Network Slicing: Market Mechanism and

Competitive Equilibria
Panagiotis Promponas, and Leandros Tassiulas
Department of Electrical Engineering and Institute for Network Science, Yale University, USA
{panagiotis.promponas, leandros.tassiulas}@yale.edu

Abstract—Towards addressing spectral scarcity and enhancing direction, network slicing is a promising technology that
resource utilization in 5G networks, network slicing is a enables the transition from one-size-fits-all to one-size-per-
arXiv:2301.02840v2 [cs.NI] 10 Jan 2023

promising technology to establish end-to-end virtual networks service abstraction [1], which is customized for the distinct
without requiring additional infrastructure investments. By
leveraging Software Defined Networks (SDN) and Network use cases in a contemporary 5G network model.
Function Virtualization (NFV), we can realize slices completely Using Software Defined Networks (SDN) and Network
isolated and dedicated to satisfy the users’ diverse Quality Function Virtualization (NFV), those slices are associated with
of Service (QoS) prerequisites and Service Level Agreements completely isolated resources that can be tailored on-demand
(SLAs). This paper focuses on the technical and economic to satisfy the diverse QoS prerequisites and SLAs. Resource
challenges that emerge from the application of the network
slicing architecture to real-world scenarios. We consider a market allocation in network slicing plays a pivotal role in load
where multiple Network Providers (NPs) own the physical balancing, resource utilization and networking performance
infrastructure and offer their resources to multiple Service [2]. Nevertheless, such a resource allocation model faces
Providers (SPs). Then, the SPs offer those resources as slices various challenges in terms of isolation, customization, and
to their associated users. We propose a holistic iterative model end-to-end coordination which involves both the core but also
for the network slicing market along with a clock auction that
converges to a robust ǫ-competitive equilibrium. At the end of the Radio Access Network (RAN) [3].
each cycle of the market, the slices are reconfigured and the SPs In a typical network slicing scenario, multiple Network
aim to learn the private parameters of their users. Numerical Providers (NPs), own the physical infrastructure and offer
results are provided that validate and evaluate the convergence their resources to multiple Service Providers (SPs). Possible
of the clock auction and the capability of the proposed market services of the SPs include e-commerce, video, gaming, virtual
architecture to express the incentives of the different entities of
the system. reality, wearable smart devices, and other IoT devices. The
SPs offer their resources as completely isolated slices to their
Index Terms—Network Slicing, Mechanism Design, Network associated users. Thereby, such a system contains three types
Economics, Bayesian Inference
of actors that interact with each other and compete for the same
resources, either monetary or networking. This paper focuses
I. I NTRODUCTION
on the technical and economic challenges that emerge from
The ascending trend of the volume of the data traffic, the application of this architecture to real-world scenarios.
as well as the vast number of connected devices, puts
pressure on the industries to enhance resource utilization in A. Related Work
5G wireless networks. With the advent of 5G networks and User Satisfaction & Sigmoid Functions: Network
Internet of Things (IoT), researchers aim at a technological applications can be separated into elastic (e.g. email, text
transformation to simultaneously improve throughput, extend file transfer) and inelastic (e.g. audio/video phone, video
network coverage and augment the users’ quality of service conference, tele-medicine) [4]. Utilities for elastic applications
without wasting valuable resources. Despite the significant are modeled as concave functions that increase with the
advances brought by the enhanced network architectures and resources with diminishing returns [4]. On the other hand,
technologies, spectral scarcity will still impede the realization the utility function for an inelastic traffic is modeled as
of the full potential of 5G technology. a non-concave and usually as a sigmoid function. Such
In the future 5G networks, verticals need distinct network non-concavities impose challenges for the optimization of
services as they may differ in their Quality of Service (QoS) a network, but are suitable with the 5G era where the
requirements, Service Level Agreements (SLAs), and key services may differ in their QoS requirements [5]. In that
performance indicators (KPIs). Such a need highlights the direction, multiple works in the literature employ sigmoid
inefficiency of the previous architecture technologies which utility functions for the network users [5]–[13]. Nevertheless,
were based on a ”one network fits all” nature. In this all of these works consider either one SP and model the
interaction between the users, or multiple SPs that compete
This paper appeared in INFOCOM 2023. for a fixed amount of resources (e.g. bandwidth).
The research work was supported by the Office of Naval Research under
project numbers N00014-19-1-2566, N00173-21-1-G006 and by the National Network Slicing in 5G Networks: Network slicing
Science Foundation under the project number CNS-2128530. introduces various challenges to the resource allocation in 5G
networks in terms of isolation, customization, elasticity, and double-sided auction, and thus appropriate for a market with
end-to-end coordination [2]. Most surveys on network slicing multiple NPs. Numerical results are provided that validate
investigate its multiple business models motivated by 5G, the and evaluate the convergence of the clock auction and the
fundamental architecture of a slice and the state-of-the-art capability of the proposed market architecture to express the
algorithms of network slicing [2], [14], [15]. Microeconomic incentives of the different entities of the system.
theories such as non-cooperative games and/or mechanism
design arise as perfect tools to model the trading of network II. M ARKET M ODEL & I NCENTIVES
infrastructure and radio resources that takes place in network
In this section we describe the different entities of the
slicing [9], [16]–[18].
network slicing market and their conflicting incentives.
Mechanism Design in Network Slicing: Multiple auction A. Market Model
mechanisms have been used to identify the business model
of a network slicing market (see a survey in [16]). Contrary A typical slicing system model [2], [3], [14], [15] consists of
to our work, the majority of the literature considers a single- multiple SPs represented by M = {1, 2, . . . , M } and multiple
sided auction, a model that assumes that a single NP owns the NPs that own RANs of possibly different RATs, represented
whole infrastructure of the market [9], [18]–[22]. For example, by a set K = {1, 2, . . . , K}. Each SP owns a slice with a
[9] considers a Vickrey–Clarke–Groves (VCG) auction-based predetermined amount of isolated resources (e. g., bandwidth)
model where the NP plays the role of an auctioneer and and is associated with a set of users, Um , that serves through its
distributes discrete physical resource blocks. We find [3] and slices. For the rest of the paper and without loss of generality
[17] to be closer to our work, since the authors employ we assume that each NP owns exactly one RAN, so we use
the double-sided auction introduced by [23] to maximize the the terms RAN and NP interchangeably.
social welfare of a system with multiple NPs. Contrary to our 1) Network Providers: The multiple NPs of the system
work, the auction proposed in [23] assumes concave utility can quantify their radio resources as the performance level of
functions for the different actors and requires the computation the same network metric (e.g., downlink throughput) [3]. Let
of their gradients for its convergence. The aforementioned x(m,k) denote the amount of resources NP k allocates to SP
assumptions might lead to an over-simplification of a more m, and the vector xm := (x(m,k) )k∈K to denote the amount of
complex networking architecture (e.g. that of the network resources m gets from every NP. Without loss of generality [3],
slicing model) where the utility function for a user with capacity Ck limitsPthe amount of resources that can be offered
inelastic traffic is expressed as a sigmoid function [9] and that from NP k, i.e., M m=1 x(m,k) ≤ Ck . Let C = (Ck )k∈K . For
of an SP as an optimization problem [3]. the rest of the paper, we assume that there is a constant cost
related to operation and management overheads induced to the
B. Contributions NP. The main goal of every NP k is to maximize its profits
Our work develops an iterative market model for the by adjusting the price per unit of resources, denoted by ck .
network slicing architecture, where multiple NPs with 2) Service Providers & Associated Users: The main goal of
heterogeneous Radio Access Technologies (RATs), own the an SP is to purchase resources from a single or multiple NPs in
physical infrastructure and offer their resources to multiple order to maximize its profit, which depends on its associated
SPs. The latter offer the resources as slices to their associated users’ satisfaction. The connectivity of a user i ∈ Um is
users. Specifically, we propose a f ive-step iterative model denoted by a vector βi = (β(k,i) )k∈K , where β(k,i) is a non-
for the network slicing market that converges to a robust ǫ- negative number representing factors such as the link quality
competitive equilibrium even when the utility functions of i.e., numbers in (0, 1] that depend on the path loss. Moreover,
the different actors are non-concave. In every cycle of the each user i of the SP m, is associated with a service class,
proposed model, the slices are reconfigured and the SPs learn c(i), depending on their preferences. We denote the set of the
the private parameters of their associated end-users to make the possible service classes of SP m as C m = {C1m , . . . , Ccmm }
equilibrium of the next cycle more efficient. The introduced and thus c(i) ∈ C m , ∀i ∈ Um . Each SP m, is trying to
market model, can be seen as a framework that suits well distribute the resources purchased from the NPs, i.e., xm ,
to various networking problems where three types of actors to maximize its profit. This process, referred to as intra-slice
are involved: those who own the physical infrastructure, those resource allocation, is described in detail in Section II-B.
who lease part of it to sell services and those who enjoy the Throughout the paper, we assume that the number of users
services (e.g. data-offloading [23]). of every SP m, i.e., |Um |, is much greater than the number
For the interaction between the SPs and the NPs and for of SPs, which is much greater than the number of NPs in
the convergence of the market to an equilibrium, we propose the market. This assumption is made often in the mechanism
an iterative clock auction. Such dynamic auctions are used design literature and is sufficient to ensure that the end-users
in the literature to auction divisible goods [24], [25]. The and the SPs have limited information of the market [23], [26].
key differentiating aspects of the proposed auction, are (i) the The latter let us consider them as price-takers. In the following
relaxation of the common assumptions that the utility functions section, we describe in detail the intra-slice resource allocation
are concave and their gradients can be analytically computed, problem from the perspective of an SP who tries to maximize
(ii) it provides highly usable price discovery, and (iii) it is a the satisfaction of its associated users.
B. Intra-Slice Resource Allocation slice resource allocation. Hence, SP m can solve the following
The problem of the intra-slice resource allocation concerns problem
the distribution of the resources, xm , from the SP m to
its associated users. Specifically, every SP m allocates a (P): max Ψm (rm , xm ) := um (rm ) − cT xm
rm ,xm
portion of x(m,k) to its associated user i, denoted as r(k,i) . s.t. ri  0, ∀i ∈ Um
Let ri := (r(k,i) )k∈K and rm := (ri )i∈Um . For ease of X
notation, the resources, ri , of a user i ∈ Um , as well as the xm  ri
i∈Um
connectivities, βi , are not indexed by m because i is assumed
to be a unique identifier for the user. Although every user i Recall that ck denotes the price per unit of resources
is assigned with r(k,i) resources from RAN k, because of its announced from every NP k. In Problem P , the objective
connectivity βi , the aggregated amount of resources it gets function Ψm can be thought of as the profit of SP m. Let the
is zi := βiT ri . Moreover, let zm := (zi )i∈Um .P In a feasible solution of the above problem be ψm ∗
.
intra-slice allocation it should hold that xm  i∈Um ri for
Problems IN-SL and P are maximization problems of
each SP m.
a summation of sigmoid functions over a linear set of
Every SP should distribute the obtained resources among
constraints. In [27] the problem of maximizing a sum of
its users to maximize their satisfaction. Towards providing
sigmoid functions over a convex constraint set is addressed.
intuition behind the employment of sigmoidal functions in the
This work shows that this problem is generally NP-hard and
literature to model user satisfaction (e.g. see [5]–[12]), note
it proposes an approximation algorithm, using a branch-and-
that by making the same assumption as logistic regression,
bound method, to find an approximate solution to the sigmoid
we model the logit1 of the probability that a user is satisfied,
programming problem.
as a linear function of the resources. Hence, the probability
In the rest of the section, we study three variations of
that user i is satisfied with the amount of resources zi , say
P [QoS sati ] z problem P. Specifically, in Section II-B1, we study the case
P [QoS sati ], satisfies log( 1−P [QoS sati ] ) = tc(i) (zi − kc(i) ) where the end-users are charged to get the resources from
and thus:
z
the SPs and in Sections II-B2 and II-B3 we regularize and
etc(i) (zi −kc(i) ) concavify P respectively, something that will facilitate the
P [QoS sati ] = tz (z −k )
, (1)
1 + e c(i) i c(i) analysis of the rest of the paper.
where kc(i) ≥ 0 denotes the prerequisite amount of resources 1) Price Mechanism in P: In this subsection we argue that
of the user i and tzc(i) ≥ 0 expresses how ”tight” this Problem P is expressive enough to capture the case where
prerequisite is. Note that the probability of a user being every user i is charged for its assigned resources. Let pi be
satisfied with respect to the value of zi , is a sigmoid function the amount of money that user i should pay to receive the zi
with inflection point kc(i) . We assume that the user’s service resources. In that case, the SPs should modify Problems IN-
class fully determines its private parameters, hence every user SL and P accordingly. First, note that user i’s satisfaction may
i ∈ c(i) has QoS prerequisite kc(i) and sensitivity parameter depend also on pi . Similarly with the previous section, we can
tzc(i) . These parameters are unknown to the users, so the SP’s express the satisfaction of user i with respect to the price pi
1
goal to eventually learn them is challenging (Section III-C). using a sigmoid function as P [price sati ] = t
p
(pi −bc(i) )
,
1+e c(i)
Given the previous analysis, the aggregated satisfaction of where bc(i) ≥ 0 is the budget of the user i for the prerequisite
resources kc(i) , and tpc(i) ≥ 0 expresses how ”tight” is this
P
the users of the SP m is um (rm ) := i∈Um ui (ri ) ( [10],
[7]), where budget. We can now model the acceptance probability function
z T [7] as P [sati ] = P [price sati ]P [QoS sati ], and hence the
etc(i) (βi ri −kc(i) ) ′
expected total revenue, or the new utility of SP m, um , is
ui (ri ) := z T . (2)
1 + etc(i) (βi ri −kc(i) ) modeled as
Note that the function ui (·) can be expressed as a function of ′ X
zi as well. With a slight abuse of notation, we switch between um (rm , pm ) := P [sati ]pi . (3)
i∈Um
the two by changing the input variable. We can write the final
optimization problem for the intra-slice allocation of SP m as: From Eq. (3), it is possible for SP m to immediately determine
(IN-SL): max um (rm ) the optimal price p̂i to ask from any user i ∈ Um . This follows
rm
from the fact that for positive pi the function admits a unique
s.t. ri  0, ∀i ∈ Um critical point, p̂. Therefore, by just adding proper coefficients
X
xm  ri to the terms of Problem IN-SL and P, we can embed a pricing
i∈Um mechanism for the end-users in the model. For the rest of the
paper, without loss of generality in our model, we assume that
In case the amount of resources obtained from every NP, xm ,
the end-users are not charged for the obtained resources.
is not given, SP m can optimize it together with the intra-
2) Regularization of P : We can regularize Problem P ,
p
1 The logit function is defined as logit(p) = log( 1−p ). with a small positive λm . In that manner, we encourage dense
solutions and hence we avoid situations where a problem in Therefore, SP m can concavify P̃ as follows:
one RAN completely disrupts the operation of the SP.
fˆi (ri , zi ) − cT xm − λm kxm k22
X
(P̂ ): max
(P̄ ): max Ψm (rm , xm ) − λm kxm k22 rm ,xm ,zm
i∈Um
rm ,xm
s.t. ri  0, ∀i ∈ Um s.t. (ri , zi ) ∈ Si , ∀i ∈ Um
X
xm 
X
xm  ri ri
i∈Um i∈Um

In the regularized problem P̄ , note that larger values of Note that P̂ is strongly concave and thus admits a unique
λm penalize the vectors xm with greater L2 norms. Let the maximizer. Let the solution and the optimal point of problem
∗ ∗
solution of Problem P̄ be ψ̄m . The Lemma below, shows that P̂ be ψ̂m and (x̂∗m , r̂m
∗
) respectively. Ultimately, we would
∗ ∗
for small λm , the optimal values ψ̄m and ψm are close. Its like to compare the solution of the concavified P̂ with the
proof is simple and thus ommited for brevity. one of the original problem P . Towards that direction, we
∗ first define the nonconcavity of a function as follows [28]:
Lemma 1. Let (rm , x∗m ) and (r̄m
∗
, x̄∗m ) be solutions of
Problems P and P̄ respectively. Then, Definition 1 (Nonconcavity of a function). We define the
∗ nonconcavity ρ(f ) of a function f : S → R with domain
ψm − λm kx∗m k22 ≤ ∗
ψ̄m ≤ ∗
ψm − λm kx̄∗m k22
S, to be
Lemma 1, proves that the regularization of P was (almost) ρ(f ) = sup(fˆ(x) − f (x)).
without loss of optimality. In the next section, we proceed by x

concavifying Problem P̄ . The new concavified problem will Let F denote a set of possibly non-concave functions. Then
be a fundamental building block of the auction analysis in define ρ[j] (F ) to be the jth largest of the nonconcavities of
Section III-A. the functions in F . The theorem below, summarizes the main
3) Concavification of P̄ : To concavify P̄ , we replace result of this section, which is that every SP can solve the
every summand of um with its tightest concave envelope, concavified P̂ instead of the original P , since the former
i.e., the pointwise infimum over all concave functions that are provides a constant bound approximation of the latter. Recall
greater or equal. For the sigmoid function ui (zi ) the concave that Ψm (r̂m ∗
, x̂∗m ) is the profit of SP m, evaluated at the
envelope, ûi (zi ), has a closed form given by solution of P̂ and that K is the number of the NPs.
ui (0) + ui (w)−ui (0)

∗
ûi (zi ) = w zi 0≤zi ≤w
, Theorem 1. Let (rm ∗
, x∗m ) and (r̄m , x̄∗m ) be solutions of
ui (zi ) w≤zi Problems P and P̄ respectively. Moreover, let F̂ :=
for some w > ki which can be found easily by bisection {ui }i∈Um . Then,
[27]. Fig. 1 depicts the concavification of the aforementioned ∗ ∗
ψm − ǫ − δ1 (λm ) ≤ Ψm (r̂m , x̂∗m ) ≤ ψm
∗
+ δ2 (λm ),
sigmoid functions for kc(·) = 100 and three different values
for tzc(·) . Note that for the lowest tzc(·) (elastic traffic) we get the where δ1 (λm ) := λm (kx∗m P k22 − kx̂∗m k22 ), δ2 (λm ) :=
K
best approximation whilst for the largest (inelastic traffic/tight λm (kx̂∗m k22 − kx̄∗m k22 ) and ǫ = j=1 ρ[j] (F̂ ).
QoS prerequisites) we get the worst. Proof:
To exploit the closed form of the envelope ûi (zi ), instead ∗
Note that ψ̄m is also given by solving P̃ and that (r̂m∗
, x̂∗m )
of problem P̄ , we will concavify the equivalent problem: ∗
with the corresponding optimal value ψ̂m , are given by solving
(P̃ ): max
X
fi (ri , zi ) − cT xm − λm kxm k22 , P̂ . Therefore, from [28, Th. 1], we have that
rm ,xm ,zm ,
i∈Um K
X
∗ ∗
s.t. (ri , zi ) ∈ Si , ∀i ∈ Um ψ̄m − ρ[j] (F̂ ) ≤ um (r̂m ) − cT x̂∗m − λm kx̂∗m k22 ≤ ψ̄m
∗
X j=1
xm  ri
i∈Um The result follows from Lemma 1.
where Si := {(ri , zi ) : ri  0, zi = } and βiT ri Remark 1. The values of δ1 and δ2 decrease as λm decreases
fi (ri , zi ) := ui (zi ) with domain Si . The following lemma and hence for small regularization penalties they can get
uses the concave envelope of the sigmoid function ui (zi ), arbitrarily close to zero.
to compute the concave envelope of fi (ri , zi ) and hence the
Remark 2. The approximation error, ǫ, depends on the K
concavification of the problem P̃ . Its proof is based on the
greatest nonconcavities of the set {ui }i∈Um . There are two
definition of the concave envelope and is omitted for brevity.
conditions that ensure negligible approximation error, i.e.,
∗
Lemma 2. The concave envelope of the function fi (ri , zi ) := ǫ << ψm : i) the end-users have concave utility functions
tz (z −k )
e c(i) i c(i) ˆ (in that case ǫ → 0) or, ii) the market is profitable enough
tz (z −k ) with domain Si , fi (ri , zi ), has the following
i ∗
1+e c(i) c(i)
for every SP m and hence ψm >> K. Condition ii) makes
closed form (with domain Si ):
the error negligible since ǫ ≤ K, and it can be satisfied for
fˆi (ri , zi ) = ûi (zi ), ∀(ri , zi ) ∈ Si . example when the supply of the market, C, is sufficiently large.
 1.0 1.0 1.0 of P and P̂ respectively (given c). Since Problem P may
admit multiple solutions, let the set Dm (c) be defined as
0.8 0.8 0.8  
∗ ∗ ∗ ∗ ∗
utility

utility

utility
0.6 0.6 0.6 Dm (c) := xm : {∃rm : {Ψm (rm , xm ) = ψm given c} .
0.4 0.4
0.4 We define a Competitive equilibrium as follows:
Sigmoid Utility 0.2 Sigmoid Utility 0.2 Sigmoid Utility
0.2 Definition 2 (Competitive equilibrium). Competitive
Concave Envelope 0.0 Concave Envelope 0.0 Concave Envelope equilibrium of the Network Slicing Market is defined to be
0 50 100 150 200 250 300 0 50 100 150 200 250 300 0 50 100 150 200 250 300 any price vector c† and allocation of the resources of the
zi zi zi
NPs x† , such that:
(a) (b) (c)
Fig. 1: Concave Envelopes of sigmoid utility functions with kc(·) = i. x†m ∈ Dm (c† ) for every SP m, and
100 and (a) tzc(·) = 0.02, (b) tzc(·) = 0.2 and (c) tzc(·) = 2. ii. C = m∈M x†m (the demand equals the supply).
P

Note that in a competitive equilibrium, every SP m gets

Theorem 1, implies that every SP can solve Problem P̂ , resources that could maximize its profit given the price vector.
which is a concave program with a unique solution, to find Because a competitive equilibrium sets a balance between the
an approximate solution to P . This observation fosters the interests of all participants, it appears to be the settling point
convergence analysis of the proposed auction in Section III-A. of the markets in economic analysis [26], [29]. Nevertheless,
since the SPs’ demands are expressed by solving a non-
III. N ETWORK S LICING M ARKET C YCLE concave program, we define an ǫ-competitive equilibrium
In this section, we study the evolution of the network slicing which will be the ultimate goal of the proposed clock auction.
market using an iterative model that consists of 5-step cycles.
Definition 3 (ǫ-Competitive equilibrium). ǫ-Competitive
We refer to the following sequence of steps as a market cycle:
equilibrium of the Network Slicing Market is defined to be
S1. |Um | prospective users appear to every SP m. any price vector ĉ† and allocation of the resources of the NPs
S2. The vector xm , i.e., the distribution of the resources from x̂† , such that:
the NPs to SP m is determined for every m. To achieve
that in a distributed fashion, an auction between the SPs i. For every SP m, there exists an ǫ ≥ 0 and a feasible intra-
†
and the NPs should be realized. slice resource allocation vector r̂m (given x̂†m ), such that:
S3. Given xm , each SP m determines the vectors ri and ∗
ψm †
− ǫ ≤ Ψm (r̂m ∗
, x̂†m ) ≤ ψm + ǫ, and
hence the amount of resources zi for every user i ∈ Um
(intra-slice resource allocation). x̂†m (the demand equals the supply).
P
ii. C = m∈M
S4. After receiving the resources, each user i determines and
reports to the SP whether the QoS received was enough Observe that the first condition of the above definition
or not to complete its application. ensures that every SP is satisfied (up to a constant) with
S5. The SPs exploit the responses of their users, to estimate the obtained resources in a sense that it operates close to its
their private parameters and hence to distribute the maximum possible profit. From Theorem P1, note that if there
resources more efficiently in the next cycle. exists a price vector ĉ† such that C = m∈M x̂∗m (ĉ† ), then
It is important for the vector xm to be determined before the prices in ĉ† with the allocation x̂† := x̂∗ (ĉ† ) form an
the intra-slice resource allocation, since the first serves as the ǫ-competitive equilibrium. Finding such a price vector, is the
capacity in the resources available to SP m. In the following, motivation of the proposed clock auction. For the rest of the
we expand upon each (non-trivial) step of the market cycle. paper we make the following assumption:

A. Step S2 - Clock Auction for the Network Slicing Market Assumption 1. The SPs calculate their demand and intra-
resource allocation by solving Problem P̂ .
In this section, we develop and analyze a clock auction
between the SPs and the NPs, that converges to a market’s This is a reasonable assumption since in Theorem 1 and the
equilibrium. Specifically, we describe the goal (Section corresponding Remarks 1 and 2, we proved that by solving
III-A1), the steps (Section III-A2), and the convergence a (strictly) concave problem, every SP can operate near its
(Section III-A3) of the auction. optimal profit. Therefore, for the rest of the paper, we call
1) Auction Goal: Note that the solutions of the problems x̂∗m (c), the demand of SP m given the prices c.
P and P̂ appear to be a function of the prices c1 , . . . , cK . Let 2) Auction Description: We propose the following clock
the demand of SP m, given the price vector c, be denoted as auction that converges to an ǫ-competitive equilibrium of the
x∗m (c) or x̂∗m (c) depending on whether SP m uses Problem Network Slicing market (Theorem 2). As we will prove in
P or P̂ to ask for resources. Let also r ∗m (c) and r̂m ∗
(c) Theorem 3, this equilibrium is robust since the convergent
be optimal intra-slice resource allocation vectors respectively. price vector is the unique one that clears the market, i.e., makes
Hence, (r ∗m (c), x∗m (c)) and (r̂ ∗m (c), x̂∗m (c)) are maximizers the demand to equal the supply.
 i. An auctioneer announces a price vector c, each The above holds true since the function h(xm ) := Um (xm ) −
component of which corresponds to the price that an NP λm kxm k22 , has as concave conjugate the function (see [31])
sells a unit of its resources.
ii. The bidders (SPs) report their demands. h∗ (s) = max {h(xm ) − cT xm },
xm 0
iii. If the aggregated demand received by an NP is greater
than its available supply, the price of that NP is increased and hence ∇h (s) = arg maxxm 0 {Um (xm ) − λm kxm k22 −
∗

and vice versa. In other words, the auctioneer adjusts the cT xm }. Therefore, V(·) is a decreasing function of time and
price vector according to Walrasian tatonnement. converges to its minimum. Note that in the convergent point
iv. The process repeats until the price vector converges. the supply equals the demand for every NP.
Note that the components of the price vector change The market might admit multiple ǫ-competitive equilibria.
simultaneously and independently. Hence different brokers Nevertheless, the equilibrium point that the clock auction
can cooperate to jointly clear the market efficiently in a converges is robust in the following sense: given Assumption
decentralized fashion [23]. Let the excess demand, Z(c), be 1, the price vector that clears the market is unique. Therefore,
the difference P between the aggregate demand and supply: essentially, in Theorem 2 we proved that the proposed clock
Z(c) = −C + m∈M x̂∗m (c). In Walrasian tatonnement, the auction converges to that unique price vector. This is formally
price vector adjusts in continuous time according to excess proposed by the following theorem.
demand as ċ = f (Z(c(t))), where f is a continuous, sign-
Theorem 3. There exists a unique price vector c† such that
preserving transformation [24]. For the rest of the paper, we P ∗ †
set f to be the identity function and thus ċ = Z(c(t)). In m∈M x̂m (c ) = C.

auctions based on Walrasian tatonnement, the payments are Towards proving Theorem 3 we provide Lemmata 4 and 5.
only valid after the convergence of the mechanism [30]. First, we show that if a component in the price vector changes,
3) Auction Convergence: Towards proving the convergence the demand of an SP who used to obtain resources from the
of the auction, we provide the lemma below which proves that corresponding NP, should change as well.
the concavified version of the intra-slice resource allocation
Lemma 4. For two distinct price vectors c, c̄ with ∃k : ck 6=
problem IN − SL, can be thought of as a concave function.
c̄k , it holds true that
The proof is ommitted as a direct extension of [3] and [31].
Lemma 3. The function Um (xm ) shown below is concave. x̂∗m (c) = x̂∗m (c̄) ⇒ x̂∗(m,k) (c) = x̂∗(m,k) (c̄) = 0.
Proof: Let such price vectors, c̄ and c, with ck 6= c̄k .
fˆi (ri , zi )
X
Um (xm ) := max
rm ,zm Since x̂∗m (c) is the optimal point of problem P̂ given c,
i∈Um applying KKT will give:
s.t. (ri , zi ) ∈ Si , ∀i ∈ Um (4)
∂{Um (xm ) − λm kxm k22 }
x̂∗(m,k) (c) = 0 or = ck . (5)
X
xm  ri ∂x(m,k) x̂∗
m (c)
i∈Um
However, x̂∗m (c̄) is optimal for P̂ given c̄. Employing a similar
Using the function Um , we can rewrite Problem P̂ as equation as (5) proves that if x̂∗m (c) = x̂∗m (c̄) then it can only
max Um (xm ) − λm − cT xm kxm k22 . hold that x̂∗(m,k) (c) = x̂∗(m,k) (c̄) = 0.
xm 0

The following theorem studies the convergence of the auction. Definition 4 (WARP property). The aggregate demand
Theorem 2. Starting from any price vector cinit , the proposed function satisfies the Weak Axiom of Revealed Preferences
(WARP), if for different price vectors c and c̄, it holds that:
clock auction converges to an ǫ-competitive equilibrium. X X
Proof: The proof relies on a global stability argument cT · x̂∗m (c̄) ≤ cT · x̂∗m (c) ⇒
similarly to [24], [29]. Let Vm (·) denote m’s net indirect m∈M m∈M

utility function: X X
c̄T · x̂∗m (c̄) < c̄T · x̂∗m (c)
Vm (c) = max {Um (xm ) − λm kxm k22 T
− c xm }. m∈M m∈M
xm 0
Lemma 5. The aggregate demand function Psatisfies ∗the WARP
T
Let a candidate Lyapunov function be V(c) := c C + for distinct price vectors c, c̄ such that m∈M x̂m (c) ≻ 0
∗
P P
m∈M m V (c). To study the convergence of the auction we and m∈M m x̂ (c̄) ≻ 0.
should find the time derivative of the above Lyapunov function: Proof: SinceP c 6= c̄ then
∗
∃k ∈ K : ck 6= c̄k . Furthermore,
o we have that m∈M x̂m (c) ≻ 0 and hence ∃m1 ∈ M
P
 n
∗
V̇(c)=ċ· C T + m∈M dc d
maxxm 0 {Um (xm )−λm kxm k22 −cT xm } . such that x̂m1 ,k (c) > 0. Using Lemma 4 we conclude that
x̂∗m1 (c) 6= x̂∗m1 (c̄). Hence, since Problem P̂ admits a unique
Hence, we deduce that: global maximum we have that:
 X 
V̇(c) = C T + {−x̂∗T T X n o
m (c)} · ċ = −Z (c(t)) · Z(c(t)). Um (x̂∗m (c)) − λm kx̂∗m (c)k22 − cT · x̂∗m (c) >
m∈M m∈M
 X n o
Um (x̂∗m (c̄)) − λm kx̂∗m (c̄)k22 − cT · x̂∗m (c̄) C. Learning the Parameters (Step S5)
m∈M
At the final step of the cycle, the SPs exploit the data they
Now, the above combined with the WARP hypothesis, obtained to learn the private parameters of their users. In that
X X fashion, the market ”learns” its equilibrium. For the rest of
cT · x̂∗m (c̄) ≤ cT · x̂∗m (c),
the paper, for generality, we assume the pricing mechanism
m∈M m∈M
introduced in Section II-B1. Therefore, for every user i, the
gives: SPs get to know whether it is satisfied by the pair of resources-
X n o price (zi , pi ). A Bayesian inference model needs the data, a
Um (x̂∗m (c)) − λm kx̂∗m (c)k22 >
m∈M
model for the private parameters and a prior distribution.
X n o Model: The observed data is the outcome of the Bernoulli
Um (x̂∗m (c̄)) − λm kx̂∗m (c̄)k22 . (6) variables sati |θc(i) ∼ Bernoulli(P [sati ]) for every user
m∈M i, where θc(i) = (tpc(i) , bc(i) , tzc(i) , kc(i) ) is the tuple of
The result follows by switching the roles of c and c̄ and the private parameters that we want to infer. Prior: Let
combine the inequalities. the prior distribution for every parameter of θc(i) have
probability density functions πtpc(i) (·), πbc(i) (·), πtzc(i) (·) and
πkc(i) (·) respectively. The SPs infer the private parameters
We can now prove Theorem 3 as follows.
θc(i) for each service class using the Bayes rule separately:
proof of Theorem 3: Towards a contradiction, assume p(θc(i) |data) ∝ Ln (data|θc(i) )π(θc(i) ), where p(θc(i) |data)
that there exist
P two distinct (non-zero) price vectors c and c̄
is the posterior distribution of θc(i) , Ln (data|θc(i) ) is the
that satisfy m∈M x̂∗m (c̄) = m∈M x̂∗m (c) = C and thus
P
likelihood of the data given our model and π(θc(i) ) is the
prior distribution. Assuming independent private parameters,
 X X 
cT · x̂∗m (c̄) − x̂∗m (c) = 0. (7)
π(θc(i) ) is the product of the distinct prior distributions, and
m∈M m∈M
for each class c we have that:
Therefore, from Lemma 5 we know that: Y
X X Ln (data|θc(i) ) = P [sati ]fi (1 − P [sati ])1−fi ,
c̄T · x̂∗m (c̄) < c̄T · x̂∗m (c), (8) i∈Ccm
m∈M m∈M
where fi is 1 when user i is satisfied and 0 when not.
which is a contradiction because of the hypothesis. The SPs can use Marcov Chain Monte Carlo (MCMC) with
Metropolis Sampling, to find the posterior distribution after
Remark 3. Theorems 2 and 3 together with Remarks 1 and 2 each market cycle. As the market evolves, the SPs exploit the
imply that if the users’ traffic is elastic, or the total capacity previous posterior distributions to find better priors for the next
C of the NPs is sufficiently large, the clock auction converges cycle.
monotonically to the unique competitive equilibrium of the IV. C ENTRALIZED S OLUTION
market.
In case there exists a centralized entity that knows the utility
At the end of step S2, the final price vector ĉ† and the final function of every SP, it can optimize the social welfare, i.e.,
demands of each SP m, x̂∗m , have been determined. the summation of the utility functions of the service and the
network providers. This centralized problem can be formulated
B. Intra-Slice Resource Allocation & Feedback (Steps S3, S4)
as follows:
At the beginning of step S3, every SP m is aware of the
convergent point x̂∗m and hence it can allocate the resources
X
(SWM): max um (rm )
rm
either by solving the sigmoid program IN − SL, or by m∈M
∗
using the convergent approximate solution, r̂m . At that step, s.t. ri  0, ∀i ∈ Um
an SP can also determine whether it will overbook network X X
resources. Overbooking, is a common practice in airlines and ri  C
m∈M i∈Um
hotel industries and is now being used in the network slicing
problem [32], [33]. This management model allocates the The SW M problem, can be solved with any chosen positive
same resources to users of the network expecting that not approximation error, using the framework of sigmoidal
everyone uses their booked capacity. In that case, SP m solves programming [27].
Problem IN −SL whilst setting increased obtained resources,
∗ ∗ V. N UMERICAL R ESULTS
xov
m = x̂m + α% ◦ x̂m , for a relatively small positive α. Here,
◦ denotes the component-wise multiplication operator. A. Auction Convergence & Parameter Tuning
During the step S4 of the cycle, each user i, receives their In this section we study the convergence of the clock
resources ri , and provide feedback on whether it was satisfied auction, as well as the impact that the various parameters have
or not. In the next step, the SPs can use the these responses on its behavior. For this simulation, we assume a small market
to learn the private parameters of the different service classes. with 3 NPs with capacities C1 = 850, C2 = 750, C3 = 755
 Auction SPP oSPP(5%) SWM
1e6
L2 Norm of the Excess Demand

L2 Norm of the Excess Demand

7 17500 1400
cinit = [0.62, 0.64, 0.58] κ = 10^{-4}
6
1200
cinit = [1.2, 1.4, 1.1] 15000
κ = 10^{-5}
1000
5 cinit = [0.2, 0.4, 0.1] κ = 10^{-6}

xm, k
12500

cinit = [0.4, 0.4, 1.1]

800
4 10000
600
3 7500 400
2 5000 200
0
1 2500
SP1/NP1 SP1/NP2 SP2/NP1 SP2/NP2
0 0
Service Provider/Network Provider
0 2 4 6 8 10 12 14 0 10 20 30 40 50 Fig. 4: Total amount of resources obtained by every SP m from
Iterations Iterations every NP k in the market, x(m,k) .
(a) (b)
Fig. 2: L2 norm of the excess demand vector throughout the clock
auction (a) for κ = 10−4 and various initialization price vectors
cinit , and (b) for cTinit = [0.62, 0.64, 0.58] and different values of B. Visualization of the Resource Allocation
κ. In this section, we get insights on the allocation of
the resources in the market. We assume 2 NPs with
cinit = [0.62, 0.64, 0.58]
C1 = C2 = 1400 and 2 SPs with 10 users each and one shared
cinit = [1.2, 1.4, 1.1] service class with tzc(i) = tpc(i) = 0.2 and kc(i) = bc(i) = 100
cinit
cinit
= [0.2, 0.4, 0.1]

= [0.4, 0.4, 1.1]

for all i. The first SP (SP1) is near the first NP (NP1)
and far from NP2 and hence, we set [β(1,1) , . . . , β(1,10) ] =
Cost of NP3

1.0

0.9 [0.99, 0.96, 0.87, 0.85, 0.82, 0.81, 0.80, 0.80, 0.70, 0.70] and
0.8

0.7
β(2,i) = 0.2, ∀i ∈ U1 . Moreover, for the users of SP2 we set
0.6

0.5
β1,i = β2,i = 0.8, ∀i ∈ U2 .
0.4
We compare the resource allocation of f our different
methods. First, ’Auction’ refers to the resource allocation that
results immediately after the auction. ’SPP’ takes x̂∗m from the
1.2
2

1.0
NP

0.3
0.4 0.8
equilibrium but performs the intra-slice of every SP by solving
of

Cost
0.5
0.6
0.7 0.6

of NP
st

0.8
0.9
IN − SL. We also study the method ’oSPP(5%)’, which
Co

1 1.0 0.4
1.1

mimics the SPP method but with 5% overbooked resources.

Fig. 3: Illustrating Theorem 2. Starting from any price vector cinit , Finally, ’SWM’ refers to the solution of the Problem SW M .
the clock auction converges to the market clearing prices c† .
Fig. 4 shows the amount of resources obtained from the
two SPs. All methods allocate the majority of the resources
of NP1 to SP1 since its users have greater connectivity with
and 5 SPs with 6 users and 3 distinct service classes each. it. Although the users of SP2 have equally high connectivity
The users’ private parameters are set as follows: for an i in with both NPs, all of the four methods were flexible enough
the first class tzc(i) = tpc(i) = 0.2, kc(i) = bc(i) = 100, for the to allocate the resources of NP2 to SP2. Note that none of the
second class tzc(i) = tpc(i) = 2, kc(i) = bc(i) = 120, and for the methods gives resources from NP2 to SP1.
third class tzc(i) = tpc(i) = 20, kc(i) = bc(i) = 150. Such values Fig. 5 depicts the intra-slice resource allocations. In Fig.
indicate that the users wish to pay a unit of monetary value 5a observe that the greater the connectivity of a user is,
for a unit of offered resources. the less resources it gets. That is because users with good
To discretize the auction, we change the cost vector connectivity factors meet their prerequisite QoS using less
according to a step value, κ, as ct+1 = ct + κZ(ct ). Fig. 2 resources and hence SP1 could maximize its expected profit
depicts the L2 norm of the excess demand vector throughout by giving them less. Note that ’SPP’ gives no resources to the
the clock auction for different cost vector initializations cinit user with the worst connectivity whereas with the overbooking,
(Fig 2a), and for different step values κ (Fig. 2b). By SP1 gets enough resources to make attractive offers to every
simulating the clock auction, we deduce that the clearing price user. Therefore, ’SPP’ might make an unfair allocation, since
vector is c†T = [0.6116, 0.6273, 0.5811]. In Fig. 2a note that when the resources are not enough, it neglects the users with
the closer the initialization cost vector is to c†T , the faster bad connectivity. In Fig. 5c, note that the homogeneity in the
the convergence becomes. Fig. 2b, connotes the need for a connectivities of the users of SP2 forces every method to fairly
proper choice of the step value κ. Clearly, κ = 10−4 gives divide the resources among them.
the fastest convergence and as we decrease the step values Fig. 6a shows the expected value of the total revenue, or the
it becomes slower. Nevertheless, since Theorem 2 is proved social welfare. ’SWM’ gives the greatest revenue among the
for the continuous case, large values of κ cannot guarantee methods that do not overbook. Nevertheless, although ’SPP’
the convergence of the auction to an equilibrium. In Fig. 3 is a completely distributed solution and was not designed to
observe that the convergence of the auction does not depend maximize the total revenue, it performs very close to ’SWM’.
on the initialization of the cost vector (Theorem 2). Moreover, a 5% overbooking leads to greater revenues.
 Auction SPP oSPP(5%) SWM Auction SPP oSPP(5%) SWM Auction SPP oSPP(5%) SWM
160
175
50 140
150
120
125 40
100
100 30
r1, i

r1, i

r2, i
80
75 60
20
50 40
25 10 20
0 0 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
User ID of SP1 User ID of SP2 User ID of SP2
(a) (b) (c)
Fig. 5: The solution of the intra-slice resource allocation problem from the perspective of the two different SPs of the market. Specifically,
how (a) SP1 distributed the resources of NP1, i.e., r1,i for every i in U1 , (b) SP2 distributed the resources of NP1, i.e., r1,i for every i in
U2 , and (c) SP2 distributed the resources of NP2, i.e., r2,i for every i in U2 .
1677.81 827.42 806.49 828.46 828.51 850.39
1575.31 1598.16 1611.54 800 805.06
1600 746.85 769.65 800

Expected Revenue of SP1

Expected Revenue of SP2

700
Expected Total Revenue

1400 700
1200 600 600
1000 500 500
800 400 400
600 300 300
400 200 200
200 100 100
0 Auction SPP oSPP(5%) SWM 0 Auction SPP oSPP(5%) SWM 0 Auction SPP οSPP(5%) SWM
Resource Allocation Method Resource Allocation Method Resource Allocation Method
(a) (b) (c)
Fig. 6: Illustrating the expected revenue (given by Eq. (3)) for the four different resource allocation methods. Fig. (a) shows the aggregated
expected revenue, Fig. (b) shows the expected revenue of SP1, and Fig. (c) shows the expected revenue of SP2.

C. Impact of Bayesian Inference

0.5
The previous results are extracted after a sufficient number
0.4

of cycles, when the SPs have learned the parameters of the

0.4
0.3

end-users. In this section, we consider an SP with 10 users

0.3
and one service class that employs Bayesian inference to learn
0.2

0.2
the private parameter tzc(i) for every i. We set the true value
of the parameter to be tzc(i) = 2. The other parameters are
0.1

set tpc(i) = 2, kc(i) = bc(i) = 120 and β1,i = 0.9, ∀i ∈ U1 . 0.1

0.0

0.0

We assume one more SP with a unique service class with 0 2 4

z
tc(i)
6 8 0 1 2 3
z
tc(i)
4 5 6

tpc(i) = tzc(i) = 0.2, kc(i) = bc(i) = 100 and β2,i = 0.9∀i ∈ U2 . (a) (b)
Finally, there are 2 NPs with C1 = C2 = 1200. Fig. 7: Posterior distribution of the unknown private parameter tzc(i)
In this example, SP1 sets as prior distribution the normal in (a) the first Market Cycle, and (b) in the second Market Cycle.
N (0.02, 2) and hence assumes elastic traffic. At the end
of each market cycle, the SP makes an estimation, t̂zc(i) , Cycle t̂zc(i) Acquired Perceived Actual
by calculating the mean of the posterior distribution. Fig. 7 Resources Revenue Revenue
depicts the histogram of the posterior distribution for the first
two market cycles. Observe that even in the third market cycle, 1 0.02 1087 530.26 699
2 1.68 1370 1160.77 1163.48
SP1 can estimate with high accuracy the actual value of the 3 2.01 1365 1161.42 1161.42
parameter. In Table I, note that the perceived revenue, i.e., the
expected revenue calculated using the estimation, is different TABLE I: Bayesian inference in different market cycles.
between the cycles that t̂zc(i) differs from tzc(i) . Hence, it is
impossible for the SPs to maximize their expected profits
when they don’t know the actual values of the parameters. consideration the heterogenity of the users’ service classes
Indeed, observe that the bad estimate of t̂zc(i) = 0.02 gives we introduce an iterative market model along with a clock
poor expected revenue compared to the last two cycles. auction that converges to a robust ǫ-competitive equilibrium.
Finally, we propose a Bayesian inference model, for the SPs
VI. C ONCLUDING R EMARKS to learn the private parameters of their users and make the
next equilibria more efficient. Numerical results validate the
In this paper we focus on the technical and economic
convergence of the clock auction and the capability of the
challenges that emerge from the application of the network
proposed framework to capture the different incentives.
slicing architecture to real world scenarios. Taking into
 R EFERENCES [20] B. Cao, W. Lang, Y. Li, Z. Chen, and H. Wang, “Power allocation in
wireless network virtualization with buyer/seller and auction game,” in
[1] Q. Zhang, F. Liu, and C. Zeng, “Adaptive interference-aware 2015 IEEE Global Communications Conference (GLOBECOM). IEEE,
vnf placement for service-customized 5g network slices,” in IEEE 2015, pp. 1–6.
INFOCOM 2019-IEEE Conference on Computer Communications. [21] K. Zhu and E. Hossain, “Virtualization of 5g cellular networks as
IEEE, 2019, pp. 2449–2457. a hierarchical combinatorial auction,” IEEE Transactions on Mobile
[2] R. Su, D. Zhang, R. Venkatesan, Z. Gong, C. Li, F. Ding, Computing, vol. 15, no. 10, pp. 2640–2654, 2015.
F. Jiang, and Z. Zhu, “Resource allocation for network slicing in 5g [22] K. Zhu, Z. Cheng, B. Chen, and R. Wang, “Wireless virtualization as
telecommunication networks: A survey of principles and models,” IEEE a hierarchical combinatorial auction: An illustrative example,” in 2017
Network, vol. 33, no. 6, pp. 172–179, 2019. IEEE Wireless Communications and Networking Conference (WCNC).
[3] Q. Qin, N. Choi, M. R. Rahman, M. Thottan, and L. Tassiulas, “Network IEEE, 2017, pp. 1–6.
slicing in heterogeneous software-defined rans,” in IEEE INFOCOM [23] G. Iosifidis, L. Gao, J. Huang, and L. Tassiulas, “A double-
2020-IEEE Conference on Computer Communications. IEEE, 2020, auction mechanism for mobile data-offloading markets,” IEEE/ACM
pp. 2371–2380. Transactions on Networking, vol. 23, no. 5, pp. 1634–1647, 2014.
[24] L. M. Ausubel and P. Cramton, “Auctioning many divisible goods,”
[4] Q.-V. Pham and W.-J. Hwang, “Network utility maximization-based
Journal of the European Economic Association, vol. 2, no. 2-3, pp.
congestion control over wireless networks: A survey and potential
480–493, 2004.
directives,” IEEE Communications Surveys & Tutorials, vol. 19, no. 2,
[25] L. M. Ausubel, P. Cramton, and P. Milgrom, “The clock-proxy auction:
pp. 1173–1200, 2016.
A practical combinatorial auction design,” Handbook of Spectrum
[5] J.-W. Lee, R. R. Mazumdar, and N. B. Shroff, “Non-convex optimization
Auction Design, pp. 120–140, 2006.
and rate control for multi-class services in the internet,” IEEE/ACM
[26] S. SHEN, “First fundamental theorem of welfare economics,” 2018.
transactions on networking, vol. 13, no. 4, pp. 827–840, 2005.
[27] M. Udell and S. Boyd, “Maximizing a sum of sigmoids,” Optimization
[6] J. Liu, “A theoretical framework for solving the optimal admissions and Engineering, pp. 1–25, 2013.
control with sigmoidal utility functions,” in 2013 International [28] ——, “Bounding duality gap for separable problems with linear
Conference on Computing, Networking and Communications (ICNC). constraints,” Computational Optimization and Applications, vol. 64,
IEEE, 2013, pp. 237–242. no. 2, pp. 355–378, 2016.
[7] A. Lieto, I. Malanchini, S. Mandelli, E. Moro, and A. Capone, [29] M. Bichler, M. Fichtl, and G. Schwarz, “Walrasian equilibria from an
“Strategic network slicing management in radio access networks,” IEEE optimization perspective: A guide to the literature,” Naval Research
Transactions on Mobile Computing, 2020. Logistics (NRL), vol. 68, no. 4, pp. 496–513, 2021.
[8] Q. Zhu and R. Boutaba, “Nonlinear quadratic pricing for concavifiable [30] C. Courcoubetis and R. Weber, Pricing communication networks:
utilities in network rate control,” in IEEE GLOBECOM 2008-2008 IEEE economics, technology and modelling. John Wiley & Sons, 2003.
Global Telecommunications Conference. IEEE, 2008, pp. 1–6. [31] S. Boyd, S. P. Boyd, and L. Vandenberghe, Convex optimization.
[9] L. Gao, P. Li, Z. Pan, N. Liu, and X. You, “Virtualization framework Cambridge university press, 2004.
and vcg based resource block allocation scheme for lte virtualization,” in [32] J. X. Salvat, L. Zanzi, A. Garcia-Saavedra, V. Sciancalepore, and
2016 IEEE 83rd Vehicular Technology Conference (VTC Spring). IEEE, X. Costa-Perez, “Overbooking network slices through yield-driven
2016, pp. 1–6. end-to-end orchestration,” in Proceedings of the 14th International
[10] G. Lee, H. Kim, Y. Cho, and S.-H. Lee, “Qoe-aware scheduling for Conference on emerging Networking EXperiments and Technologies,
sigmoid optimization in wireless networks,” IEEE Communications 2018, pp. 353–365.
Letters, vol. 18, no. 11, pp. 1995–1998, 2014. [33] C. Marquez, M. Gramaglia, M. Fiore, A. Banchs, and X. Costa-Perez,
[11] M. Hemmati, B. McCormick, and S. Shirmohammadi, “Qoe-aware “Resource sharing efficiency in network slicing,” IEEE Transactions on
bandwidth allocation for video traffic using sigmoidal programming,” Network and Service Management, vol. 16, no. 3, pp. 909–923, 2019.
IEEE MultiMedia, vol. 24, no. 4, pp. 80–90, 2017.
[12] L. Tan, Z. Zhu, F. Ge, and N. Xiong, “Utility maximization resource
allocation in wireless networks: Methods and algorithms,” IEEE
Transactions on systems, man, and cybernetics: systems, vol. 45, no. 7,
pp. 1018–1034, 2015.
[13] S. Papavassiliou, E. E. Tsiropoulou, P. Promponas, and P. Vamvakas,
“A paradigm shift toward satisfaction, realism and efficiency in wireless
networks resource sharing,” IEEE Network, vol. 35, no. 1, pp. 348–355,
2020.
[14] I. Afolabi, T. Taleb, K. Samdanis, A. Ksentini, and H. Flinck,
“Network slicing and softwarization: A survey on principles, enabling
technologies, and solutions,” IEEE Communications Surveys &
Tutorials, vol. 20, no. 3, pp. 2429–2453, 2018.
[15] S. Vassilaras, L. Gkatzikis, N. Liakopoulos, I. N. Stiakogiannakis, M. Qi,
L. Shi, L. Liu, M. Debbah, and G. S. Paschos, “The algorithmic aspects
of network slicing,” IEEE Communications Magazine, vol. 55, no. 8,
pp. 112–119, 2017.
[16] U. Habiba and E. Hossain, “Auction mechanisms for virtualization
in 5g cellular networks: basics, trends, and open challenges,” IEEE
Communications Surveys & Tutorials, vol. 20, no. 3, pp. 2264–2293,
2018.
[17] D. Zhang, Z. Chang, F. R. Yu, X. Chen, and T. Hämäläinen, “A double
auction mechanism for virtual resource allocation in sdn-based cellular
network,” in 2016 IEEE 27th Annual International Symposium on
Personal, Indoor, and Mobile Radio Communications (PIMRC). IEEE,
2016, pp. 1–6.
[18] F. Fu and U. C. Kozat, “Wireless network virtualization as a sequential
auction game,” in 2010 Proceedings IEEE INFOCOM. IEEE, 2010,
pp. 1–9.
[19] H. Ahmadi, I. Macaluso, I. Gomez, L. DaSilva, and L. Doyle,
“Virtualization of spatial streams for enhanced spectrum sharing,” in
2016 IEEE Global Communications Conference (GLOBECOM). IEEE,
2016, pp. 1–6.