Network Slicing For Guaranteed Rate Services: Admission Control and Resource Allocation Games

This article has been accepted for publication in a future issue of this journal, but has not been
fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2018.2859918, IEEE
Transactions on Wireless Communications
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. XX, NO. X, XXXXXXX XXXX 1
Network Slicing for Guaranteed Rate Services:

Admission Control and Resource Allocation Games
P. Caballero, A. Banchs, Senior Member, IEEE, G. de Veciana, Fellow, IEEE, X. Costa-Pérez, Member, IEEE and
A. Azcorra, Senior Member, IEEE
Abstract—Technologies to enable network slicing are expected to be “sliced” into logical networks, which may be customized
to be a key component of next generation mobile networks. to support one or more specific services. This provides a
Their promise lies in enabling tenants (such as mobile operators basis for efficient infrastructure sharing among diverse entities,
and/or services) to reap the cost and performance benefits of
sharing resources while retaining the ability to customize their so-called tenants, each owning a slice. Tenants could be
own allocations. When employing dynamic sharing mechanisms, traditional or virtual mobile network operators acquiring a
tenants may exhibit strategic behavior, optimizing their choices in network slice from an infrastructure operator to support their
response to those of other tenants. This paper analyzes dynamic business, as well as new players that simply view connectivity
sharing in network slicing when tenants support inelastic users as a service, such as Over-The-Top (OTT) service providers
with minimum rate requirements. We propose a NEtwork Slicing
(NES) framework combining (i) admission control, (ii) resource which provision network slices to ensure quality of service to
allocation and (iii) user dropping. We model the network slicing their end-customers.
system with admitted users as a network slicing game; this is A major element underlying network slicing is a mechanism
a new class of game where the inelastic nature of the traffic for resource allocation amongst slices. One of the approaches
may lead to dropping users whose requirements cannot be
met. We show that, as long as admission control guarantees
considered in 3GPP suggests that base station resources could
that slices can satisfy the rate requirements of all their users, be statically partitioned based on fixed ‘network shares’ [3].
this game possesses a Nash Equilibrium. Admission control However, given that slices’ loads may be non-uniform accross
policies (a conservative and an aggressive one) are considered, space and varying in time, sharing gains can be achieved
along with a resource allocation scheme and a user dropping by dynamically allocating resources to slices based on their
algorithm, geared at maintaining the system in Nash Equilibria.
We analyze our NES framework’s performance in equilibrium,
current needs (while respecting their overall network shares).
showing that it achieves the same or better utility than static At the same time, tenants should retain the ability to operate
resource partitioning, and bound the difference between NES and their slices autonomously and, in particular, to customize the
the socially optimal performance. Simulation results confirm the allocation of resources to their users. This suggests the need
effectiveness of the proposed approach. for a flexible framework for resource sharing, wherein (i)
Index Terms—Wireless networks, Network slicing, Multi- tenants indicate their preferences to the infrastructure (e.g.,
tenant networks, Resource allocation, Guaranteed rate services, by dynamically subdividing their network share amongst their
Inelastic Traffic. users), and (ii) base station resources are allocated to slices
according to such preferences (e.g., proportionally to the
I. I NTRODUCTION shares assigned to the users).
It is widely agreed among the relevant industrial commu- Under such a resource allocation model, it is to be expected
nity [1] and ongoing standardization efforts [2] that enabling that tenants might exhibit strategic behavior, adjusting their
network slicing is a key technological requirement for 5G mo- preferences to current demands at the different base stations
bile networks. Such technology enables wireless infrastructure so as to maximize their performance (subject to their share of
the network). This could potentially have adverse effects on the
Manuscript received January 18, 2018; revised May 15, 2018; accepted July network; e.g., the overall network efficiency might be harmed,
8, 2018. Date of publication xxxx xx, xxxx; date of current version xxxx xx,
xxxx. The work of University of Texas at Austin was supported in part by a gift or tenants’ preferences (and the corresponding requests) might
from Cisco. The work of University Carlos III of Madrid was supported by the exhibit oscillations. While this problem has been studied in [4]
H2020 5G-MoNArch project (Grant Agreement No. 761445) and the 5GCity for the case of elastic users, in many cases tenants’ traffic
project of the Spanish Ministry of Economy and Competitiveness (TEC2016-
76795-C6-3-R). The work of NEC Europe Ltd. was supported by the H2020 will be inelastic in nature, wherein a user must either be
5G-Transformer project (Grant agreement no. 761536). The associate editor guaranteed a minimum rate or her utility decreases sharply.
coordinating the review of this paper and approving it for publication was When attempting to satisfy such user requirements, tenants’
X. Wang. (Corresponding author: Pablo Caballero.)
P. Caballero and G. de Veciana are with The University of Texas at behavior may differ substantially from that in [4], affecting
Austin, Austin, TX 78712 USA (e-mail: pablo.caballero@utexas.edu; gus- both network efficiency and stability. The focus of this paper is
tavo@utexas.edu). thus on the analysis of resource allocation for network slicing
A. Banchs and A. Azcorra are with the University Carlos III of Madrid,
28911 Leganés, Spain, and also with the IMDEA Networks Institute, 28918 when tenants support inelastic users.
Leganés, Spain (e-mail: banchs@it.uc3m.es; azcorra@it.uc3m.es).
X. Costa-Pérez is with NEC Europe Ltd., 69115 Heidelberg, Germany (e- Related work: The resource allocation mechanism analyzed
mail: xavier.costa@neclab.eu). in this paper corresponds to a Fisher market, which is a
0000–0000/00$00.00
c 2018 IEEE
1536-1276 (c) 2018 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 publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2018.2859918, IEEE
2 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. XX, NO. X, XXXXXXX XXXX
standard framework in economics. In such markets, buyers (in concave utility functions; among the few works on this topic it
our case slices) have fixed budgets (in our case corresponding is worth mentioning [21], which uses potential games to prove
to pre-agreed network shares) and bid for resources within convergence of Best Response Dynamics to a region around
their budget (according to their preferences), which are then the NE for finite strategy games [22].
allocated to buyers proportionally to their bids [5]. Within the In the specific context of Fisher market-like frameworks,
Fisher Market framework, our model falls in the category of to the best of our knowledge our work is the first attempt to
buyers that anticipate the impact of their bids [6]. The analysis analyze resource allocation for inelastic traffic. In particular,
of Fisher markets under such price-anticipating buyers has this work addresses the following gap in the literature of
been limited, so far, to the case of buyers with linear [6] or resource allocation models: the analysis of budget-constrained
concave [4], [7] utility functions. resource allocation under price-anticipating users with inelas-
A related resource allocation model often considered in the tic utilities. The nature of inelastic utility functions leads to
networking field is the so-called ‘Kelly’s mechanism’, which a new class of non-cooperative games, where a slice prefers
allocates resources to players proportionally to their bids to drop users whose rate requirements cannot be met, rather
[8]. This model has also been analyzed for price-anticipating than allocating them insufficient resources. The nature of such
players [9]. However, in Kelly’s mechanism players respond games differs substantially from the ones previously analyzed
to their payoff (given by the utility minus cost) whereas in our in the literature for elastic traffic.
model tenants’ behavior is only driven by their utilities (since On the 5G standardization front, network slicing is currently
they have a fixed budget, i.e., the network share). Moreover, being specified by 3GPP [2]. In particular, 3GPP’s SA5 is
Kelly’s model has mainly been studied for concave utility working on the definition of a management and orchestration
functions. framework to support network slicing [23], [24]. While these
The topic of network slicing is currently attracting substan- efforts do not specifically address dynamic resource allocation,
tial attention from the research community. One of the main which is our focus here, the algorithms we propose are in line
issues investigated is the resource allocation across different with this framework. One of the key features of our approach
slices, which is the focus of this paper. A number of works is the ability of tenants to customize their allocations; there
have been devoted to the resource allocation among different is wide consensus in the standardization community that this
operators or tenants sharing the same wireless infrastructure is needed to efficiently satisfy their very diverse requirements
(see e.g. [10]–[12]), and in [13], the authors focus on resource (see, e.g., [25] for examples of possible vertical tenants).
allocation of processing resources in network slicing in the Key contributions: The rest of the paper is organized as
context of C-RAN; see [14] for a survey on resource slicing follows. In Section II we present our system model, and
in virtual wireless networks. In contrast to our paper, all these propose the Network Slicing (NES) framework to address
works have focused on elastic traffic. resource allocation in such system. NES consists of three mod-
In the context of network slicing, there are some works ules: admission control, weight allocation and user dropping.
which have considered inelastic traffic. The algorithm pro- Section III focuses on the admission control module: it finds
posed in [15] attempts to satisfy the demands of all slices the requirements to ensure stability and proposes two policies,
but does not account for the resources each slice is entitled to. a conservative and an aggressive one, to perform admission
Similarly, [16]–[18] propose algorithms to meet requests from control. Section IV presents the other two modules: a resource
all tenants, but do not account for elastic demands and do not allocation mechanism and a strategy to drop users when
consider budget constraints. In [19], the authors propose an rate guarantees are infeasible, and analyzes the convergence
algorithm to trade resources among tenants, but their approach of the resulting dynamics. We then study in Section V the
involves complex negotiations and relies on heuristic consid- performance of NES versus two benchmark allocations: static
erations rather than a well-established analytical framework. resource partitioning and the social optimal. Throughout the
In contrast to all these works, our approach supports both paper, we present analytical results that support the design of
elastic and inelastic services and is based on fixed budgets, NES, including (i) the existence of a Nash Equilibrium and the
corresponding to the network shares; this is in line with one convergence of Best Response Dynamics, (ii) the effectiveness
of the scenarios considered in 3GPP [3] and does not involve of admission control and protection from other slices, (iii)
pricing individual requests, which may represent an advantage the user selection and weight allocation choices, and (iv)
in practical deployments. the gains over static slices and loss over social optimal. We
In this work, we build on the Fisher Market mechanism further evaluate the performance of NES via simulation in
for resource allocation across slices and analyze the game Section VI, confirming that it provides substantial gains in
resulting from the interaction of several non-cooperative slices terms of utility, throughput performance and reduced blocking
aiming to maximize their own network utility given a fixed probability while incurring an acceptable complexity.
budget. This problem has been addressed in the context of
concave utility functions: [7] ensures the existence of Nash
Equilibria (NE) for this type of utility functions, [20] proves II. N ETWORK S LICING MODEL
the existence of a NE for price-taking players, [4] shows the We consider a wireless network consisting of a set of
convergence of Best Response Dynamics for certain classes resources B (the base stations or sectors) shared by a set of
of concave functions and [6] shows they may not converge network slices O (each operated by a different tenant). At a
for linear utilities. Much less attention has been paid to non- given point in time, the network supports a set of active users
CABALLERO et al.: NETWORK SLICING FOR GUARANTEED RATE SERVICES: ADMISSION CONTROL AND RESOURCE ALLOCATION GAMES 3
U (the customers or devices), which can be subdivided into management system defined by 3GPP [24], and hence should
subsets Ubo , Ub and U o , corresponding to the users of slice be readily available.
o at base station b, the users at base station b, and the users In order to avoid the indeterminate form resulting from
of slice o, respectively. We consider that the association of having all the weights at a base station equal to 0 in (1),
users with base stations is fixed (e.g., by a pre-specified user we will require weights to exceed a fixed lower bound (i.e.,
association policy) and let b(u) denote the base station that wu ≥ δ, ∀u). This bound can be arbitrarily small; indeed, in
user u is (currently) associated with. practice it should be set as small as possible, to allow slices
the highest possible flexibility while avoiding zero weights.
Accordingly, in the rest of the paper we assume that δ is so
A. Resource allocation model
small that its effect can be neglected, except for Theorem 2,
Following a similar approach as [4], [10], in our model where this assumption is required to prove the existence of a
each slice o is allocatedPa network share so (corresponding Nash Equilibrium.
to its budget) such that o∈O so = 1. The slice is at liberty In the case where a slice o is the only one with users at
to distribute its share amongst its users, assigning them non- a given base station b, such a slice would simply set wu to
negative weights (corresponding to the bids): the minimum possible value for these users, allowing them to
wu for u ∈ U o , such that
X
wu ≤ so . receive all the resources of this base station while minimizing
u∈U o
the consumed share. To avoid dealing with this special case,
hereafter we shall assume that all base stations have users
We let wo = (wu : u ∈ U o ) be the weights of slice o, w = from at least two slices. Note that this assumption is made
(wu : u ∈ U) those of all slices and w−o = (wu : u ∈ U \ to simplify the expressions and discussion, and does not limit
U o ) the weights of P all users excluding those of slice o. We the generality of our analysis and algorithm, which indeed
further let lb (w) = u∈Ub wu denote the load at base station supports base stations with all users from the same slice.
b, dob (wo ) =
P
u∈U o wu the aggregate weight of slice o at
b
b, and aob (w−o ) = u∈Ub \U o wu the aggregate weight of all
P
b
other slices (excluding o) at b. We shall allocate each user a B. Slice utility
fraction of the base station’s resources in proportion to her Network slices may support services and customers with
weight wu . different needs, or may wish to differentiate the service they
We let cu denote the achievable rate for user u, defined as provide from competing slices. To that end, we assume that
the product of (i) the average rate per resource unit achieved each slice has a private utility function, U o , that reflects the
by the user, and (ii) the total amount of resources available slice’s performance according to the preferences and needs of
at the base station. Note that this depends on the modulation its users. The slice utility consists of the sum of the individual
and coding scheme selected for the current radio conditions, utilities of its users, Uu , i.e.,
which accounts for noise as well as the interference from the X
neighboring base stations. Following similar analyses in the U o (w) = Uu (ru (w)).
u∈U o
literature [10], [26], [27], we shall assume that cu is fixed for
each user at a given time. For inelastic traffic, we assume each user u requires a
We further let ru denote the rate allocated to user u. Under guaranteed rate γu , hereafter referred to as the user’s minimum
our model, ru is given by cu times the fraction of the base rate requirement. Following standard practice, we shall model
station’s resources allocated to the user. Given that users are inelastic traffic utility functions as1
allocated a fraction of resources proportional to their weights,
we have that ru is a function of the weights w given by: Uu (ru (w)) = φu fu (ru (w)), for ru (w) ≥ γu ,
wu wu where fu (·) is a concave2 utility function associated with
ru (w) = P cu = cu . (1)
v∈Ub(u) w v lb(u) (w) the user, and P φu is the relative priority of user u (where
φu ≥ 0 and u∈U o φu = 1). The relative priorities reflect
When implementing the proposed resource allocation mech- the importance that users are given by the tenant of their slice;
anism, a slice may assign a non-zero weight to some users they drive, jointly with the load at the respective base stations,
while others may be dropped. To decide the setting of the the weights assigned to the users, which in turn determine the
users’ weights, we assume that each slice o is aware of the rate allocation.
aggregate weight of the other tenants at each base station, Note that the above utility function is only defined for
i.e., aob (w−o ). It is worth noting that for the mechanism under rates above the minimal requirement, as performance degrades
study we have that (i) a slice only sees the aggregate weight of drastically if this guarantee is not met. Note also that the
the other slices, and hence can learn very limited information above definition includes elastic traffic, which corresponds to
about the other slices; in particular, the weights of each tenant
are not disclosed, and (ii) the mechanism needs to store very 1 Inelastic traffic utility functions are typically modeled as a discontinuous
limited data; indeed, it is sufficient to keep the total load function [28] or a sigmoidal one [29]. In this paper we adopt the former
of each base station, as a tenant can obtain aob (w−o ) by model, which aims at providing users with a guaranteed rate, and thus is
aligned with the Guaranteed Bit Rate (GBR) class of 3GPP [30].
simply subtracting its weight from the base station’s load. Such 2 Note that, even when f (·) is concave, we are dealing with non-concave
u
information is already considered within the network slicing utilities, due to the minimum rate requirement.
the special case γu = 0; thus, the results of this paper apply b) Static Slicing Allocation (SS): By static slicing (also
to mixes of elastic and inelastic traffic. known as static splitting [32]) we refer to a complete parti-
While most of our results hold for arbitrary fu (·) functions, tioning of resources based on the network shares so , o ∈ O.
in some cases we will focus on the following widely accepted In this setting, each slice o receives a fixed fraction so of each
family of utility functions (see α-fairness, [31]): resource, which is shared among its users proportionally to
their weights,
(ru )1−αo
(
(1−αo ) , αo 6= 1 wu
fu (ru ) = (2) russ (wo ) = P so cu , ∀u ∈ U o , ∀o ∈ O, (3)
log(ru ), αo = 1, v∈U o wv
b(u)
where the αo parameter sets the level of concavity of the where we note that, in this case, the rate of a user depends
user utility functions, which in turn determines the underlying only on the weights of the other users in her slice, i.e., wo .
resource allocation criterion of the slice. Particularly relevant A slice can then unilaterally optimize its weight allocation as
cases are αo = 0 (maximum sum), αo = 1 (proportional follows:
fairness), αo = 2 (minimum potential delay fairness) and
αo → ∞ (max-min fairness). max U o (wo )
wo ≥0
In our model for slice behavior, a tenant proceeds as follows X
to optimize its performance. First, it maximizes the number s.t. wu = so , russ (w) ≥ γu , ∀u ∈ U o .
u∈U o
of users that see their rate requirement met, selecting as many
users as can be possibly served. Second, it maximizes the where we have abused notation to indicateP that in this case
utility U o (w) obtained from the users that have been selected. the slice’s utility, given by U o (wo ) = u∈U o Uu (russ (wo )),
Note that the above framework is sufficiently flexible to ac- depends only on wo . We shall denote the resulting optimal
commodate different network slicing models, including those weights resulting from static slicing by wo,ss .
under study in 3GPP [24]. For instance, in the case where
tenants are Mobile Virtual Network Operators (MVNOs), the D. Network slicing framework
users of a tenant may have different service demands (e.g.,
elastic and inelastic users). Alternatively, we can also support a In this paper, we introduce our NEtwork Slicing (NES)
model where different slices are deployed for specific services; framework to address the resource allocation problem in the
in this case, we may have some slices with only elastic users context of the above system. NES manages both users and
and others with only inelastic users. resources in network slices, as mobile users come and go.
The proposed framework comprises the following modules:
1) Admission control: the purpose of this module is to ensure
C. Baseline allocations that admitted users will see their rate requirements met
Below we introduce two approaches to resource allocation during their lifetime with a sufficiently high probability,
that we will use as benchmarks to assess the performance of even after there are changes in the network.
the proposed framework. For now, we shall assume the users’ 2) Weight allocation: this module determines how to allocate
rate requirements can be met, and thus focus on the weight weights to the users, with the goal of maximizing the
allocation that maximizes the slice’s utility. slice’s utility.
a) Socially Optimal Allocation (SO): If slices were to 3) User dropping: while admission control aims at ensuring
share their utility functions with a central authority, one that all rate requirements are always met, when users re-
could in principle consider a (share-constrained) allocation of associate or see a change in their radio conditions, or
weights (and resources) that optimizes the overall performance when other slices admit more users, it could happen that
of the network, expressed in terms of the network utility U (w) a slice can no longer keep all its users while meeting their
defined as the sum of the slices’ utilities (see [4], [10]): requirements; in that case, this module decides which
X users to drop.
U (w) := U o (w). The design of the admission control module is presented
o∈O in Section III, while that of the weight allocation and user
dropping modules is presented in Section IV.
The above is referred to as the socially optimal allocation,
In order to analyze the stability of the NES framework,
which is given by the following maximization:
we assume that slices are competitive (strategic and selfish),
max U (w) i.e., each attempts to unilaterally optimize its own utility, and
w≥0 model the behavior of the weight allocation and user dropping
X
s.t. wu = so , , ∀o ∈ O, wu ≥ δ, modules as a non-cooperative game. Note that this game only
u∈U o considers admitted users, i.e., admission control is not part of
ru (w) ≥ γu , ∀u ∈ U. the game. It may be played at a point in time when admitted
users may have re-associated or seen a change in their radio
We shall denote the resulting optimal weights and resource conditions, or new users may have been admitted; as a result,
allocation in the socially optimal setting by w∗ and r∗ = when playing the game we may not be able to meet all
(ru∗ (w∗ ) : u ∈ U), respectively. rate requirements. Thus, the game involves slices deciding
(i) which set of users to serve when the rate requirements making some of the users in the first slice infeasible (see the
of all users cannot be met, and (ii) how to allocate weights Appendix for the proof of all the theorems).
amongst the slice’s users, in response to other slices’ decisions.
Theorem 1. When slices cannot satisfy all of their users’ rate
Hereafter we refer to this game as the network slicing game;
requirements, the existence of a NE cannot be guaranteed for
its formal definition is stated as follows:
the network slicing game.
Definition 1. Consider a set of slices o ∈ O, each with a
The problem identified by the above theorem can be over-
set of admitted users u ∈ U o . In the network slicing game,
come by applying an admission control scheme that avoids
each slice selects which subset of users to serve within the
such situations. According to the following theorem, a NE
set U o and their associated weight allocation wo such that
exists as long as admission control is able to guarantee that
(i) as many users as possible are served (meeting their rate
a slice can satisfy the rate requirements of all its users under
requirements), and (ii) the slice’s utility U o is maximized for
any feasible weight allocation of the other slices (including
the selected subset of users.
future allocations when possibly new users may have been
admitted). Note that in this case the resulting game focuses on
III. A DMISSION CONTROL FOR SLICED NETWORKS maximizing slice utilities while meeting the rate requirements
In order to meet user rate requirements, NES needs to apply of all users. This result implies that, as long as proper admis-
admission control on new users, rejecting them when the slice sion control is implemented and ensures that rate requirements
cannot guarantee with a very high probability that it will be can always be satisfied, the stability of the system can be
able to satisfy the rate requirements of all its users during their guaranteed.
lifetime. Note that this only applies to new users; in case the Theorem 2. Suppose admission control ensures that, for any
user rate requirements can no longer be satisfied as a result feasible weight allocation of the other slices, each slice o has
of users moving, or other tenants changing their allocations, a weight allocation wo such that its users’ rate requirements
this is handled by the user dropping module described in are met. Then, the network slicing game has a (not necessarily
Section IV-A. unique) NE.
In the following, we analyze the implications of applying
admission control on the system stability, and propose two Note that the above theorem guarantees the existence of a
different admission control algorithms, Worst-case admission NE when all slices are elastic; indeed, elastic slices have a rate
control (WAC) and Load-driven admission control (LAC). requirement equal to 0, and therefore their rate requirements
These two algorithms correspond to different trade-offs be- can always be satisfied. This leads to the following result.
tween slice isolation and efficiency: while WAC provides Corollary 1. When all slices are elastic, the network slicing
perfect isolation, guaranteeing that a slice will never need game has a NE.
to drop users because of changes in the other slices’ loads,
LAC achieves a higher efficiency at the cost of providing In the following, we propose two alternative admission con-
more relaxed guarantees on isolation (yet ensuring that the trol policies (one more aggressive and one more conservative)
probability of dropping a user remains sufficiently low). that aim at ensuring that the conditions given by Theorem 2 are
met. Note that it is ultimately up to the tenant to choose and
customize its admission control strategy, and hence each tenant
A. Nash Equilibrium existence
may independently apply its own admission control policy.
A critical question is whether the network slicing game
defined in Section II-D possesses a Nash Equilibrium (NE),
B. Worst-case admission control (WAC)
i.e., there exists a choice of users and associated weight
allocation w such that no slice can unilaterally modify its The WAC policy is devised to ensure that the rate re-
choice to improve its utility. In the following, we analyze the quirements of all users are always met, independently of
requirements on admission control policies in order to ensure the behavior of the other tenants. To that end, under the
that a NE exists after admission control is applied. Note that, if WAC policy a slice admits users as follows: it conservatively
the game does not have a NE, strategic slice behavior may lead assumes it has access to only a fraction so of resources at each
to system instability affecting the practicality of the proposed base station, and admits users only if their requirements can
approach. be satisfied with these resources. Given that a user needs a
The following theorem shows that if admission control fraction γu /cu of the base station’s resources to meet her rate
cannot ensure that slices can satisfy the rate requirements of requirement, this policy imposes that for slice o the following
all their users, the network slicing game may not have a NE. constraint is satisfied at each base station b:
The proof of the theorem exhibits a case where instability X γu
≤ so . (4)
arises when there is no weight allocation such that the rate o
cu
u∈Ub
requirements of all the users of a given slice are met given
feasible allocations for the other slices. Note that in a dynamic The WAC policy aims at ensuring that (4) is satisfied at all
setting such a situation could arise, when a slice initially times. However, even if this condition holds when a new user
admits users for which the requirements are feasible, and is admitted, it may be subsequently violated upon changes in
subsequently other slices admit additional users to their slice, the slice, e.g., due to mobility of users or changes in their cu .
To provide robustness against such changes, we follow the opt to be more aggressive in admitting users without running
approach in [33] for single-tenant networks. Specifically, we significant risks. To this end, we propose the Load-driven
add a guard band to (4) aimed at ensuring that the condition Admission Control (LAC) policy, where a slice measures
will continue to hold with high probability after any changes. the current load across base stations and performs admission
Thus, a slice admits a new user request as long as the following control decisions based on the measured loads (assuming that
holds X γu they will not change significantly).3
≤ ρw · so , The following theorem provides a basis for the design of
cu
o u∈Ub the LAC policy. It gives a necessary and sufficient condition
where ρw < 1 parametrizes the guard band: the smaller this that has to be satisfied to meet the rate requirements of the
parameter, the larger the guard band. In practice, this parameter slice’s users, given the current weight allocations of the other
may be set to different values by different slices based on the slices. This constraint is shown to be less restrictive than the
slice specifics, such as the fluctuations of cu or user association one imposed by (4), implying that LAC (potentially) allows
(where larger fluctuations will require a larger guard band) or the admission of more users than WAC.
the desired level of assurance to its users (stricter guarantees Theorem 4. Consider a slice o comprising users with rate
will require a larger guard band). The reader is referred to [33] requirements γ o = (γu : u ∈ U o ), and suppose the aggregate
for a discussion on how to set this parameter. weight of the other slices is given by ao . Then, a weight
In the following, we analyze the properties of WAC under allocation wo that meets slice o’s rate requirements exists if
the assumption that (4) is satisfied with this policy. The and only if the following is satisfied:
theorem below shows that, as long as this condition is satisfied, P
a slice will always be able to meet its users’ rate guarantees X u∈U o γu /cu
P b ao ≤ so . (5)
independent of the setting of the other slices. Thus, a high 1 − u∈U o γu /cu b
b∈B b
degree of protection to the choices and changes in other slices
is provided. The theorem also shows that if the slice deviates where Ubo is the subset of users of slice o associated with base
from the proposed policy, it is not protected from the other station b, according to the given user association policy.
slices’ choices, implying that this policy represents a necessary Moreover, if the rate requirements satisfy (4), then the above
condition to provide protection. condition is satisfied.
Theorem 3. Consider a slice o with users having rate require- The central idea of the LAC policy is as follows. Upon
ments γ o = (γu : u ∈ U o ), then the following hold: receiving a request of a new user u with a rate requirement
1) If (4) is satisfied, there exists at least one weight alloca- γu , slice o assesses the current ao values in the network
tion wo such that ∀u ∈ U o ru (w) ≥ γu , for any feasible and checks whether (5) would be satisfied with the new user.
allocation of the other slices’ aggregate weights ao . According to the theorem, as long as (5) is satisfied, the rate
2) If (4) is not satisfied, slice o is not protected, as there is requirements can be met if the ao values do not change.
a feasible ao allocation such that slice o is not able to However, in practice ao may change due to the response of the
meet the rate requirements of its admitted users. other slices to slice o, or to changes in the other slices (e.g.,
the admission of new users). We shall address this uncertainty
Note that combining this result with Theorem 2, it follows
by following a similar approach to WAC: when admitting a
that a NE exists when all slices run WAC. Indeed, the above
new user, we verify that (5) is satisfied with a sufficiently large
theorem ensures that a slice can find an allocation that meets
guard band, i.e.,
the rate requirements of all its users for any feasible ao , which
comprises all the possible allocations of the other slices w−o .
P
X u∈U o γu /cu
Theorem 2 guarantees that when this holds, a NE exists. Thus, P b ao ≤ ρl · so , (6)
1 − u∈U o γu /cu b
we have the following corollary: b∈B b
Corollary 2. If (4) is satisfied by all slices, then the network where ρl < 1 is the parameter providing the guard band for
slicing game has a NE. LAC. Note that, in addition to other considerations, in this case
the setting of ρl will need to account for observed statistical
Note that Corollary 2 imposes more conservative conditions
fluctuations of ao , larger fluctuations requiring a larger guard
than Theorem 2; for instance, if a slice never has users at
band.
a given base station, according to Theorem 2 such a slice
The following theorem shows that, as long as the chosen
cannot place any weight on this base station; in contrast, the
value for ρl is sufficiently conservative, LAC is effective in
arguments behind (4) account for, and protect the slice against,
guaranteeing that the rate requirements of all users are met.
such possibility.
Theorem 5. There exists a ρl value sufficiently small such
C. Load-driven admission control (LAC) that the rate requirements of all the users of slice o can be
met independent of how the other slices change their weights.
While the WAC policy protects a given slice from the
others, it may be overly conservative in some cases where base 3 Note that many similar (load-driven) admission control algorithms have
stations are lightly loaded or where some slices are unlikely to been proposed in the literature [34], [35] in the context of single-tenant
use resources at certain base stations. In those cases, one may networks. In this paper, we apply this concept to a network slicing setting.
The following corollary follows from the above result and Algorithm 1: MaxSubset Algorithm.
Theorem 3. Indeed, as long as every slice satisfies either (4) Initialize: Ũ o = ∅
and (6), Theorems 3 and 5 guarantee that all slices can choose while Ũ o 6= U o do
a weight allocation that satisfies the rate requirements of all
u∗ = argminu0 {∆ω o (Ũ o , u0 ) | u0 ∈ U o \ Ũ o }
their users. Furthermore, Theorem 2 guarantees that when this
if ω o (Ũ o ∪ {u∗ }) ≤ so then Ũ o := Ũ o ∪ {u∗ };
holds there exists a NE. These implies that, as long as all
else return;
slices run either WAC or LAC, the system can be expected to
end
be stable.
Corollary 3. If either (4) or (6) holds for every slice (the
o
latter with a sufficiently small ρl ), then there exists a NE. Pu cu /lb = γu ; summing these over u ∈ Ũb and isolating
w
u∈Ũ o wu yields
P
o o o −o u∈Ũ o γu /cu
IV. W EIGHT ALLOCATION AND USER DROPPING FOR ωb (Ũ ) = ab (w ) P b .
N ETWORK S LICING 1 − u∈Ũ o γu /cu
b
P
where we are assuming u∈Ũ o γu /cu ≤ 1 (otherwise we let
Once a slice decides which users to admit, possibly fol-
ωbo (Ũ o ) = ∞).
lowing one of the admission control policies presented above,
We further let ω o (Ũ o ) = b∈B ωbo (Ũ o ) denote the aggre-
P
it needs to determine the weight allocation of the admitted
gate minimal weight requirement for the slice, and for any
users. In NES, this is determined based on a sequence of
user u0 ∈ U o we define the marginal aggregate weight of the
best responses, where in each round a slice chooses its best
user u0 given candidate subset Ũ o as
response given the choices of the other slices. A slice’s best
response involves the following two steps: (i) user subset ∆ω o (Ũ o , u0 ) = ω o (Ũ o ∪ {u0 }) − ω o (Ũ o ).
selection, to determine which subset of users to serve, and Building on the above notation, we present a greedy solution
(ii) weight allocation, to set the weights of the users in in Algorithm 1, which provides as output the set of selected
the selected subset. In the following, we first present the users Ũ o . The following theorem confirms the effectiveness of
algorithms to perform the user subset selection and weight this algorithm.
allocation, and then analyze the convergence of the sequence
of best responses. Theorem 6. The MaxSubsetSelection algorithm results in a
subset of users that maximizes the number of users the slice
can serve and still meet their minimal rate requirements.
A. User subset selection Alternatively, slices might apply a PriorityUserSelection
When a slice cannot satisfy the rate requirements of all its algorithm to customize their user subset selection policy by
users, it needs to decide which subset to serve. Note that, while assigning users a priority order. Such an ordering may depend,
admission control aims at ensuring that rate requirements of e.g., on the users’ traffic class, the revenue they generate, how
all users can always be satisfied, in practice this can only be long users have been in the system, and/or their current signal
ensured with a (very) high probability due to the unpredictable to noise ratio, among other factors. To this end, the algorithm
nature of the mobile network; thus, in some unlikely cases it simply adds users sequentially to the subset to be served in
may happen that the rate requirements of some users cannot be order of decreasing priority until no more can be added, i.e.,
met. When this happens, the slice has to drop those users. Note ω o (Ũ o ∪ {u∗ }) > so .
that this yields a novel paradigm for managing the resources
B. Weight allocation
of a slice, where changes in one part of the network may lead
to dropping users in another part. Once a slice has selected a set of users whose requirements
Below we present the algorithms for two possible ap- can be satisfied, it sets their weights as follows. Given the
proaches for user selection: (i) MaxSubsetSelection, which aggregate weights of the other slices, aob (w−o ), a slice chooses
maximizes the cardinality of the subset of served users (thus wo such that the its utility is maximized, i.e.,
o
X
minimizing user dropping); and (ii) PriorityUserSelection, wo = arg max U o (w0 , w−o ),
which uses a priority ordering on a slice’s users (enabling w0o
u∈Ũ o
a slice to customize its users’ service). wu0 γu
To realize MaxSubsetSelection we use a greedy algorithm s.t.: ≥ , ∀u ∈ Ũ o ,
ab (w ) + lbo (w0 o )
o −o cu
which at each step adds the user which needs the smallest X
additional weight to meet the selected users’ rate requirements. wu0 ≥ δ, ∀u ∈ Ũ o , wu0 ≤ so .
To that end, let Ũ o be a candidate subset of the admitted users u∈Ũ o
by slice o, U o , and let ωbo (Ũ o ) be the minimum aggregate 0o −o

where, for convenience, we write U (w , w ) = U o (w) too
weight required to satisfy the rate requirements the candidate highlight dependencies on other slices weights.
subset’s users on base station b, Ũbo . The value of ωbo (Ũ o ) Note that as long as utility functions fu (·) are concave in
can be computed as follows. The minimum weight wu needed the allocated user rates, the above maximization corresponds
to satisfy the rate requirement of user u ∈ Ũbo must satisfy to a (computationally tractable) convex optimization problem.
C. Convergence of best response dynamics While the theorem assumes the same set of admitted users
With NES, we determine users’ weight allocation based for static slicing and NES, we argue that the result holds
on a sequence of best responses. The proposed algorithm in general. Indeed, a tenant is free to choose any admission
implements the best response computed above in rounds: slices control policy, including that employed by static slicing, and
update the weight allocation of their users wo , sequentially, it is to be expected that it will apply the policy that maximizes
one at a time and in the same fixed order, in response to its utility. Thus, it follows that the level of satisfaction of the
the other slices weights ao . Following standard game theory tenant will be greater with NES, under the chosen admission
terminology, we refer to this iterative process as Best Response policy, than with static slicing.
Dynamics.
The following theorem shows that the above dynamics may B. Loss over the socially optimal allocation
not converge. In particular, the proof of the theorem considers We now study the difference in the utility achieved under
an instance satisfying the conditions of Theorem 2, i.e., a socially optimal resource allocation vs. that achieved under
feasible instance under admission control, and shows that, even NES. We focus on the case where fu (·) follows (2) for αo = 1
though a NE is guaranteed to exist under such conditions, Best and αo = 2, which are two highly relevant settings in practice
Response Dynamics do not converge. (corresponding to proportional and minimum delay potential
fairness, respectively). To perform the comparison, we define
Theorem 7. Consider a game instance such that, for each
the Loss over the Social Optimal (LSO) as follows. For αo = 1
slice o ∈ O there exists an allocation satisfying the rate .
we define LSO = U (w∗ ) − U (ŵ), where w∗ is the socially
requirements of all its users for any possible allocation of the
optimal weight allocation and ŵ is the weight allocation with
other slices. Even though a NE is guaranteed to exist under .
NES, while for αo = 2 we define it as LSO = UU(w (ŵ)
∗ ) . Note
these conditions, Best Response Dynamics may not converge.
that these definitions are adjusted to the type of utility function:
While the above theorem shows that convergence cannot for αo = 1, utilities are logarithmic in the rate, and hence by
be ensured, our simulation results show that in practice Best subtracting utilities we capture the ratio between rates, while
Response Dynamics converge quickly to a region close to the for αo = 2 utilities are inversely proportional to the rates, and
NE, and hence we can simply force the system to halt after a hence the ratio between rates is obtained by dividing utilities.
number of best response rounds and use the weights obtained The following theorem provides a bound on the LSO and
in the last round. Specifically, following the results provided gives an instance for which the LSO is close to this bound,
in Section VI-D, in our simulations we halt the system after showing that the bound is tight.
7 rounds.
Theorem 9. Let user utilities fu (·) follow (2), γu be the
From the above, it can be seen that NES incurs an ac- ¯ largest
minimum rate guarantee in the network, c̄u be the
ceptable computational load, as its execution involves solving
possible achievable rate and ε = γu /c̄u . Under a given set of
a sequence of convex optimization problems (each of which ¯
admitted users, we have that:
scales with the number of users of the slice and number of base
stations) for a limited number of times (namely, the number 1) If αo = 1 ∀o ∈ O, then LSO ≤ − log(ε) and there is an
of slices in the network multiplied by 7). Moreover, the instance for which LSO ≥ − 21 log(2ε).
above computations may be possibly performed at centralized 2) If αo = 2 ∀o ∈ O, then LSO ≤ 1ε and there is an
1
controllers, as the resource allocation does not need to be instance for which LSO ≥ 3ε .
implemented in the base stations before the sequence of Note that, according to the above results, the bound on the
optimizations converges or stops. Also, resources may be re- LSO relaxes as we decrease the minimum rate requirement
allocated only periodically to alleviate the overhead associated in the network, and becomes unbounded in the case where
to the reconfiguration of base stations. Quantitative results on we have elastic traffic with no rate guarantees, i.e., γu = 0.
the computational load are provided in Section VI-E. However, in a well provisioned network all users should expe-
rience a sufficiently large rate, and in this case the LSO should
V. A NALYSIS OF THE NES FRAMEWORK be low according to the above result. This is corroborated
by our simulation results, which show that in practice NES
In the following, we analyze the performance achieved by performance is close to optimal and LSO is very small.
the NES approach proposed above as compared to the two
baseline allocations given in Section II-C: (i) the socially VI. P ERFORMANCE EVALUATION
optimal allocation, and (ii) static slicing. Our analysis assumes
We next evaluate the performance of NES via simulation.
that NES reaches a Nash equilibrium.
Unless otherwise stated, the mobile network setup of our
simulator follows the IMT-A evaluation guidelines for dense
A. Gain over static slicing ‘small cell’ deployments [36], considering a network with
19 base stations disposed in a hexagonal grid layout with 3
The result below shows that NES outperforms static slicing.
sectors, i.e., |B| = 57. User mobility follows the Random
Theorem 8. For the same set of admitted users, the utility Waypoint (RWP) model. The users arrive to the network
achieved by an operator under NES is never lower than the following a Poisson Process with intensity λ arrivals/sec, and
utility that this operator would obtain under static slicing. their holding times are exponentially distributed. Users’ SINR
15.3 100
15.2
15.1 80
15
14.9 60
14.8
14.7 40
14.6
14.5 20
14.4
0
0 25 50 75 100 3 4 5 6 7 8 9 10
Fig. 1. Performance of NES in terms of network utility as compared to the Fig. 2. Throughput gains over SS for different traffic types (elastic, inelastic),
two benchmark allocations (SS and SO). utility functions (αo ) and network load (λ).
is computed based on physical layer network model specified most relevant αo values in practice, (ii) elastic and inelastic
in [36] (which includes path loss, shadowing, fast fading and slices, where all users are either elastic and inelastic, and (iii)
antenna gain) and user association follows the strongest signal different arrival rates λ, yielding different network loads. We
policy. The achievable rate for a user u, cu , is determined conclude from the results that (i) gains are very substantial,
based on the thresholds reported in [37]. Unless otherwise ranging from 100% to 20%, (ii) they decrease with the load,
stated, the rate requirement of the inelastic users is set to as already observed in [4], and (iii) they are fairly insensitive
γu = 0.5 Mbps, we have αo = 1 for all slices, there are to the fraction of inelastic traffic and choice of utility function.
5 slices in the network with equal shares, the arrival rate is
λ = 5 (equally split among slices) and the average holding C. Blocking probability
time is 1 minute. In the simulations, we consider both slices
In addition to improving the performance of admitted users,
with mixed traffic of different types (Sections VI-A and VI-C)
one of the key advantages of the dynamic resource allocation
as well as slices dedicated to one specific traffic type (Section
implemented by NES is that it allows admitting more users
VI-B). All confidence intervals are below 1%.
while meeting their rate requirements. In order to assess the
achieved improvement, we evaluate the blocking probability
A. Network utility (i.e., the probability that a new user cannot admitted) under
We first analyze the network utility achieved by NES NES versus SS. For NES, we consider the two admission
as compared to the two benchmark solutions presented in policies proposed in Section III (WAC and LAC), while for
Section II-C (namely, SS and SO). To ensure that the rate SS we apply the policy given in [33]. For all settings, we drop
requirements of admitted users are always met, we adopt the users based on the MaxSubsetSelection algorithm, and adjust
WAC admission control policy with ρw = 1 and suppress the guard bands to ensure that the probability of dropping an
user movements yielding changes in base station associations admitted user is no more than 1%. To increase the offered
and/or cu values. To analyze the impact of inelastic traffic, load sufficiently so that we can observe the behavior of the
we vary the fraction of inelastic traffic arrivals, θ, yielding blocking probability, we set γu = 1 Mbps and an average
an arrival rate of θλ for inelastic users and of (1 − θ)λ for holding time of 2 minutes. The results are given in Fig. 3 as
elastic ones. The results, depicted in Fig. 1, show that (i) NES a function of the fraction of inelastic user arrivals (θ). They
outperforms very substantially SS, providing very high gains, show very high gains over SS for both approaches (WAC and
and (ii) it performs almost optimally, very close to the SO. LAC), and confirm that, by behaving more aggressively, LAC
Moreover, this holds independently of the mix of elastic and is able to admit many more users than WAC.
inelastic users present in the network.
D. Convergence to the NE
B. Throughput gains To better understand the dynamics of NES, we have eval-
To give a more intuitive measure of the gains achieved by uated a very large number of randomly generated scenarios
NES, we define the throughput gain over SS, ∆, as follows: (namely 104 scenarios) with the following settings: (i) a
it is the value such that, if we increase the rate of all users uniform number of slices between 2 and 10, i.e., |O| ∼
in SS by ∆, we reach the same network utility as NES (e.g., U (2, 10), (ii) a number of users per slice of |U o | ∼ U (0, 350),
∆ = 100% means that SS achieves the same utility as NES (iii) inelasticity level θ ∼ U (0, 100)(%), (iv) minimum rate
when multiplying all user rates by 2). Fig. 2 illustrates the requirements γu ∼ U (0, 3) Mbps, and (v) the shares so
throughput gains for (i) αo = 1 and αo = 2, which are the two proportional to the number of users. We have found that a
2.5
90 2 slices, elastic
5 slices, elastic
80 10 slices, elastic
2
2 slices, inelastic
70 5 slices, inelastic
10 slices, inelastic
60
1.5
50
40
1
30
20 0.5
10
0 0
0 20 40 60 80 100 100 300 500 700 900 100 300 500 700
Fig. 3. Blocking probability for new arrivals for the two policies proposed Fig. 5. Computational times of the proposed approach as a function of the
and the SS benchmark. number of slices and users in the network.
probabilities below 1%. Results confirm that NES can be

applied to practical settings, as complexity is roughly linear
with the size of the network and computational times remain
low even for large size problems; for instance, for a network
with 9000 users the time falls below 2.5 seconds. We further
observe that inelastic traffic slightly increases complexity but
does not challenge the practicality of the approach. Finally, we
note that the computational time values provided here could
be further improved by optimizing the code, parallelizing tasks
and/or increasing the machine computational power.
F. Slice differentiation
We next analyze the ability of NES to deploy slices
providing a customized service. To this end, we consider a
Fig. 4. Box plot for the RMSE of the weight allocation at a given round with scenario with 4 slices with different requirements: (i) slice
respect to the NE weight allocation. 1 provides rate requirements of γu = 1 Mbps with WAC,
(ii) slice 2 provides γu = 0.5 Mbps with WAC, (iii) slice 3
provides γu = 0.5 Mbps with LAC, and (iv) slice 4 provides
vast majority (97.6%) of scenarios converge to the NE after no minimum rate requirements. All slices have the same
100 rounds. For such scenarios, Fig. 4 shows the difference share, the arrival rate is of λ = 10 equally split among the
between the weight allocation at a given round and the one slices, and admission control is configured to provide dropping
at the NE in terms of mean squared error (RMSE), providing probabilities below 1%. Fig. 6 shows the empirical CDF of the
a box plot with the median (red), 95% percentile (box), 99% user rates for each slice as well as the blocking probabilities
percentile (whisker) and outliers (red crosses). We observe that (≈ 47.2%, 16.7%, 3.58% and 0%, respectively). We observe
the RMSE decreases exponentially in the number of rounds. that (i) the minimum rate requirements are satisfied for all
After 7 rounds we are already very close to the NE (the median slices; (ii) as the rate requirements increase, so does the
is below 10−4 ), which justifies our choice in Section IV-C. blocking probability, yielding an overall improvement of the
Additional results, not included for space reasons, show that user rate distribution, and (iii) by employing LAC, we achieve
user rates exhibit a very similar behavior to the weights. a dramatic reduction of the blocking probability while paying
a very small prices in terms of user rate distribution. We
E. Computational load conclude that NES is effective in enabling slice differentiation.
Next we evaluated the computational complexity of the NES
algorithm when the system halts after 7 rounds (as given by VII. C ONCLUSIONS
the configuration chosen in this paper). Fig. 5 shows the com- In this paper we proposed and analyzed a framework for
putational times for a dual-core 2.9GHz i7 processor for elastic network slicing that relies on network shares and allows
and inelastic traffic and different numbers of slices and users, slices to customize resource allocations to their users. This
when the number of base stations is scaled with the number of framework results in a network slicing game where each slice
users and admission control is adjusted to ensure that dropping unilaterally reacts to the settings of the others. While this
1 Proof of Theorem 2
0.8
Let W be the convex and compactP set of feasible weights
w satisfying (i) wu ≥ δ ∀u, and (ii) u∈Uo wu = so ∀o and
let us consider the mapping w → w̃ = Γ(w), where w̃o is the
0.6 best response of slice o to w−o . We next show that this map-
ping satisfies the conditions of Kakutani’s theorem: i) Γ(w) is
0.4
non-empty, ii) Γ(w) is a convex-valued correspondence, and
iii) Γ(w) has a closed graph. Conditions i) and ii) follow
from the fact that the best response of a slice to w−o is a
0.2 unique allocation w̃o . This implies that that w̃ exists and is a
single point (and hence a convex set). Condition iii) is shown
0
by proving that w̃o is a continuous function of w−o for all
0.2 0.5 1 2 3 4 5 10 slices. Consider the set of users for which ru > γu and the set
for which ru = γu . As long as these sets do not change, w̃o
Fig. 6. Blocking probability and empirical CDF of the user rates for a scenario
can be expressed as a continuously differentiable function of
of 4 slices with different requirements. {w̃o , w−o }, and it follows from the implicit function theorem
that w̃o is a continuous function of w−o . When some user
moves from set ru > γu to ru = γu (or viceversa), such
game has been previously studied for elastic traffic, the slices’ user satisfies both the equation for ru = γu and the one for
behavior changes substantially when users have minimum rate ru > γu , providing continuity over the transitions. Since all
requirements, and so does the outcome of the game. Indeed, the conditions of Kakutani’s theorem are satisfied, we have
we have shown that (in contrast to the elastic case) this game that the mapping Γ has at least one fixed point, which implies
may not have a Nash Equilibrium and, even when it has a NE, that at least one NE exists.
Best Response Dynamics may not converge to the equilibrium. To show that the NE is not necessarily unique, we provide
In spite of this (apparently) negative result, we have shown an example with multiple NEs. Consider a scenario with three
that as long as admission control is applied (which is to be slices (1,2,3) and three base stations (a,b,c). Let the first slice
expected under inelastic traffic), we can guarantee that a NE have users in base stations a and c (users 1a, 1c), the second
exists. We have proposed algorithms for admission control, slice in a and b (2a, 2b) and the third slice in b and c (3b, 3c).
weight allocation and user dropping, which jointly bring the Let φ1a = φ1b = 1/2, φ2a = φ3c = 1 and φ2b = φ3b = 0.
system to a NE. We have further analyzed performance at Also, let γu = 1/2 for users 2b and 3b, γu = 0 for all other
the equilibrium, showing that it is close to the social optimal users and cu = 1 for all users. It can be seen that all the weight
and provides substantial gains over static slicing. Based on allocations satisfying w1a = w1b = 1/6, w2b = w3b = w and
these results, our main conclusion is that the proposed NES w2a = w3c = 1/3 − w for w ∈ [δ, 1/3 − δ] correspond to a
framework provides an effective and implementable scheme for NE, which shows that multiple NE exist for this example.
dynamically sharing resources across slices, both for elastic
and for inelastic traffic.
Proof of Theorem 3
A PPENDIX : P ROOFS OF THE T HEOREMS
The result of 1) follows directly from Lemma 1 in [4]. If
Proof of Theorem 1 users are admitted at base stations such that under static slicing
Consider a setting with two base stations (a and b) and their rate guarantees are met, i.e. russ ≥ γu , then it follows
two slices (1 and 2), each slice with one user associated to by the above mentioned lemma that there exists an allocation
base station a and another user associated to base station b. satisfying ru ≥ russ ≥ γu , which proves the first part of the
We refer to these users as U = {1a, 1b, 2a, 2b}. Let the rate theorem.
requirements of slice 1 be γ1a = γ1b = 2C/3, the users of To prove 2), we proceed as follows. Suppose slice o admits
slice 2 have no minimum rate requirements, and s1 = s2 = users are such that their associated
P rate requirements violate
1/2. We show that this game has no NE by contradiction. We (4) at some base station b, i.e., u∈U o γu /cu > so . If all other
b
necessarily have that either w2a ≤ 1/4 or w2b < 1/4. Let us slices place their entire share at that base station, we have
assume that w2a < 1/4 and w2b > 1/4. Since in this case P
slice 1 can only meet the rate requirements of user 1a, its best X ru u∈Ubo wu
response will concentrate its weight on this user, w1a = 1/2. =P ≤ so ,
o
u∈Ub
cu u∈U o wu + 1 − so
However, the best response of slice 2 to such allocation of b
slice 1 is to concentrate its share on user 2a. Thus, w2a > 1/4, P P
which contradicts the initial assumption. Following a similar which implies u∈Ubo ru /cu < u∈Ubo γu /cu and hence
argument, it can be seen that if we assume w2a = 1/4 or necessarily ru < γu for some u, proving the second part of
w2a > 1/4, we also reach a contradiction. the theorem.
Proof of Theorem 4 Proof of Theorem 6

Recall that the rate of user u is given by ru = wu cu /lb(u) . The proof goes by contradiction. Let Ũ o be the set of
If we add the rates P of the users of slice o at a given base users selected by the MaxSubsetSelection algorithm, and let us
station b and isolate u∈U o wu , we obtain assume that there exists an alternative feasible user selection
b
P Û o such that |Û o | > |Ũ o |. If we take the set Û o and substitute
X u∈U o ru /cu each user by another one in the base station with smaller
wu = P b ao .
o
1 − u∈U o ru /cu b γu /cu , the resulting set Ū o is feasible and has the same number
u∈Ub b
of users as the original one. Note that set Ū o necessarily has
PBy summing the above over all base stations and noting that some base station b with more users than set Ũ o – otherwise
u∈U o wu = so , we obtain |Û o | > |Ũ o | would not hold. Let us assume that there exists
P
X u∈U o ru /cu some other base station b0 with fewer users. In this case,
P b ao = so . (7) let us remove user u from one of the base stations with
1 − u∈U o ru /cu b
b∈B b
more users, b, and add user u0 in one of the base stations
We now prove that as long as (5) is satisfied, there exists with fewer users, b0 . The resulting set remains feasible, as
a weight allocation wo that meets the rate requirements of all ∆ωbo0 (Ū o , u0 ) ≤ ∆ωbo0 (Ū o , u) – otherwise MaxSubsetSelection
users. Let us consider the weight allocation satisfying4 would have chosen a different subset of users. We can do this
(γu /cu )lb(u) until there are no base station with fewer users than in Ũ o .
wu = P so , ∀u ∈ U o . (8) The result of these operations is a feasible set where all base
v∈U o (γv /cv )lb(v)
stations have as many users or more than Ũ o , and overall it has
Note that with the above weight allocation, the rates ru
more users. However, this yields a contradiction: if such set
are proportional to γu , which means that either we have ru ≥
was feasible, the MaxSubsetSelection algorithm would have
γu ∀u or ru < γu ∀u. The latter yields a contradiction; indeed,
selected more users.
if ru < γu ∀u it follows that
P P
X u∈Ubo γu /cu o
X u∈U o ru /cu Proof of Theorem 7
ab > P b ao = so ,
1 − u∈U o ru /cu b
P
1 − u∈U o γu /cu Let us consider a scenario with three base stations (a,b,c)
b∈B b b∈B b
and three slices (1,2,3), with s1 = s2 = s3 = 1/3 and
which contradicts (5). Hence, it follows that ru ≥ γu ∀u.
any arbitrary α1 , α2 , α3 values. Let slice 1 have two users
We next prove that if (5) is not satisfied, then there ex-
associated to base stations a and b (u1a ,u1b ), slice 2 two
ists no weight allocation meeting the rate requirements. The
users associated to base stations b and c (u2b ,u2c ) and slice
proof goes by contradiction. Assume (5) is P not satisfied but
3 two users associated to base stations a and c (u3a ,u3c ). Let
u ≥ γu ∀u. From the latter, it follows that
rP u∈Ub ru /cu ≥ cu = 1 ∀u, γ1a = γ2b = γ3c = 1/2, γ1b = γ2c = γ3a = 0,
u∈Ub γu /cu ∀b. Combining this with (7) yields φ1a = φ2b = φ3c → 0 and φ1b = φ2c = φ3a → 1. The NE
P
X u∈U o γu /cu of this instance is wu = 1/6 ∀u. However, if we start with
P b ao ≤ so , w3c = w < 1/6 and w3a = 1/3 − w, and perform a best
1 − u∈U o γu /cu b
b∈B b
response cycle starting starting with slice 1 followed by 2 and
which contradicts that assumption that (5) is not satisfied. 3, it can be seen that this leads to an endless cycle where
Finally, we show that if the rate requirements satisfy (4), each slice takes a weight allocation of either {w, 1/3 − w}
then
P they surely satisfy (5). The lhs of (5) increases with or {1/3 − w, w} at each step (none of which corresponds to
u∈Ubo γu /cu . As long as this value is no larger than so , we the NE). Hence, Best Response Dynamics do not converge for
have that the following equation gives a sufficient condition this instance of the game.
so P o
for (5) to be satisfied: 1−s o b∈B ab ≤
Pso .
The abovePis surely satisfied since b∈B aob = 1 − so . As Proof of Theorem 8
(4) imposes u∈U o γu /cu ≤ so , it follows that as long as (4) The proof follows from Lemma 1 of [4], which shows that,
b
is satisfied, (5) is also satisfied. given a slice o and a feasible weight allocation w−o for the
other slices, there exists a weight allocation wo for slice o,
Proof of Theorem 5
ao
possibly dependent on w−o , such that the resulting weight
Let us take ρl = minb 1−P bo γu /cu . Then, from (6) it allocation w satisfies ru (w) ≥ russ for all u ∈ U o . Therefore,
u∈U
P P b
follows that b∈B u∈U o γu /cu ≤ so . From this, we have there exists a weight allocation that provides the same utility
b
that condition (4) is satisfied. According to Theorem 3, as as static slicing. Since the weight allocation chosen by NES
long as this condition is satisfied, there exists a choice of is the one that maximizes the slice’s utility, it surely provides
wo that satisfies the rate requirements of all users of slice a utility no smaller than that under static slicing.
o independent of the weight setting of the other slices, which
completes the proof. Proof of Theorem 9
4 The existence of such an allocation follows from applying Brouwer fixed-
We start for αo = 1. To prove the bound on the LSO, we
point theorem to the function f : W → W, where wu = fu (w) is given first note that
P !
by (8) and W isPthe set of weights satisfying u∈U o wu = so and wu ≥ X X wu∗
(γu /cu )aob so / v∈U o (γv /cv ) (recall that aob 6= 0 ∀b, as weights cannot ∗
U (w ) = so φu log P ∗ cu
be zero). o∈O u∈Uo u0 ∈Ub(u) wu0
X X
≤ so φu log (c̄u ) . [7] J. B. Rosen, “Existence and Uniqueness of Equilibrium Points for
o∈O u∈Uo Concave N-Person Games,” Econometrica, vol. 33, no. 3, pp. 520–534,
Jul. 1965.
Furthermore, from the minimum rate constraint it follows that [8] F. P. Kelly, A. K. Maulloo, and D. Tan, “Rate control for communication
! networks: Shadow prices, proportional fairness, and stability,” Journal
X X ŵu of the Operational Research Society, vol. 49, pp. 237–252, Mar. 1998.
U (ŵ) = so φu log P cu [9] R. Johari and J. N. Tsitsiklis, “Efficiency Loss in a Network Resource
u0 ∈Ub(u) ŵu
0
o∈O u∈Uo Allocation Game,” Mathematics of Operations Research, vol. 29, no. 3,
X X pp. 407–435, Aug. 2004.
≥ so φu log γu . [10] R. Mahindra et al., “Radio Access Network sharing in cellular net-
o∈O u∈Uo ¯ works,” in Proc. of IEEE ICNP, Oct. 2013.
[11] J. Zheng, P. Caballero, G. de Veciana, S. J. Baek, and A. Banchs,
Combining the above two equations, we obtain U (w∗ ) − “Statistical multiplexing and traffic shaping games for network slicing,”
U (ŵ) ≤ log(c̄u /γu ) = − log(ε), which completes the first in Proc. of WiOpt 2017, May 2017.
part of the proof. ¯ [12] P. Caballero, A. Banchs, G. de Veciana, and X. Costa-Prez, “Multi-
Tenant Radio Access Network Slicing: Statistical Multiplexing of Spatial
To show that the above bound is tight, we consider the Loads,” IEEE/ACM Transactions on Networking, vol. 25, no. 5, pp.
following network instance. We have two slices with shares 3044–3058, Oct 2017.
s1 = s2 = 1/2 and two base stations. The first slice has two [13] Y. L. Lee, J. Loo, T. C. Chuah, and L.-C. Wang, “Dynamic Network
Slicing for Multitenant Heterogeneous Cloud Radio Access Networks,”
users in the first base station (weights w11 and w12 ) and the IEEE Transactions on Wireless Communications, vol. 17, no. 4, pp.
second slice has one user in the first base station (w21 ) and 2146–2161, Apr. 2018.
another one in the second base station (w22 ). All users have [14] R. Matias, J. Baliosian, J. Serrat, and J.-L. Gorricho, “Resource Slicing
in Virtual Wireless Networks: A Survey,” IEEE Transactions on Network
cu = c̄u , and the rate requirements are γ11 = c̄u (1/2 − ε) and Service Management, vol. 13, no. 3, pp. 462–466, Sep. 2016.
for the first user and γu = γu = c̄u ε for the other ones. [15] J. Kwak, J. Moon, H. W. Lee, and L. B. L, “Dynamic network slicing
Furthermore, let φ11 → 0, φ12¯ → 1, φ21 → 0 and φ22 → 1. and resource allocation for heterogeneous wireless services,” in Proc. of
IEEE PIMRC, Oct. 2017.
In the allocation employed by NES (which corresponds to [16] W. Guan, X. Wen, L. Wang, Z. Lu, and Y. Shen, “A Service-oriented
the NE) we have w11 = 1/2 − ε, w12 = ε, w21 → 1/2 and Deployment Policy of End-to-End Network Slicing Based on Complex
w22 → 0, which yields U (ŵ) = 21 log(εc̄u )+ 21 log(c̄u ). In the Network Theory,” IEEE Access, vol. 6, pp. 19 691–19 701, Apr. 2018.
[17] O. Narmanlioglu, E. Zeydan, and S. S. Arslan, “Service-Aware Multi-
social optimal,
1 the following weight allocation: w11 =
we have
Resource Allocation in Software-Defined Next Generation Cellular
1 ε ε
2 − ε 2 + 2(1−ε) , w12 = 1/2 − w11 , w21 = 2(1−ε) and Networks,” IEEE Access, vol. 6, pp. 20 348–20 363, Mar. 2018.
[18] M. Jiang, M. Condoluci, and T. Mahmoodi, “Network slicing manage-
w22 = 1/2 − w21 , from which U (w∗ ) = 12 log ((1/2)c̄u ) + ment & prioritization in 5G mobile systems,” in Proc. of European
1 ∗ 1
2 log(c̄u ). This yields U (w ) − U (ŵ) = − 2 log (2ε) , which Wireless, May 2016.
terminates the proof for αo = 1. [19] O. U. Akguel, I. Malanchini, V. Suryaprakash, and A. Capone, “Service-
To prove the LSO bound for αo = 2, we note that Aware Network Slice Trading in a Shared Multi-Tenant Infrastructure,”
in Proc. of IEEE GLOBECOM, Dec. 2017.
X X 1 [20] L. Zhang, “Proportional response dynamics in the Fisher market,”
U (w∗ ) ≥ − so φu Theoretical Computer Science, vol. 412, no. 24, pp. 2691–2698, May
c̄u 2011.
o∈O u∈Uo
[21] D. Monderer and L. S. Shapley, “Potential games,” Games and economic
and behavior, vol. 14, no. 1, pp. 124–143, 1996.
X X 1 [22] O. Candogan, A. Ozdaglar, and P. A. Parrilo, “Dynamics in near-
U (ŵ) =≤ − so φu .
γu potential games,” Games and Economic Behavior, vol. 82, pp. 66 –
o∈O u∈Uo
¯ 90, 2013.
[23] 3GPP, “Study on management and orchestration of network slicing for
Combining these two equations we obtain UU(w (ŵ) 1
∗ ) ≤ ε , which next generation network (Release 15),” TR 28.801 V1.2.0, May 2017.
completes the first part of the proof. The tightness of the bound [24] 3GPP, “Management of 5G networks and network slicing; Concepts, use
is proven by considering the same network instance as for cases and requirements (Release 15),” TS 28.530, v0.6.0, Apr. 2018.
[25] 5GPPP White paper, “5G Empowering vertical industries,” 2016.
αo = 1:
[26] A. Gudipati, L. E. Li, and S. Katti, “RadioVisor: A Slicing Plane for
U (ŵ) − 12 εc̄1u − 12 c̄1u 1
ε +1 1 Radio Access Networks,” in Proc. of HotSDN, 2014.
= = ≥ . [27] V. Sciancalepore et al., “Interference coordination strategies for content
U (w∗ ) − 12 (1/2)c̄
1
u
− 12 c̄1u 1
1/2 + 1
3ε update dissemination in LTE-A,” in Proc. of IEEE INFOCOM, May
2014.
[28] S. Shenker, “Fundamental Design Issues for the Future Internet,” IEEE
R EFERENCES Journal of Selected Areas in Communications, vol. 13, no. 7, pp. 1176–
[1] NGMN Alliance, “Description of Network Slicing Concept,” NGMN 1188, Sep. 2006.
5G P1, Jan. 2016. [29] P. Hande, S. Zhang, and M. Chiang, “Distributed rate allocation for
[2] 3GPP, “Study on Architecture for Next Generation System,” TR 23.799, inelastic flows,” IEEE/ACM Transactions on Networking, vol. 15, no. 6,
v0.5.0, May 2016. pp. 1240–1253, Dec. 2007.
[3] ——, “Study on Radio Access Network (RAN) sharing enhancements,” [30] 3GPP, “Technical Specification Group Services and System Aspects;
TS 22.101 V15.1.0, Jun. 2017. Policy and charging control architecture,” 3GPP TS 23.203, Jun. 2016.
[4] P. Caballero, A. Banchs, G. de Veciana, and X. Costa-Perez, “Network [31] J. Mo and J. Walrand, “Fair end-to-end window-based congestion
Slicing Games: Enabling Customization in Multi-Tenant Networks,” in control,” IEEE/ACM Transactions on Networking, vol. 8, no. 5, pp. 556–
Proc. of IEEE INFOCOM, May 2017. 567, Oct. 2000.
[5] L. Zhang, “Proportional response dynamics in the Fisher market,” [32] X. Costa-Perez, J. Swetina, T. Guo, R. Mahindra, and S. Rangarajan,
Theoretical Computer Science, vol. 412, no. 24, pp. 2691–2698, May “Radio access network virtualization for future mobile carrier networks,”
2011. IEEE Communications Magazine, vol. 51, no. 7, pp. 27–35, Jul. 2013.
[6] M. Feldman, K. Lai, and L. Zhang, “The Proportional-Share Allocation [33] R. Ramjee, D. Towsley, and R. Nagarajan, “On Optimal Call Admission
Market for Computational Resources,” IEEE Transactions on Parallel Control in Cellular Networks,” Wireless Networks, vol. 3, no. 1, pp. 29–
and Distributed Systems, vol. 20, no. 8, pp. 1075–1088, Aug. 2009. 41, Mar. 1997.
[34] J. Kim and A. Jamalipour, “Traffic management and QoS provisioning Gustavo de Veciana (S’88-M’94-SM’01-F’09) re-
in future wireless IP networks,” IEEE Personal Communications, vol. 8, ceived his B.S., M.S, and Ph.D. in electrical engi-
no. 5, pp. 46–55, Oct. 2001. neering from the University of California at Berkeley
[35] R. D. Callaway, M. Devetsikiotis, and C. Kan, “Design and imple- in 1987, 1990, and 1993 respectively, and joined the
mentation of measurement-based resource allocation schemes within the Department of Electrical and Computer Engineering
realtime traffic flow measurement architecture,” in Proc. of IEEE ICC, where he is currently a Cullen Trust Professor of
Jun. 2004. Engineering. His research focuses on the analysis
[36] ITU-R, “Report ITU-R M.2135-1, Guidelines for evaluation of radio and design of communication and computing net-
interface technologies for IMT-Advanced,” Technical Report, Dec 2009. works; data-driven decision-making in man-machine
[37] 3GPP, “Evolved Universal Terrestrial Radio Access (E-UTRA); Physical systems, and applied probability and queueing the-
layer procedures,” TS 36.213, v12.5.0, Rel. 12, Mar. 2015. ory. Dr. de Veciana served as editor and is currently
serving as editor-at-large for the IEEE/ACM Transactions on Networking. In
2009 he was designated IEEE Fellow for his contributions to the analysis
and design of communication networks. He currently serves on the board of
Pablo Caballero received his B.S. in telecom- trustees of IMDEA Networks Madrid.
munications and his M.S. in telematics engineer-
ing respectively from the University Carlos III of
Madrid in 2013 and 2015, respectively. In 2015, he
joined the Wireless Networking and Communica- Xavier Costa-Pérez (M’01) is Head of 5G Net-
tions Group at the University of Texas at Austin to works R&D at NEC Laboratories Europe, where
pursue his Ph.D. under the supervision of Profs. Gus- he manages several projects focused on 5G mo-
tavo de Veciana and Albert Banchs. Previously, bile core, backhaul/fronthaul and access networks.
Pablo worked as Research Assistant at IMDEA His team contributes to NEC projects for prod-
Networks Institute and as Research Intern at NEC ucts roadmap evolution, to European Commission
Laboratories Europe. His research interests lie in the R&D collaborative projects as well as to open-
design and performance evaluation of communication networks, game theory source projects and related standardization bodies,
and algorithm analysis. and has received several R&D Awards for successful
technology transfers. Dr. Costa-Pérez has served on
the Program Committees of several conferences and
holds multiple patents. He received both his M.Sc. and Ph.D. degrees in
Albert Banchs (M’04-SM’12) received his Telecommunications from the Polytechnic University of Catalonia (UPC-
M.Sc. and Ph.D. degrees from the Polytechnic BarcelonaTech) and was the recipient of a national award for his Ph.D. thesis.
University of Catalonia (UPC-BarcelonaTech) in
1997 and 2002, respectively. He is currently a Full
Professor with the University Carlos III of Madrid
(UC3M), and has a double affiliation as Deputy Arturo Azcorra (SM’02) received his M.Sc. degree
Director of the IMDEA Networks institute. Before in Telecommunications from UPM in 1986 and his
joining UC3M, he was at ICSI Berkeley in 1997, Ph.D. in 1989. In 1993, he obtained an MBA with
at Telefonica I+D in 1998, and at NEC Europe Ltd. honors. He is IEEE Senior Member and ACM SIG-
from 1998 to 2003. Prof. Banchs is Editor of IEEE COMM Member. He has been visiting researcher at
Transactions on Wireless Communications and MIT and UC Berkeley. He has participated in 49 re-
IEEE/ACM Transactions on Networking. His research interests include the search projects since the first Framework Programme
performance evaluation and algorithm design in wireless and wired networks. of the EU. Azcorra has been the project coordinator
of 5G-TRANSFORMER, 5G-Crosshaul, CARMEN,
CONTENT and E-NEXT. More information in: http:
//en.wikipedia.org/wiki/Arturo Azcorra

Network Slicing For Guaranteed Rate Services: Admission Control and Resource Allocation Games

Uploaded by

Copyright:

Available Formats

Network Slicing For Guaranteed Rate Services: Admission Control and Resource Allocation Games

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Network Slicing For Guaranteed Rate Services: Admission Control and Resource Allocation Games

Uploaded by

Copyright:

Available Formats

This article has been accepted for publication in a future issue of this journal, but has not been

Network Slicing for Guaranteed Rate Services:

by slice o, U o , and let ωbo (Ũ o ) be the minimum aggregate 0o −o

probabilities below 1%. Results confirm that NES can be

Proof of Theorem 4 Proof of Theorem 6

You might also like

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.