Competitive Caching of Contents in 5G Edge Cloud Networks

Competitive Caching of Contents
in 5G Edge Cloud Networks

Francesco De Pellegrini, Antonio Massaro, Leonardo Goratti, and Rachid El-Azouzi Fondazione Bruno Kessler, via Sommarive, 18 I-38123 Povo, Trento, Italy; CERI/LIA, University of Avignon, 339, Chemin des Meinajaries, Avignon, France. This research received funding from the European Union’s H2020 Research and Innovation Action under Grant Agreement No.671596 (SESAME project).

The surge of mobile data traffic forces network operators to cope with capacity shortage. The deployment of small cells in 5G networks is meant to reduce latency, backhaul traffic and increase radio access capacity. In this context, mobile edge computing technology will be used to manage dedicated cache space in the radio access network. Thus, mobile network operators will be able to provision OTT content providers with new caching services to enhance the quality of experience of their customers on the move.

In turn, the cache memory in the mobile edge network will become a shared resource. Hence, we study a competitive caching scheme where contents are stored at given price set by the mobile network operator.

We first formulate a resource allocation problem for a tagged content provider seeking to minimize the expected missed cache rate. The optimal caching policy is derived accounting for popularity and availability of contents, the spatial distribution of small cells, and the caching strategies of competing content providers. It is showed to induce a specific order on contents to be cached based on their popularity and availability.

Next, we study a game among content providers in the form of a generalized Kelly mechanism with bounded strategy sets and heterogeneous players. Existence and uniqueness of the Nash equilibrium are proved. Finally, extensive numerical results validate and characterize the performance of the model.

Mobile Edge Computing, Caching, Convex Optimization, Kelly Mechanism, Nash Equilibrium.

I Introduction

The recent boom of mobile data traffic is causing unprecedented stress over mobile networks. In fact, the global figures for such traffic reached 3.7 exabytes per month at the end of . They are ascribed mostly to over-the-top (OTT) video content providers (CP) such as Vimeo, YouTube and NetFlix. Forecasts predict that the world’s mobile data traffic will reach monthly exabytes by , of which will be video [1].

As a consequence, capacity shortage has become a real threat for mobile network operators (MNOs). Solutions involving the deployment of small cell (SC) base stations [2] have been receiving large consensus from industry and academia for next 5G systems. SCs are low power secondary base stations with limited coverage, to which user equipments (UEs) in radio range can connect, hence increasing spatial reuse and network capacity.

However, SCs are connected to a mobile operator’s core network via backhaul technologies such as, e.g., DSL, Ethernet or flexible millimeter-wave links. In order to avoid potential bottlenecks over the backhaul connection to SCs, mobile edge caching solutions have been devised. Actually, the primary goal of mobile edge caching is precisely to circumvent the limited backhaul connection of SCs [3] and ensure fast adaptation to radio link conditions.

From the network management standpoint, in order to handle a large number of SCs and associated memory caches, MNOs will rely on the emerging mobile edge computing (MEC) [4] 5G technology. MEC platforms are designed to enable services to run inside the mobile Radio Access Network (RAN) increasing proximity to mobile users, drastically reducing round trip time and thus improving the user experience.

Ultimately, CPs will be able to leverage on the MEC caching service offered by 5G MNOs. Contents can be replicated directly on lightweight server facilities embedded in the radio access network in proximity of SCs. In this context, the design of effective mobile edge caching policies requires to factor in popularity, number of contents, cache memory size as well as spatial density of small cells to which UEs may associate to. Indeed, due to storage limitations, allocation of contents on mobile edge caches has become an important optimization problem [5, 6, 7, 8, 9, 10, 11].

In this paper, we consider a scheme in which CPs can reserve mobile edge cache memory from a MNO. The MNO will provide a multi-tenant environment where contents can be stored at given price and will assign the available caching resources to different OTT content providers. In turn, this engenders competition of CPs for cache utilization.

First, we study the single CP optimization problem: under a given spatial distribution of SCs, the CP decides the optimal cache memory share to be reserved to different classes of contents. This permits to identify the minimum missed cache rate as a function of the purchased memory. Also, the optimal caching policy defines an order among contents jointly determined by two attributes: by the demand rate, i.e., the contents’ popularity, and by the concurrent effect of contents with similar popularity, i.e., the contents’ availability.

Finally, the competition among CPs is formulated using a new generalized Kelly mechanism with bounded strategy set. CPs trade off the cost for caching contents in the radio access network versus the expected missed cache rate. We show that the game admits a Nash equilibrium, and we prove that it is unique. Further properties of the game, including convergence and the revenue of the MNO, are investigated numerically.

The manuscript is organized as follows. In Sec. II we provide the related work and we outline the main contributions. Sec. III introduces the mathematical model developed throughout the paper. In Sec. IV the optimal content caching strategy is devised and in Sec. V we obtain the key characterization of the optimal missed cache rate. Sec. VI provides the analysis of the caching game. Numerical validation is performed in Sec. VII. Finally, Sec. VIII provides closing remarks.

Ii Related Works and main contribution

In [5] the authors consider a device-to-device (D2D) network and derive throughput scaling laws under cache coding and spatial reuse. Content delay is optimized in [12] by performing joint routing and caching, whereas in [6] a distributed matching scheme based on the deferred acceptance algorithm provides association of users and SC base stations based on latency figures. Similarly to our model, in [7] SC base stations are distributed according to a Poisson point process. Contents to be cached minimize a cost which depends on the expected number of missed cache hits.

In [8] a model for caching contents over a D2D network is proposed. A convex optimization problem is obtained and solved using a dual optimization algorithm. In our formulation we have obtained closed form solutions and properties of the optimal cost function.

In [9] a coded caching strategy is developed to optimize contents’ placement based on SC association patterns. In [13] a Stackelberg game is investigated to study a caching system consisting of a content provider and multiple network providers. In that model, the content providers lease their videos to the network providers to gain profit and network providers aim to save the backhaul costs by caching popular videos. In [10] the authors model a wireless content distribution system where contents are replicated at multiple access points – depending on popularity – so as to maximally create network-coding opportunities during delivery. Finally, [11] proposes proactive caching in order to take advantage of contents’ popularity. The scheme we develop in this work can also be applied to proactive caching.

Since content demand patterns are typically not known apriori, practical caching algorithms perform local content replacement policies  [14, 15]. Those rule how contents are replaced when the cache memory is full: heuristics including replacing least frequently used contents (LFU), last recently used contents (LRU) and several other variants have been proposed in literature. In our development we assume perfect knowledge of contents’ popularity, namely, the demand rates: recent results [15] show that by online estimation of the contents’ popularity, it is possible to achieve optimality, i.e., to minimize the missed cache rate. We leave the online estimation of the contents’ demand rates as part of future works.

Main results. The main contributions obtained in this work are the following:

  • a model is introduced which accounts for the contents’ characteristics, the spatial distribution of small cells, the price for cache memory reservation and the effect of competing content providers under multi-tenancy;

  • using such model, by convex optimization, the optimal caching policy is found to possess a waterfilling-type of structure which induces an ordering of contents depending on contents’ popularity and availability;

  • a competitive game is formulated where the price for cache memory reservation is fixed by the network provider. It is proved to be a new type of Kelly mechanism with bounded strategy set and it is showed to admit a unique Nash equilibrium.

To the best of the authors’ knowledge, this work is the first one to study mobile edge caching under a competitive scheme. This appears a crucial aspect in order to define new business models of 5G MNOs for the emerging MEC technology.

Symbol Meaning
number of content classes
intensity, i.e., spatial density of small-cells
set of content providers,
covering radius of UEs
storage capacity of a local edge cache unit (number of caching slots)
total storage capacity of the deployment
number of contents of class for content provider
popularity, i.e., demand rate for contents of class of content provider
availability, i.e.,
caching rate of content provider ,
total caching rate
total caching rate of competing content providers;
mobile network provider’s own caching rate
caching policy for content provider , ,
share of cache memory occupied by content provider
share of cache memory for -th class contents of content provider
maximum caching rate for content provider
price per caching slot for content provider
TABLE I: Main notation used throughout the paper

Iii System Model

Let us consider a MNO serving a set of content providers, where . Each CP serves his customers leveraging the MNO network.

Contents served to the customers of a tagged CP belong to different popularity classes, based on their demand rate or popularity . The -th popularity class thus features contents and content requests per day. Thus, we follow a multi-level popularity model similar to the one proposed in [10, 16]. In such model, files are divided into different popularity classes, and files within each class are equally popular.

We assume that each SC is attached to a local edge caching server, briefly cache. Multiple caches are aggregated by connecting them through the MNO backhaul and managed using a local MEC orchestrator, thus forming a seamless local edge cache unit as in Fig. 1. caching slots represent the available memory on such local edge cache unit; the total cache space across the whole deployment is hence where is the number of local edge cache units. For the sake of simplicity, each content is assumed to occupy one caching slot; since we assume , we rely on fluid approximations to describe the dynamics of cache occupation.

Fetching a non cached content from the remote CP server beyond the backhaul comes at unitary cost; such cost may represent the content’s access delay or the throughput to fetch the content from the remote server. Conversely, such cost is negligible if the user associates to a small cell storing a cached copy of the content. However, such cache should be reached by connecting to a SC within the UE radio range . SCs are distributed according to a spatial Poisson point process with intensity .

The following assumptions characterize the caching process:
i. each CP can purchase edge-caching service from the MNO and issue caching slot requests per day; we call the caching rate, where ;
ii. MNO will reserve caching slots per day for her own purposes;
iii. reserved slots expire after days for ;
iv. in order to attain caching slots per day, CP bids , and the MNO grants caching slots per day, where is such that . In our analysis we assume for the sake of simplicity111We refer to [17] for an in depth discussion of the connection between mechanisms and fair share of resources of the type studied in this paper..
v. CPs are charged based on the caching rate ;
vi. demand rates per content class are uniform across the MNO’s network.

The MNO will thus accommodate memory slots for CP according to


so that the whole cache memory occupation will be ruled by


where is the total caching rate. The corresponding dynamics for the fraction of reserved cache memory, assuming is

The MNO, in order to ensure full memory utilization, will choose such that . It follows from a simple calculation that in steady state, the fraction of the caching space for content provider is


Because contents’ requests are uniform across the MNO’s network, same fraction of cache space is occupied by CP in each local edge cache unit.

In particular, CP will split his reserved memory among content classes according to a proportional share allocation with weighting coefficients , , where . We define the caching policy of CP .

Then, the fraction of local edge cache memory occupied by contents of class from content provider is


Finally, a tagged content of class of content provider is found in the memory of a local edge cache with probability . In the rest of the paper, we will assume for the sake of simplicity.

Now, we want to quantify the probability for a given requested content not to be found in the local edge cache memory, i.e., the missed cache probability.

Under the Poisson assumption, the probability for a tagged UE not to find any SC within a distance is . Applying a thinning argument, the probability not to find a content of class of CP within distance is .

The expected missed cache rate (MCR) is thus


It depends on caching rate and on caching policy . Also, accounts for the fact that other content providers share the same cache space. In the next section we shall describe the optimal caching policy attained when CP aims at minimizing (5), for a fixed value of the caching rate.

Fig. 1: Local edge cache unit providing memory slots.

Iv Optimal Caching Policy

In order to analyse the model introduced before, we need to characterize the CPs’ response to competitors’ actions, i.e., . Hence, we assume that each CP aims at minimizing his own MCR, and that the network provider guarantees full information to CPs, i.e., storage capacity, spatial density of SCs and total caching rate. We defer the study of the system under partial information at the content provider’s side to later works.

We hence consider the following resources allocation problem for the single CP:

Definition 1 (Optimal Caching Policy).

Given opponents’ strategy profile the optimal caching policy of is the solution of


subject to the following constraints:


It is immediate to observe that is a strictly convex function in the single content provider control , so that a unique solution exists [18]. In order to solve for the constrained minimization problem in equations (6) and (7) we can write the Lagrangian for player as follows

For notation’s sake, we have defined ; we define this quantity availability. Furthermore, since constraints are affine, the Karush Kuhn Tucker (KKT) conditions provide the solution of the original problem [18].

Hereafter, we enlist the KKT conditions:

Using a standard argument [18], by complementary slackness, implies ; let us define index set .

Iv-a Popularity sorted case

Let us discuss a simplified setting where popularity is the main driver for the CPs. Let us first assume that the indexes are sorted according to contents’ popularity , i.e., . We also assume : more popular contents are also less abundant. This assumption will be relaxed in the next section, where we derive the general solution; we will see that there exists a natural order combining popularity and availability of contents which determines whether a content class is cached or not.

In the case at hand, the structure of the optimal allocation follows from

Lemma 1.

Let and for , then implies

From the previous statement we can deduce the following

Corollary 1.

Under the assumptions of Lemma. 1 there exists such that response for and otherwise

The stationarity conditions can be used in order to determine the optimal content allocation in closed form. Let , then , so that


Finally, due to the constraint saturation


From Corol. 1, the optimal solution corresponds to the maximal such that the corresponding solving (8) lies in .

We hence observe that when for all , i.e., when availability is same for all classes, the optimal caching policy depends on contents’ popularity only. The smaller the request rate , i.e., the less popular the content class, the smaller the cache share reserved to contents of that class. Such optimal policy is analogous to the optimal content replacement MIN algorithm [19]. In fact, MIN, assumes full information about the future, replaces first contents which will be requested farthest in the future.

Iv-B General solution

The solution to the KKT conditions can be formulated as a waterfilling-like solution [18]. In fact, from stationarity conditions, writes as

which can be specialized into the following two cases.

Case i: . In this case for any . Hence, by complementary slackness, .

Case ii: . It is always possible to find satisfying the stationarity condition and a that satisfies the complementary slackness condition: just set and

Finally, let . For notation’s sake, the solution writes


It is immediate to recognize a waterfilling solution in logarithmic scale. Let . Indeed is strictly increasing in . Also, for , and . Thus, there exists a unique positive satisfying our problem.

Actually, the solution is determined in polynomial time : let be the permutation of the indexes which sorts in increasing order, i.e., . For every choice , one can determine a value of

for . Then, consider the only , compatible with (9). We observe that is equivalent to state that : clearly, if , indeed , so that we can generalize Prop. 1 as follows

Corollary 2 (Threshold structure).

There exists such that for and otherwise.

Remark 1 (Contents’ Order).

The existence of a threshold structure in a waterfilling-type of solution is not surprising; what we learn instead is that the natural order which determines which content classes are cached or not is given by the values . Hence, the index sorting which orders the content classes with decreasing is the order by which a content provider prioritizes content classes to be cached as the cache memory available increases.

In the rest of the paper we assume content classes sorted according to .

V Optimal Missed Cache Rate

CPs who optimize contents to be cached, for a given value of , minimize the expected MCR in the caching policy . In the game model presented in the next section we shall leverage on the convexity properties of the optimized MCR , defined as


where . As already proved, the minimum in (10) is unique, hence is well defined. Hereafter we thus demonstrate its convexity in .

Actually, convexity can be derived for a class of functions wider than the posynomial expression appearing in (10). We first need the following fact, whose proof is found in the Appendix.

Lemma 2.

Let be non increasing, with domain . Let be convex. Then is convex on .

We can now derive the general conditions for the convexity of the optimal missed cache rate

Theorem 1.

Let , convex and decreasing in each variable for , then

is convex and decreasing in .


In order to prove convexity for , we consider perspective function : is known to be convex if is convex [18, pp.89]. In the next step, let , and consider the function

which is convex since it is obtained by minimizing over the simplex which is a convex set. Now using Lemma 1, we conclude that is convex in the first variable, and by affinity so does .

In order to prove the monotonicity of in , let us consider and for some and the respective optimal caching policy and . We write

where the first inequality follows from monotonicity and the second from optimality. ∎

The case in (10) satisfies the assumptions by letting .

For presentation’s sake, in Sec. VI we shall identify . There, we also need the following result, whose proof is found in the Appendix.

Lemma 3 (Limit solution for ).

There exists such that, for any , and the optimal MCR is


V-a The case

For two classes of contents, , the expression for can be derived in simple closed form. This sample case retains the main properties of the optimal policy and it is useful in order to provide insight into the structure of the optimal MCR. First, we write the expression of the optimal MCR


For the sake of notation, we denote . The (unconstrained) minimum of the right hand term is attained at


When , the utility function of is

where the constant appearing on the first term is


Incidentally, the convexity of for can be verified directly from the convexity of and by composition with an affine function, which preserves convexity.

We are interested in characterizing precisely the behavior of the expected MCR as a function of . In particular, we want to assess the influence of the system parameters.

Now, we can obtain the following result

Proposition 1.

i. Assume . Let , and define threshold for content


then it holds


where the corresponding optimal cache policy is in the first case, in the second case and constant as in (14)
ii. Let , then case holds for any with associated expected MCR defined as in case i.
iii. If , both i. and ii. hold with role of content and reversed.

The proof follows by inspection of (13) considering as an unconstrained minimizer. First, we observe that if , then , i.e., the first content class is always cached. The other conditions follow by imposing .

Discussion: availability, popularity and competition

Hereafter we draw insight from Prop. 1. First, as seen there, the optimal caching rate depends solely on a few system parameters, namely and for . Actually, when then : contents of type are always cached because . The fact that contents of type are cached depends on the sign of , which in turn determines the actual structure of the waterfilling solution.

From Prop. 1, means that contents of type are either less popular () and/or less available () than contents of type . The availability of contents of type determines whether they will be eventually cached. In practice, when , there exists a critical value of the CP caching rate , i.e., the threshold (15). Above such value, contents of type are cached, below that they are not cached. For the sake of consistency, in the case when , while for , .

Furthermore, increases linearly with both the MNO caching rate and the competitors’ aggregate caching rate : competition for edge caching resources tends to prevent caching of contents with smaller product . Actually, under higher competition figures, optimal caching policies are of the type , and . It is interesting to observe that, as detailed in case ii., not always there exists a caching rate such that it is worth caching the least profitable content class.

Fig. 2: Case : (a) Increasing value of as a function of , for and (b) Region of switch on of content .

We have provided a pictorial representation of the results of this section in Fig. 2 for the case . In Fig. 2 the value of has been fixed at different values and the corresponding behavior of the threshold value has been reported as a function of . For , it holds since there is no switch-on value of for class . Fig. 2 represents the region where the switch-on of the less popular content is possible as it can be derived from the expression (15) as a function of and .

Vi Game Model for Content Providers

So far the caching rate has been input for the CPs in order to decide how to optimize the caching policy . Let MNO propose to CPs costs per caching slot. CP strategy in turn is the number of caching slots he reserves per day, with convex and compact strategy set . The best response of CP depends on his contents, and his opponents’ strategies. It is the minimizer of the cost function : it solves


Here and opponents’ strategy profile writes .

The appearing in (17) is a general caching policy and we shall consider two cases.

Caching Rate Optimizers. In this case, the best response of content providers is decided for a fixed caching policy . I.e., each content provider decides beforehand the caching policy for any given caching rate . Let : it is convex and decreasing and . Hence, if all players are caching rate optimizers, the game is a variant of the Kelly mechanism [20]. The basic Kelly mechanism allocates a divisible resource among players proportionally to the players’ bids, in our case the equivalent required caching rates. Here, compared to the standard formulations in literature [21, 22, 20, 23] our formulation combines three specific features which render it non standard:

  • bounded compact and convex strategy set;

  • is equivalent to a bidding reservation, as described in [23];

  • prices may depend on the player, i.e., the game is a generalized Kelly mechanism [20]

We denote the Kelly mechanism in the form outlined above a generalized Kelly mechanism with reservation and bounded strategy set.

Simultaneous Optimizers. In this case . When players are simultaneous optimizers, the structure of the game still resembles the Kelly mechanism [22]. For , the game corresponds to the case of caching rate optimizers. For , the fact that the game is actually a Kelly mechanism is proved formally in the following

Lemma 4 (Kelly form for Simultaneous Optimizers).

If players are simultaneous optimizers, the game (17) is a generalized Kelly mechanism with reservation and bounded strategy set.

The proof of the above result is found in the Appendix. Here, it is sufficient to observe that even in the case of a simultaneous optimizer CP , the optimal MCR can be expressed as where is convex and continuously differentiable in .

Vi-a Existence and uniqueness of the Nash Equilibrium

In the general case, the game may comprise a mixture of both CPs who are caching rate optimizers and who are simultaneous optimizers. From the above discussion, the game is still a generalized Kelly mechanism with reservation and bounded strategy set.

In order to characterize the possible equilibria, we describe first the best response of each player

Lemma 5 (Best response).

Given the opponent CPs’ strategy profile :
i. It holds if and only if where

ii. Let , then , where .

The above statement follows from the fact that the objective function in (17) is convex and thus has a unique minimum in . The expression of in the case of simultaneous optimizers is derived from the expression (11) reported in Lemma 1.

The zero and the saturated Nash equilibria are easily characterized in the following

Proposition 1 (Trivial Nash Equilibria).

i. is the unique Nash equilibrium iff if is a simultaneous optimizer and if is a caching rate optimizer.
ii. is the unique Nash equilibrium if and only if it holds for all .

We observe that in the original Kelly mechanism, the strategy vector is never a Nash equilibrium [21, 22].

In our case, it may be the Nash equilibrium and this is the effect of the term due the MNO’s usage of the cache. In fact, the physical interpretation is provided by the condition i. in Prop. 1. No CP has incentive to start caching at give price when the marginal revenue for starting caching, i.e., represented by the product of demand and availability, does not exceed the value of the cache share reserved to the MNO operations, the term . Conversely, at low prices a saturated Nash equilibrium is expected.

In the general case, the presence of a bounded strategy set requires a specific proof for the uniqueness of the Nash equilibrium, as seen in the following.

Theorem 2 (Existence and Uniqueness).

The game has a Nash equilibrium and it is unique.

We describe a brief outline of the full proof of the above result which is found in the Appendix. In order to prove the existence of Nash equilibria of the game, it is sufficient to observe that:

  • the multistrategy set is a convex compact subset of ;

  • is convex conditionally to the opponents strategy, both for simultaneous optimizers and caching rate optimizers;

Hence, the existence of Nash equilibria is a direct consequence of the result of Rosen [24], originally formulated for –persons concave games. With respect to the uniqueness, and are always unique from Prop. 1. Once we excluded those trivial cases, the uniqueness can be derived by extending an argument [23] to the case of a bounded strategy set. Such proof applies to both the case of simultaneous optimizers and of caching-rate optimizers. However, it requires cost functions to be continuously differentiable in , which is not straightforward for simultaneous optimizers.

Finally, the proof of uniqueness applies also to the context where part of the players are simultaneous optimizers and the others are caching-rate optimizers.

We further observe that from the proof of Thm. 2 we can derive a simple bisection algorithm to calculate the unique solution of the game. It will be used in the numerical section where we shall provide further characterization of the game via quantitative measures, There, we are describing the pricing operated by the MNO and the convergence to the Nash equilibrium when CPs are myopic cost minimizers.

Vii Numerical Results





Fig. 3: (a) Milan downtown base stations deployment, detail of the area considered; (b) CP optimal cost: theoretical prediction (c) CP optimal cost: outcome of the simulation (d) CP optimal caching policy for varying and fixed value of m. Settings are: , , , , , , , , , , .





Fig. 4: (a) Cost function for simulataneous optimizer CP as varys, all parameters are the same as in Fig 3; (b) Dynamic of the cost function for a 3-simoultaneous optimizers game. Settings are: , , , , , (c) detail of the corresponding restpoint; (d) revenue of the MNO for increasing uniform price. Parameters are: , , , , , , , , m.

In this section we provide numerical description and validation of the model.222Both the Python scripts and the dataset used for validation can be downloaded at First we validate the models’ assumptions against a real world scenario. Then, we focus on the single player’s actions, having fixed the remaining players’ strategies. Finally we provide numerical characterization of the game introduced in the previous section.
Point Process. The model introduced in Sec. III assumes that SCs are distributed according to a spatial Poisson process of given intensity . Hence, we have tested the performance of the optimal caching policy in the case the SCs spatial deployment does not adhere to the assumption of a Poisson point distribution. In order to do so, we have been comparing the theoretical results with the outcome of a simulation performed over a real dataset. The real dataset (source is the sample distribution of the cell towers deployed in downtown Milan over a Kms area, as depicted in Fig. 3: it includes the location of cell towers corresponding to base stations per square Km. The distribution of base stations in a very densely populated urban area has been used as a reasonable approximation for a SC deployment.

The sample spatial density has been used in the model in order to evaluate, under the same spatial density of SCs, the theoretical CP’s cost function for increasing values of the covering radius m in the following cases (see Fig. 3): a) the CP performs a uniformly random caching policy , for constant caching rate b) the CP performs a popularity-based caching policy, i.e., , for constant c) the CP is a caching rate optimizer adopting a popularity based caching policy d) the CP is a simultaneous optimizer.

The results in Fig. 3 refer to a simulation encompassing the same strategies under the sample point distribution of Fig. 3. The simulation has been performed by repeatedly selecting a random UE position in the playground, and measuring the sampling frequency of missed cache events upon requesting contents from SCs within the UE’s radio range.

By comparing the results in Fig. 3 and Fig. 3, we observe that the Poisson distribution – as expected due to the non-uniform spatial density of the sample real-world deployment – tends to slightly underestimate the cost incurred by CPs. However, the theoretical and the simulated results are very close and the relative performance of the caching policies match the prediction of the theoretical model. This result confirms that the proposed model performs well even in real world scenarios: under a non-Poisson point process for the SC spatial distribution a rational optimizing player would choose the proposed optimal strategy over other possible strategies.

Cost function. In the next experiment we describe the optimal caching policy (Fig. 3) and the cost function (Fig. 4) in the case . In particular, Fig. 3 reports on the characteristic waterfilling structure of the optimal caching as the parameter increases. As predicted by the model, the water-filling solution has a threshold structure. The value of determines the content classes that become active: for large all content classes are cached, whereas for small values only some are cached. In Fig. 4 we have reported the typical convex shape of the cost function corresponding to the same setting and for increasing values of . It is worth noting how the actions of opponents, reflected in the value of , affect the shape of ’s cost function.

Convergence to the Nash equilibrium. In Fig. 4 and Fig. 4 we have simulated CPs who are simultaneous optimizers. They behave as myopic players: each one of them, chosen at random, optimizes his own cost function based on the opponents’ profile. Numerical simulations show that, after a small number of iterations the game stabilizes on the same restpoint irrespective of initial strategies. As depicted in Fig. 4 the restpoint is indeed a minimum for each CP’s cost function, i.e., it is the Nash equilibrium of the game. This behavior suggests that the game has the finite improvement property [25], even though we could not identify analytically a potential for the game. Hence, the system would naturally converge to his unique Nash equilibrium if each player optimizes independently its own cost function against the opponents.

Finally, we have drawn in Fig 4 the daily revenue of the MNO at the Nash equilibrium as a function of the caching price , uniform for all CPs. Because the MNO’s total revenue depends on the Nash equilibrium, she could try to optimize her revenue by leveraging the CPs’ cost structure. We observe numerically that the total revenue appears to have a unique maximum at a certain maximizer price . This suggests the existence of a unique Stackelberg equilibrium for the proposed scheme. This provides the possibility to compute the global restpoint of the system when both CPs and MNO behave strategically.

Viii Conclusions

A model for mobile edge caching in 5G networks has been presented. OTT content providers compete for the cache memory made available by a MNO at given price. Several features of the system are captured, including popularity and availability of contents, spatial distribution of small cells, competition for cache memory and the effect of price. CPs can optimize the allocation of contents in order to reduce customers’ aggregated missed cache rate. We have found that the optimal caching policy is of waterfilling type. Also, it is showed to give priority to contents based on popularity and availability. We have confirmed the validity of the caching policy optimization on real-world traces.

Finally, the competition for the shared caching memory can be formulated as a convex –persons game. This game is a new form of the Kelly mechanism with bounded strategy set, where each CP trades off the expected missed cache rate for the price paid to the MNO in order to reserve cache memory space. The existence and uniqueness properties of the Nash equilibrium are demonstrated. Also, numerical results indicate that when CPs are myopic optimizers, the system converges to a unique restpoint which is the Nash equilibrium.

Furthermore, from numerical results, this game appears to have a unique Stackelberg equilibrium, a relevant feature for the MNO in order to maximize her revenue at the optimal price. To this respect, an interesting research direction is to develop online algorithms by which the MNO can learn over time such optimal price.


  • [1] “Cisco visual networking index: Global mobile data traffic forecast update, 2015–2020 white paper,” CISCO, White Paper, February 3 2016.
  • [2] J. Andrews, S. Buzzi, W. Choi, S. Hanly, A. Lozano, A. Soong, and J. Zhang, “What will 5G be?” IEEE JSAC, vol. 32, no. 6, pp. 1065–1082, June 2014.
  • [3] X. Wang, M. Chen, T. Taleb, A. Ksentini, and V. Leung, “Cache in the air: exploiting content caching and delivery techniques for 5G systems,” IEEE Comm. Mag., vol. 52, no. 2, pp. 131–139, February 2014.
  • [4] MEC ETSI Industry Specification Group, “ETSI DGS/MEC-IEG004: Mobile-Edge Computing (MEC) – Service Scenarios,” Available online:
  • [5] M. Ji, G. Caire, and A. F. Molisch, “Fundamental limits of distributed caching in D2D wireless networks,” in Proc. of IEEE ITW, Sevilla, Spain, Sept. 9-13 2013, pp. 1–5.
  • [6] F. Pantisano, M. Bennis, W. Saad, and M. Debbah, “Cache-aware user association in backhaul-constrained small cell networks,” in Proc. of IEEE WiOPT, Hammamet, Tunisia, May 12-16 2014, pp. 37–42.
  • [7] B. N. Bharath, K. G. Nagananda, and H. V. Poor, “A learning-based approach to caching in heterogenous small cell networks,” IEEE Transactions on Communications, vol. 64, no. 4, pp. 1674–1686, April 2016.
  • [8] H. J. Kang and C. G. Kang, “Mobile device-to-device (D2D) content delivery networking: A design and optimization framework,” Journal of Communications and Networks, vol. 16, no. 5, pp. 568–577, Oct 2014.
  • [9] A. Sengupta, S. Amuru, R. Tandon, R. Buehrer, and T. Clancy, “Learning distributed caching strategies in small cell networks,” in Proc. of IEEE ISWCS, Barcelona, Spain, Aug. 26-29 2014, pp. 917–921.
  • [10] J. Hachem, N. Karamchandani, and S. Diggavi, “Multi-level coded caching,” in Proc. of IEEE INFOCOM, Hong-Kong, RPC, April 26th - June 1st 2015, pp. 756 – 764.
  • [11] E. Bastug, M. Bennis, and M. Debbah, “Living on the edge: The role of proactive caching in 5G wireless networks,” IEEE Communications Magazine, vol. 52, no. 8, pp. 82–89, Aug 2014.
  • [12] M. Dehghan, A. Seetharam, B. Jiang, T. He, T. Salonidis, J. Kurose, D. Towsley, and R. K. Sitaraman, “On the complexity of optimal routing and content caching in heterogeneous networks,” in Proc. of IEEE INFOCOM, Hong-Kong, RPC, April 26 - May 1 2015, pp. 936–944.
  • [13] J. Li, W. Chen, M. Xiao, F. Shu, and X. Liu, “Efficient video pricing and caching in heterogeneous networks,” IEEE Transactions on Vehicular Technology, vol. PP, no. 99, pp. 1–1, 2015.
  • [14] W. Ma and D. H. Du, “Design a progressive video caching policy for video proxy servers,” IEEE Trans. Multimedia, vol. 6, no. 4, pp. 599–610, 2004.
  • [15] S. Li, J. Xu, M. van der Schaar, and W. Li, “Popularity-driven content caching,” in Proc. of IEEE INFOCOM, San Francisco, CA, 10-15 April 2016.
  • [16] J. Hachem, N. Karamchandani, and S. Diggavi, “Effect of number of users in multi-level coded caching,” in IEEE ISIT, Hong-Kong, RPC, June 14 - 19 2015, pp. 1701–1705.
  • [17] T. Baçar and G. Olsder, Dynamic Noncooperative Game Theory, 2nd ed.   Society for Industrial and Applied Mathematics, 1998.
  • [18] S. Boyd and L. Vandenberghe, Convex Optimization.   New York, NY, USA: Cambridge University Press, 2004.
  • [19] L. A. Belady, “A study of replacement algorithms for a virtual-storage computer,” IBM Syst. J., vol. 5, no. 2, pp. 78–101, Jun. 1966.
  • [20] Y. Yang, R. T. B. Ma, and J. C. S. Lui, “Price differentiation and control in the Kelly mechanism,” Elsevier Perf. Evaluation, vol. 70, no. 10, pp. 792–805, October 2013.
  • [21] A. Reiffers-Masson, Y. Hayel, and E. Altman, “Game theory approach for modeling competition over visibility on social networks,” in Proc. of IEEE COMSNETS, Bangalore, India, Jan 7-10 2014, pp. 1–6.
  • [22] R. Johari, “Efficiency loss in market mechanisms for resource allocation,” Ph.D. dissertation, Department of Electrical Engineering and Computer Science, 2004.
  • [23] R. Maheswaran and T. Baçar, “Efficient signal proportional allocation (espa) mechanisms: Decentralized social welfare maximization for divisible resources,” IEEE J.Sel. A. Commun., vol. 24, no. 5, pp. 1000–1009, Sep. 2006.
  • [24] J. B. Rosen, “Existence and uniqueness of equilibrium points for concave N-person games,” Econometrica, vol. 33, no. 3, July 1965.
  • [25] D. Monderer and L. S. Shapley, “Potential games,” Games and Economic Behavior, vol. 14, no. 1, pp. 124 – 143, 1996.

Appendix A Appendix

A-a Proof of Lemma 1


Let the optimal allocation, and let us assume that It is sufficient to write the generic cost as

from which it is immediate to see that a response identical to but where and is better off. In fact we can write