bmin , then the scheduling module cell is O(NlGBR (Dlmax −1))+O(Rl ·max{NlGBR , NlNGBR }),
in PARS can guarantee that there is no packet dropping of where Dlmax and Rl are the maximum tolerant delay of
GBR flows due to expiration if the following equation holds flows (in frames) and the number of PRBs (in a TTI) in the
for any k: cell.
 Proof: The scheduling module conducts the three-layer
Rk × bmin ≥ gi,j (k). (22) schedulers in sequence (referring to Fig. 2), so we analyze
∀fi,j

Proof: The scheduling module uses FLS in its 1st-layer the computational complexity of each scheduler separately.
scheduler and sets the sampling interval Mi,j of each GBR Specifically, the 1st-layer scheduler adopts FLS to compute
flow fi,j to dmax max the GBR transmission need, which relies on Eq. (6) to do
i,j − 1, where di,j is the maximum tolerant
delay. According to Lemma 1, if the eNB can allocate enough the computation. Obviously, the computation of gi,j (k) for
resource to meet the transmission demand gi,j (k) of fi,j by a flow fi,j requires (Mi,j − 1) multiplications and (3(Mi,j −
Eq. (6), each of fi,j ’s packet must be delivered before the 1)+1) sums, so the complexity of Eq. (6) is O(Mi,j ). Because
deadline dmax there are NlGBR GBR flows, so we have to repeat Eq. (6) for
i,j . In fact, Eq. (22) indicates the condition of
whether the eNB has sufficient PRBs, where the left term NlGBR times, which spends time of O(NlGBR · Mmax ), where
is the total number of data bits that can be sent out in the Mmax = max{Mi,j }. As mentioned in Section IV-B.1, we set
kth frame, while the right term is the amount of overall GBR Mi,j = dmax
i,j − 1, so Mmax = Dl
max
− 1. Thus, the complexity
transmission need that should be satisfied in the frame. Based of the 1st-layer scheduler will be O(NlGBR (Dlmax − 1)).
on priority rule 1 in Section IV-B.2, GBR flows can always The 2nd-layer scheduler first uses the enhanced M-LWDF
obtain PRBs for transmission first in the 2nd-layer scheduler. scheme to allocate PRBs to GBR flows. From Eq. (9), it takes
This rule implies that the eNB must allocate enough PRBs for O(NlGBR ) time to allocate each PRB because we have to
each GBR flow to avoid dropping its packets, no matter there check every GBR flow. If there still remain PRBs, we use
exist non-GBR flows. Then, the 3rd-layer scheduler deals with the enhance max-CQI scheme to distribute PRBs among non-
GBR and non-GBR PRB reallocation independently. Thus, GBR flows. Based on Eq. (10), it requires O(NlNGBR ) time
there is no possibility that some GBR PRBs will be reallocated to assign a PRB since we should examine every non-GBR
to non-GBR flows. Therefore, the scheduling module can meet flow. When NlGBR ≥ NlNGBR , the worst case occurs if all
the delay requirement of every GBR flow if Eq. (22) holds for PRBs are allocated to GBR flows. If NlGBR < NlNGBR , the
any k, thereby proving this theorem. worst case occurs when the resource is given to only non-GBR
On the other hand, our pricing module enhances the NLP flows. Thus, the complexity of the 2nd-layer scheduler will be
method discussed in Section III-B, which employs the linearity O(Rl · max{NlGBR , NlNGBR }).
factor to estimate the amount of money charged to users: In the 3rd-layer scheduler, two cases are considered. For
the GBR case, the eNB reallocates at most ξGBR · Rl PRBs to
fL (x) = A × (ê − ê−Bx ), (23) silver and bronze UEs by Eqs. (12) and (13), so the complexity
is O(ξGBR · Rl · (NlGBR,S + NlGBR,B )), where NlGBR,S and
where A decides the base level of price while B adjusts
the deduction of price when using high bit rate or volume NlGBR,B respectively denote the number of GBR flows owned
transfers. It has been shown in [19] that the linearity factor of by silver and bronze UEs. Similarly, the non-GBR case will
NLP charging increases operator revenue while considering spend time of O(ξN GBR ·Rl ·(NlNGBR,S +NlNGBR,B )), where
price elasticity. Moreover, Theorem 2 shows that our pricing NlNGBR,S and NlNGBR,B are the number of non-GBR flows
module can further improve revenue than the NLP method. owned by silver and bronze UEs, respectively.
Theorem 2: With the same result of PRB allocation, the Thus, the total complexity is O(NlGBR (Dlmax − 1)) +
pricing module in PARS can receive revenue no less than NLP. O(Rl · max{NlGBR , NlNGBR }) + O(ξGBR · Rl · (NlGBR,S +
Proof: By comparing the pricing equations of NLP and NlGBR,B )) + O(ξN GBR · Rl · (NlNGBR,S + NlNGBR,B )).
the pricing module in Eq. (2) and Eq. (16), it is apparent Because ξGBR ≤ 1, ξN GBR ≤ 1, NlGBR,S + NlGBR,B ≤
that the pricing module will ask users for extra charge Pe . NlGBR , and NlNGBR,S + NlNGBR,B ≤ NlNGBR , we can
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE/ACM TRANSACTIONS ON NETWORKING

simplify the complexity to O(NlGBR (Dlmax − 1)) + O(Rl · TABLE IV

max{NlGBR , NlNGBR }), thereby proving the lemma. S IMULATION PARAMETERS
Lemma 3: Suppose that N is the total number of flows in
an LTE network. Then, the computational complexity of the
pricing module is O(N ) in the worst case.
Proof: The pricing module uses Eq. (16) to compute
the fee of each flow, where the variable charge Pv (li ), extra
charge Pe , and QCI index y can be determined by the flow
itself. Moreover, the network load can be calculated once
by Eq. (18). Thus, it takes O(N ) time to compute the charges
for all flows in the network by Eq. (16). Then, calculating the
fee to each UE requires O(N ) time in Eq. (20), because each
flow belongs to only one UE. Due to the same reason, it also
spends O(N ) time to find the values of CG , CS , and CB by
Eq. (21). Therefore, the worst-case complexity of the pricing
module will be O(N ) + O(N ) + O(N ) = O(N ).
Theorem 3: Given N GBR GBR flows and N N GBR non-
GBR flows in an LTE network, the computational complex-
ity of the PARS framework is O(N GBR (D − 1)) + O(R ·
max{N GBR , N N GBR }) + O(N GBR + N N GBR ), where D is
the maximum tolerant delay of all flows (in frames) and R is
the maximum number of PRBs supported by a cell (in a TTI)
in the worst case.
Proof: Based on the discussion in Section IV-E, each eNB
conducts the scheduling module independently. Let L denotes
the set of all cells in the LTE network. By Lemma 2, the
overall complexity to conduct the scheduling module will be
 mechanism, we use the two-level scheme in [9] to be their
O(NlGBR (Dlmax − 1)) + O(Rl · max{NlGBR , NlNGBR }) scheduling solution, which first adopts FLS to find the amount
l∈L of GBR transmission need and then uses the PF solution to
= O(N GBR (D − 1)) + O(R · max{N GBR , N NGBR }). allocate PRBs. For NLP, SCP, and PARS, the load threshold δ
(24) is 50%. Also, we set both ξGBR and ξN GBR to 0.1 in PARS.
For the PED model, the scaling constant λ in Eq. (14) is
By combining Eq. (24) with Lemma 3, where N is replaced set to 2 × 105 according to [25]. The basic unit for price is
by (N GBR + N N GBR ), we can thus prove the theorem. called monetary unit (abbreviated to ‘mu’). Remark 4 gives a
discussion on our simulation setting.
VI. P ERFORMANCE E VALUATION Remark 4 (Setting in Simulations): LTE-Sim is a popular
We use LTE-Sim [31] to verify the efficiency of the PARS open-source simulator to model LTE behavior. It encom-
framework. Table IV lists the parameters of our simulations. passes critical aspects of LTE networks, including E-UTRAN
We consider an LTE macro-cell where the eNB distributes (evolved universal terrestrial radio access) and EPS (evolved
100 PRBs among UEs in every TTI. Each UE follows the packet system) [33]. Thus, LTE-Sim can provide sophisticated
random direction model [32] to move in the cell with a velocity simulations for not only the wireless communication but also
of 3km/h to imitate human walking. We also vary the number core network in LTE. Hence, we choose to use LTE-Sim for
of UEs to evaluate the effect of different network loads, where performance evaluation.
the number of golden, silver, and bronze UEs are equal. Moreover, to provide a general simulation environment, we
Each UE has two GBR flows: 8.4kbps VoIP flow (QCI = 1) set parameters in Table IV as follows:
and 242kbps H.264 video flow (QCI = 2). It also has a non- • The eNB-related parameters are determined based on the
GBR flow: 400kbps constant-bit-rate (CBR) flow (QCI = 6). common setting of an LTE macro-cell eNB [34].
In each single experiment, the number of golden/silver/bronze • We select GBR and non-GBR flows by consulting the
UEs and their flows do not change. However, we apply the LTE standard [23].
PED model to reflect the relationship between price and traffic • The channel-related parameters are set according to the
demand. Specifically, we make a UE adjust the demand of its LTE specification in [35].
flows in each period (= 100 TTIs). Then, the UE will use • All of the pricing methods (including PARS) use the FLS
Eq. (14) to recalculate the demand of its flows based on the scheme for scheduling, so they can allocate PRBs based
average price that it has to pay in the previous period. on the same amount of GBR transmission need.
We compare our PARS framework with the flat-rate, fixed- • We set the pricing-related parameters of the flat-rate,
PRB, NLP, and SCP methods discussed in Section III-B. fixed-PRB, NLP, and SCP methods by following their
Since these four methods do not have resource scheduling original simulation setting in [20] and [36].
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

WANG AND TSAI: PRICING-AWARE RESOURCE SCHEDULING FRAMEWORK FOR LTE NETWORKS 11

TABLE V
I MPROVEMENT R ATIO BY OUR PARS F RAMEWORK

network load L exceeds the threshold δ in the SCP method,

no matter what type of service is used. It will prevent low-
level UEs from using more network resource. In this case,
SCP inevitably incurs the lowest efficiency ψ(t) among all
methods. Except for MT, our PARS framework can have
higher spectral efficiency than other methods due to three
reasons. First, it flexibly uses the pricing-based M-LWDF
and max-CQI methods to schedule resource for GBR and
non-GBR flows, respectively. Second, PARS employs the
opportunistic resource reallocation mechanism to adaptively
adjust the usage of some PRBs to improve channel utilization.
In particular, when high-level UEs encounter bad channel
condition, their PRBs can be given to other UEs that have
better channel quality. Third, by taking the price elasticity
Fig. 5. Comparison on system performance and operator revenue. (a) spectral into consideration, PARS can encourage users to well utilize
efficiency. (b) operator revenue. such additional resource. These three designs help significantly
improve system performance of the PARS framework.
Fig. 5(b) shows the result of operator revenue, where the
Based on the above parameter setting in our simulations, unit is 106 mu (denoted by ‘Mmu’). Although achieving the
we can provide more fair comparison among different highest spectral efficiency, the revenue of MT is kept quite low,
methods. especially when there are more UEs. The major reason is that
MT does not allocate PRBs based on user level. When high-
A. System Performance Versus Operator Revenue level UEs has ‘slightly’ bad channel condition, they may be
We first measure both system performance and operator starved as most resource will be given to bronze UEs whose
revenue by different methods. To evaluate system performance, users pay less money. The flat-rate method employs a fixed
we employ the concept of spectral efficiency: charging policy, so the only way to increase the revenue is to
 avg increase the number of users. However, it would also congest
r (t)
ψ(t) = ∀i i , (25) the network. Both SCP and PARS result in the highest revenue,
Bd since they will charge users for extra money. However, SCP
where riavg (t) is the average data rate of a UE ui at the ‘punishes’ low-level users for using PRBs when L > δ
measuring time t (in particular, 100 seconds in our simulation), according to Eq. (3). It will discourage bronze UEs to well
and Bd is the downlink channel bandwidth. Specifically, the utilize network resource even when golden and silver UEs do
spectral efficiency is the total information rate that can be sent not have much traffic demand. On the contrary, our PARS
over a given bandwidth in the LTE network, and its unit is framework adaptively adjusts the extra charge Pe in Eq. (19)
‘bps/Hz’. Apparently, higher efficiency ψ(t) indicates that the by considering the PED model, and such charge occurs only
eNB can use the downlink channel more efficiently to transmit when low-level UEs use additional PRBs originally assigned
data, thereby achieving better system performance. to high-level UEs. That is why PARS can have higher spectral
To better illustrate the tradeoff between system performance efficiency than SCP in Fig. 5(a), even though the values of
and operator revenue, we develop a method, called maximum their revenue Fig. 5(b) are close to each other.
throughput (MT), which has a bias in favor of the spectral We summarize the improvement ratio by PARS in Table V
efficiency. In particular, MT uses the max-CQI scheme for and give our observations in this experiment as follows:
resource allocation, where each PRB is always given to the UE • The MT method optimizes system performance by greed-
with the best channel condition. Besides, MT adopts the fixed- ily matching each PRB with the UE that achieves the
PRB method as its pricing mechanism, so it also classifies highest data rate. However, without considering user level
users into golden, silver, and bronze levels. and pricing, it will greatly decrease operator revenue.
Fig. 5(a) presents the result of spectral efficiency. It is • The SCP method seeks to maximize operator revenue by
expected that the MT method will result in the highest asking low-level users to pay for more money when the
efficiency ψ(t), because it only picks the UE for each PRB network load becomes heavy. Nevertheless, it ignores the
that can achieve the highest data rate. On the other hand, case where most of the network load is contributed by sil-
users will be charged with much more money when the ver or bronze UEs, thereby hurting system performance.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

12 IEEE/ACM TRANSACTIONS ON NETWORKING

Fig. 6. Comparison on network throughput by different types of flows. (a) VoIP flows. (b) video flows. (c) CBR flows

• Our PARS framework considers both scheduling and in SCP. On the other hand, PARS lets low-level UEs use
pricing modules to balance between system performance the PRBs allocated to high-level UEs if they have better
and operator revenue. The scheduling module works on channel quality, so it could increase CBR throughput.
the basis of user level but allows exchanging the usage of To sum up, our PARS framework can achieve higher video
some PRBs to improve performance. The pricing module and CBR throughput compared with the flat-rate, fixed-PRB,
takes the PED model into account and asks users for extra NLP, and SCP methods in the experiment.
charge depending on their service types. Thus, PARS can
keep high revenue (almost the same with that of SCP) at
C. Network Throughput by Users
the expense of degrading a small amount of throughput
(i.e., around 10% of performance loss than MT). Next, we study the throughput of different UEs, as given in
Fig. 7. Two scenarios are considered: light-load (48 UEs) and
B. Network Throughput by Flows heavy-load (96 UEs). In the light-load scenario, all methods
have similar VoIP and video throughput (referring to Fig. 7(a)
We then evaluate the throughput of different flows, as shown
and (b)), because they allocate PRBs to GBR flows first and
in Fig. 6. Since all methods use FLS to estimate the amount
the eNB has sufficient resource to satisfy GBR demand. Thus,
of GBR transmission need and let GBR flows receive resource
the effect of user level becomes insignificant. For CBR service
first, VoIP and video throughput increases as the number of
in Fig. 7(c), golden UEs have higher throughput than others
UEs grows. However, video flows have quite larger demand
due to their high priority. Since PARS achieves higher spectral
(242kbps) than VoIP flows (8.4kbps), so video throughput
efficiency in Fig. 5(a), there remain more PRBs for CBR
starts decreasing when there are more than 72 UEs. On the
transmission. Such PRBs are given to UEs by their price-based
other hand, because non-GBR flows can receive resource only
weight WiP , so PARS can have much higher CBR throughput
after the eNB has met GBR demand, CBR throughput will
for both golden and silver users than other methods.
significantly decrease when the number of UEs increases.
In the heavy-load scenario, user level has less impact on
Since the PED model reflects the relationship between
VoIP throughput in Fig. 7(d), as VoIP service has small
demand and price, the traffic demand of flows would dynam-
traffic demand. On the contrary, video service requests a large
ically interact with the behavior of different methods. From
amount of data transmission, so golden UEs are allocated
Fig. 6, we have some observations about such interaction:
with more PRBs for video transmission, followed by silver
• VoIP flows: VoIP throughput of each method is similar
and bronze UEs, as shown in Fig. 7(e). However, our PARS
due to two reasons. First, VoIP has the smallest price framework allows low-level UEs to get additional PRBs when
elasticity Ed = 1.3 in Eq. (14), which means that users high-level UEs incur bad channel condition, thereby resulting
may not significantly reduce VoIP traffic even when the in much higher video throughput for bronze users than other
price is raised. Second, a VoIP flow has light traffic load, methods. On the other hand, since the limited network resource
so it is easy to satisfy VoIP demand. is competed by many UEs, non-GBR CBR service can only
• Video flows: Comparing with VoIP, video flows have
receive few PRBs for transmission, as shown in Fig. 7(f).
larger price elasticity (Ed = 1.7) and very heavy traffic
load. For the SCP method, since it substantially raises
the (extra) price when the network load becomes heavy, D. QoS Support for GBR Flows
users (especially for low-level ones) are thus inclined to We finally investigate QoS support for GBR flows by dif-
decrease their video demand. On the contrary, our PARS ferent methods. Here, we measure the packet loss rate of VoIP
framework takes the PED model into account, so it does and video flows due to expiration, as shown in Fig. 8. Based on
not burden users with much money on receiving video the LTE standard [7], when the latency of a VoIP/video packet
service. It therefore can improve video throughput. exceeds 100ms/150ms, it is dropped. Obviously, a lower loss
• CBR flows: With the largest price elasticity (Ed = 2.1) rate implies that the method well supports QoS for GBR flows.
and traffic load, CBR throughput decreases drastically As discussed earlier, VoIP flows have small demand, so
when there are more UEs. Such effect is more obvious their loss rate can be kept low in Fig. 8(a). However, when
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

WANG AND TSAI: PRICING-AWARE RESOURCE SCHEDULING FRAMEWORK FOR LTE NETWORKS 13

Fig. 7. Comparison on network throughput by different levels of users. (a) VoIP flows (48 UEs). (b) video flows (48UEs). (c) CBR flows (48 UEs).
(d) VoIP flows (96 UEs). (e) video flows (96 UEs). (f) CBR flows (96 UEs).

framework can better utilize PRBs for transmission and allow

UEs to exchange the usage of PRBs by their channel quality,
PARS thus greatly reduces the video loss rate, as shown
in Fig. 8(b).

VII. C ONCLUSION
Both issues of resource scheduling and pricing are important
in LTE research. However, previous studies deal with them
independently and face a difficult choice between improving
system performance or increasing operator revenue. To con-
quer this difficulty, we develop the PARS framework with
two tightly-coupled modules. The scheduling module adopts
a three-layer strategy to apply priority rules to PRB allocation
based on user level and flow type. It also allows the eNB
to exchange the usage of some PRBs to improve channel
utilization. The pricing module keeps the PED model in
mind and charges users for extra money depending on the
service they use, so as to increase revenue without significantly
degrading user demand. Through LTE-Sim experiments, we
compare PARS with the flat-rate, fixed-PRB, NLP, and SCP
methods. The results demonstrate that PARS strikes a good
balance between performance and revenue, and also supports
QoS for VoIP and video service.
Fig. 8. Comparison on GBR packet loss rate. (a) VoIP flows. (b) video flows.
R EFERENCES
[1] S. Srikanth, P. A. Murugesa Pandian, and X. Fernando, “Orthog-
onal frequency division multiple access in WiMAX and LTE:
the number of UEs exceeds 72, the VoIP loss rate starts A comparison,” IEEE Commun. Mag., vol. 50, no. 9, pp. 153–161,
increasing in flat-rate, NLP, and SCP. On the other hand, video Sep. 2012.
flows incur a significantly high loss rate when the number [2] Z. Shen, A. Papasakellariou, J. Montojo, D. Gerstenberger, and F. Xu,
“Overview of 3GPP LTE-advanced carrier aggregation for 4G wireless
of UEs grows, as it is not easy for the eNB to satisfy their communications,” IEEE Commun. Mag., vol. 50, no. 2, pp. 122–130,
large demand. Since the scheduling module of our PARS Feb. 2012.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

14 IEEE/ACM TRANSACTIONS ON NETWORKING

[3] Y. C. Wang and C. A. Chuang, “Efficient eNB deployment strategy [25] S. Lanning, D. Mitra, Q. Wang, and M. Wright, “Optimal planning for
for heterogeneous cells in 4G LTE systems,” Comput. Netw., vol. 79, optical transport networks,” Phil. Trans. Roy. Soc. London A, vol. 358,
pp. 297–312, Mar. 2015. no. 1773, pp. 2183–2196, Aug. 2000.
[4] F. Capozzi, G. Piro, L. A. Grieco, G. Boggia, and P. Camarda, [26] Global Knowledge, Using Voice Over IP (VoIP) in Mobile Networks.
“Downlink packet scheduling in LTE cellular networks: Key design (2011). [Online]. Available: http://www.globalknowledge.nl/
issues and a survey,” IEEE Commun. Surveys Tut., vol. 15, no. 2, [27] Ericsson, Voice and Video Calling Over LTE. (2014). [Online]. Avail-
pp. 678–700, May 2013. able: http://www.ericsson.com/
[5] L. A. DaSilva, “Pricing for QoS-enabled networks: A survey,” IEEE [28] European Telecommunications Standards Institute, Policy and Charging
Commun. Surveys Tut., vol. 3, no. 2, pp. 2–8, 2nd Quart., 2000. Control Architecture (Release 8), 3GPP, document TS 23.203, 2010.
[6] N. G. Mankiw, Principles of Economics, 7th ed. Boston, MA, USA: [29] B. Soret, H. Wang, K. I. Pedersen, and C. Rosa, “Multicell cooperation
Cengage Learning, Inc., 2014. for LTE-advanced heterogeneous network scenarios,” IEEE Wireless
[7] European Telecommunications Standards Institute, Policy and Charging Commun., vol. 20, no. 1, pp. 27–34, Feb. 2013.
Control Architecture (Release 13), 3GPP, document TS 23.203, 2015. [30] J. M. Liang, Y. C. Wang, J. J. Chen, J. H. Liu, and Y. C. Tseng, “Energy-
[8] H. Luo, S. Ci, D. Wu, J. Wu, and H. Tang, “Quality-driven cross-layer efficient uplink resource allocation for IEEE 802.16j transparent-relay
optimized video delivery over LTE,” IEEE Commun. Mag., vol. 48, no. 2, networks,” Comput. Netw., vol. 55, no. 16, pp. 3705–3720, Nov. 2011.
pp. 102–109, Feb. 2010. [31] LTE Simulator. (2016) [Online]. Available: http://telematics.poliba.
[9] G. Piro, L. A. Grieco, G. Boggia, R. Fortuna, and P. Camarda, “Two- it/index.php/en/lte-sim
level downlink scheduling for real-time multimedia services in LTE [32] W. H. Yang, Y. C. Wang, Y. C. Tseng, and B. S. P. Lin, “Energy-efficient
networks,” IEEE Trans. Multimedia, vol. 13, no. 5, pp. 1052–1065, network selection with mobility pattern awareness in an integrated
Oct. 2011. WiMAX and WiFi network,” Int. J. Commun. Syst., vol. 23, no. 2,
[10] M. Iturralde, A. Wei, T. Ali-Yahiya, and A. L. Beylot, “Resource pp. 213–230, Feb. 2010.
allocation for real time services using cooperative game theory and a [33] G. Piro, L. A. Grieco, G. Boggia, F. Capozzi, and P. Camarda,
virtual token mechanism in LTE networks,” Wireless Pers. Commun., “Simulating LTE cellular systems: An open-source framework,” IEEE
vol. 72, no. 2, pp. 1415–1435, Sep. 2013. Trans. Veh. Technol., vol. 60, no. 2, pp. 498–513, Feb. 2011.
[11] B. Liu, H. Tian, and L. Xu, “An efficient downlink packet scheduling [34] E. Dahlman, S. Parkvall, and J. Skold, 4G: LTE/LTE-Advanced for
algorithm for real time traffics in LTE systems,” in Proc. IEEE Consum. Mobile Broadband. Amsterdam, The Netherlands: Elsevier, 2013.
Commun. Netw. Conf., Jan. 2013, pp. 364–369. [35] European Telecommunications Standards Institute, LTE; Evolved Uni-
[12] W. K. Lai and C. L. Tang, “QoS-aware downlink packet scheduling versal Terrestrial Radio Access (E-UTRA); Radio Frequency (RF)
for LTE networks,” Comput. Netw., vol. 57, no. 7, pp. 1689–1698, Requirements for LTE Pico Node B, 3GPP, document TR 136, vol. 931,
May 2013. 2011.
[13] C. Wang and Y. C. Huang, “Delay-scheduler coupled throughput- [36] A. Belghith, S. Trabelsi, and B. Cousin, “Realistic per-category pricing
fairness resource allocation algorithm in the long-term evolution wireless schemes for LTE users,” in Proc. IEEE Int. Symp. Modeling Optim.
networks,” IET Commun., vol. 8, no. 17, pp. 3105–3112, Nov. 2014. Mobile, Ad Hoc, Wireless Netw., May 2014, pp. 429–435.
[14] Y. C. Wang and S. Y. Hsieh, “Service-differentiated downlink flow
scheduling to support QoS in long term evolution,” Comput. Netw.,
vol. 94, pp. 344–359, Jan. 2016.
[15] M. S. Mushtaq, S. Fowler, A. Mellouk, and B. Augustin,
“QoE/QoS-aware LTE downlink scheduler for VoIP with power saving,”
J. Netw. Comput. Appl., vol. 51, pp. 29–46, May 2015. You-Chiun Wang (SM’13) received the Ph.D.
[16] F. Khan and N. Baker, “Charging data dimensioning in 3G mobile degree in computer science from National Chiao-
networks,” in Proc. IEE Int. Conf. 3G Mobile Commun. Technol., Tung University, Taiwan, in 2006. He is currently
Oct. 2004, pp. 183–187. an Associate Professor with the Department of Com-
[17] J. Cushnie, D. Hutchison, and H. Oliver, “Evolution of charging puter Science and Engineering, National Sun Yat-sen
and billing models for GSM and future mobile Internet services,” in University, Taiwan. He has authored over 60 papers
Quality of Future Internet Services. Berlin, Germany: Springer, 2002, and chapters. His research interests include wireless
pp. 312–323. communications, mobile computing, and sensor net-
[18] P. Marbach, “Analysis of a static pricing scheme for priority services,” works. He served as TPC members of over 100 con-
IEEE/ACM Trans. Netw., vol. 12, no. 2, pp. 312–325, Apr. 2004. ferences including INFOCOM, ICDCS, and WCNC.
[19] E. Wallenius and T. Hamalainen, “Pricing model for 3G/4G net-
works,” in Proc. IEEE Int. Symp. Pers. Indoor Mobile Radio Commun.,
Sep. 2002, pp. 187–191.
[20] U. Mir and L. Nuaymi, “LTE pricing strategies,” in Proc. IEEE Veh.
Technol. Conf., Jun. 2013, pp. 1–6.
[21] K. J. Astrom and B. Wittenmark, Computer Controlled Systems: Theory
Tzung-Yu Tsai received the M.S. degree in com-
and Design, 3rd ed. New York, NY, USA: Dover, 2012. puter science from National Sun Yat-sen University,
[22] H. Lee, S. Vahid, and K. Moessner, “A survey of radio resource man- Taiwan, in 2015. His research interest aims at
agement for spectrum aggregation in LTE-advanced,” IEEE Commun.
LTE networks.
Surveys Tut., vol. 16, no. 2, pp. 745–760, 2nd Quart., 2014.
[23] European Telecommunications Standards Institute, Evolved Universal
Terrestrial Radio Access (E-UTRA); Physical Layer Procedures, 3GPP,
document TS 36.213, 2012.
[24] R. Giuliano and F. Mazzenga, “Exponential effective SINR approxima-
tions for OFDM/OFDMA-based cellular system planning,” IEEE Trans.
Wireless Commun., vol. 8, no. 9, pp. 4434–4439, Sep. 2009.