302 IEEE TRANSACTIONS ON CLOUD COMPUTING, VOL. 9, NO.

1, JANUARY-MARCH 2021

Price-Based Resource Allocation for Edge

Computing: A Market Equilibrium Approach
Duong Tung Nguyen , Student Member, IEEE, Long Bao Le ,
Senior Member, IEEE, and Vijay Bhargava, Life Fellow, IEEE

Abstract—The emerging edge computing paradigm promises to deliver superior user experience and enable a wide range of Internet
of Things (IoT) applications. In this paper, we propose a new market-based framework for efficiently allocating resources of
heterogeneous capacity-limited edge nodes (EN) to multiple competing services at the network edge. By properly pricing the
geographically distributed ENs, the proposed framework generates a market equilibrium (ME) solution that not only maximizes the
edge computing resource utilization but also allocates optimal resource bundles to the services given their budget constraints. When
the utility of a service is defined as the maximum revenue that the service can achieve from its resource allotment, the equilibrium can
be computed centrally by solving the Eisenberg-Gale (EG) convex program. We further show that the equilibrium allocation is Pareto-
optimal and satisfies desired fairness properties including sharing incentive, proportionality, and envy-freeness. Also, two distributed
algorithms, which efficiently converge to an ME, are introduced. When each service aims to maximize its net profit (i.e., revenue minus
cost) instead of the revenue, we derive a novel convex optimization problem and rigorously prove that its solution is exactly an ME.
Extensive numerical results are presented to validate the effectiveness of the proposed techniques.

Index Terms—Market equilibrium, Fisher market, fairness, algorithmic game theory, edge computing, fog computing

1 INTRODUCTION

T HE last decade has witnessed an explosion of data traffic

over the communication network attributed to the rap-
idly growing cloud computing and pervasive mobile devi-
latency, reliability, security, mobility, and localization of
new systems and applications (e.g., embedded artificial
intelligence, mission-critical communication, 5G wireless
ces. This trend is expected to continue for the foreseeable systems) [1]. To this end, edge computing (EC) [2], also
future with a whole new generation of applications includ- known as fog computing (FC) [1], has emerged as a
ing 4K/8K UHD video, tactile Internet, virtual/augmented novel computing paradigm that complements the cloud
reality (VR/AR), and a variety of IoT applications [1]. As and addresses many shortcomings in the traditional
the cloud infrastructure and number of devices continue to cloud model.
expand at an accelerated rate, a tremendous burden will be In EC, storage, computing, control, and networking
put on the network. Hence, it is imperative for operators to resources are placed closer to end-users, things, and sen-
develop innovative solutions to meet the soaring traffic sors. The size of an EN is flexible ranging from smart-
demand and accommodate diverse requirements of various phones, smart access points (AP), base stations (BS) to edge
services and use cases in future networks. clouds [3]. For example, a smartphone is the edge between
Thanks to the economy of scale and supercomputing wearable devices and the cloud, a home gateway is the edge
capability advantages, cloud computing will likely con- between smart appliances and the cloud, a telecom central
tinue to play a prominent role in the future computing office is the edge between mobile devices and the core net-
landscape. However, cloud data centers (DC) are often work. By providing elastic resources and intelligence at the
geographically distant from the end-user, which induces edge, EC offers many remarkable capabilities, such as local
enormous network traffic, along with significant commu- data processing and analytics, distributed caching, location
nication delay and jitter. Therefore, despite the immense awareness, resource pooling and scaling, enhanced privacy
power, cloud computing alone is facing growing limita- and security, and reliable connectivity. EC is also a key
tions in satisfying the stringent requirements in terms of enabler for ultra-reliable low-latency applications (e.g., AR,
autonomous driving). A myriad of benefits and other use
cases (e.g., offloading, caching, advertising, healthcare,

D.T. Nguyen and V. Bhargava are with the Department of Electrical and
smart homes/grids/cities) of EC can be found in [1], [2], [3].
Computer Engineering, University of British Columbia, Vancouver, BC
V6T 1Z4, Canada. E-mail: {duongnt, vijayb}@ece.ubc.ca. Today, EC is still in the developing stages and presents

L.B. Le is with INRS-EMT, Universit
e du Qu
ebec, Montr
eal, Qu
ebec H5A many new challenges, such as network architecture design,
1K6, Canada. E-mail: [email protected]. programming models and abstracts, IoT support, service
Manuscript received 29 July 2017; revised 10 May 2018; accepted 1 June 2018. placement, resource provisioning and management, security
Date of publication 6 June 2018; date of current version 5 Mar. 2021. and privacy, incentive design, and reliability and scalability
(Corresponding author: Duong Tung Nguyen.)
Recommended for acceptance by G. Min. of edge devices [1], [2], [3]. In this paper, we focus on the EC
Digital Object Identifier no. 10.1109/TCC.2018.2844379 resource allocation problem. Unlike cloud computing, where
2168-7161 © 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See ht_tps://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY KURUKSHETRA. Downloaded on April 25,2022 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 NGUYEN ET AL.: PRICE-BASED RESOURCE ALLOCATION FOR EDGE COMPUTING: A MARKET EQUILIBRIUM APPROACH 303

computational capacity of large DCs is virtually unlimited Similarly, an application provider (e.g., Uber, Pokemon
and network delay is high, EC is characterized by relatively Go) or a sensor network may own a number of ENs in a city
low network latency but considerable processing delay due and need to allocate the edge resources to handle requests
to the limited computing power of ENs. Also, there are a of different groups of users/sensors. The budget can be
massive number of distributed computing nodes compared decided based on criteria such as the populations of users/
to a small number of large DCs. Additionally, ENs may sensors in different areas and/or payment levels (subscrip-
come with different sizes (e.g., number of computing units) tion fees) of different groups of users. Another example is
and configurations (e.g., computing speed) ranging from a that a university (or other organizations) can grant different
smartphone to an edge cloud with tens/hundreds of servers. virtual budgets to different departments or research labs so
These nodes are dispersed in numerous locations with vary- that they can fairly share the edge servers on the campus.
ing network and service delay towards end-users. The first model may also emerge in the setting of cloud fed-
On the other hand, different services may have different eration at the edge where several companies (i.e., services)
requirements and properties. Some services can only be pool their resources together and each of them contributes a
handled by ENs satisfying certain criteria. Furthermore, dif- fixed portion of resource of every EN. Here, the budgets are
ferent services may be given different priorities. While proportional to the initial contributions of the companies.
every service not only wants to obtain as much resource as Instead of resource pooling, these companies may agree
possible but also prefers to be served by its closest ENs with upfront on their individual budgets, and then buy/rent a
low response time, the capacities of ENs are limited. Also, given set of ENs together.
due to the diverse preferences of the services towards the In these scenarios, it is important to consider both fair-
ENs, some nodes can be under-demanded while other are ness and efficiency. Thus, conventional schemes such as
over-demanded. Thus, a fundamental problem is: given a set social welfare maximization, maxmin fairness, and auction
of geographically distributed heterogeneous ENs, how can we effi- models may not be suitable. In particular, a welfare maximi-
ciently allocate their limited computing resources to competing zation allocation often gives most of the resources to users
services with different desires and characteristics, considering ser- who have high marginal utilities while users with low mar-
vice priority and fairness? This work introduces a novel mar- ginal utilities receive a very small amount of resources,
ket-based solution framework which aims not only to even nothing. Similarly, in auction models, the set of losers
maximize the resource utilization of the ENs but also to are not allocated any resource. Hence, these solutions can
make every service happy with the allocation decision. be unfair to some users. On the other hands, a maxmin fair-
The basic idea behind our approach is to assign different ness solution often allocates too many resources to users
prices to resources of different ENs. In particular, highly with low marginal utilities, hence, it may not be efficient.
sought-after resources are priced high while prices of under- To strive the balance between fairness and efficiency, we
demanded resources are low. We assume that each service advocate the General Equilibrium Theory [4], with a specific
has a certain budget for resource procurement. The budget focus on the Fisher market model [5], as an effective solu-
can be virtual or real money. Indeed, budget is used to cap- tion concept for this problem. Specifically, the first model
ture service priority/differentiation. It can also be inter- can be cast as a Fisher market in which services act as
preted as the market power of each service. Given the buyers and ENs act as different goods in the market. For the
resource prices, each service buys the favorite resource bun- linear additive utility function as considered in this work,
dle that it can afford. When all the resources are fully allo- given resource prices, a service may have an infinite set of
cated, the resulting prices and allocation form a market optimal resource bundles, which renders difficulty in
equilibrium (ME). If there is only one EN, an ME can be found designing distributed algorithms. We suggest several meth-
easily by adjusting the price gradually until demand equals ods to overcome this challenge. Moreover, we show that the
supply or locating the intersection of the demand and supply obtained allocation is Pareto-optimal, which means there is
curves. However, when there are multiple heterogeneous no other allocation that would make some service better off
ENs and multiple services with diverse objectives and differ- without making someone else worse off [6]. In other words,
ent buying power, the problem becomes challenging since there is no strictly “better” allocation. Thus, a Pareto-opti-
the services have more options to buy resources. We consider mal allocation is efficient.
two distinct market models in this work. We furthermore link the ME to the fair division literature
In the first model, the money does not have intrinsic [7] and prove that the allocation satisfies remarkable fair-
value to the services. Given resource prices, each service ness properties including envy-freeness, sharing-incentive,
aims to maximize its revenue from the allocated resources, and proportionality, which provides strong incentives for
without caring about how much it has to pay as long as the the services to participate in the proposed scheme. Indeed,
total payment does not exceed its budget. This model arises these properties were rarely investigated explicitly in the
in many real-world scenarios. For example, in 5G networks, ME literature. Envy-freeness means that every service prefers
the Mobile Edge Computing (MEC) servers of a Telco are its allocation to the allocation of any other service. In an
shared among different network slices, each of which runs a envy-free allocation, every service feels that its share is at
separate service (e.g., voice, video streaming, AR/VR, con- least as good as the share of any other service, and thus no
nected vehicles, sensing) and serves a group of customers service feels envy. Sharing-incentive is another well-known
who pay for the service. The Telco can allot different budg- fairness concept. It ensures that services get better utilities
ets to the slices depending on their importance and/or than what they would get in the proportional sharing scheme
potential revenue generation (e.g., the total fee paid by the that gives each service an amount of resource from every
users/subscribers of each slice). EN proportional to its budget. Note that proportional

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY KURUKSHETRA. Downloaded on April 25,2022 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 304 IEEE TRANSACTIONS ON CLOUD COMPUTING, VOL. 9, NO. 1, JANUARY-MARCH 2021

sharing is an intuitive way to share resources fairly in terms
Centralized solution. The unique ME in the first model
of quantity. For the federation setting, sharing-incentive can be determined by the EG program. We also
implies that every service gets better off by pooling their prove some salient fairness features of the ME.
resources (or money) together. Finally, it is natural for a ser-
Decentralized algorithms. We introduce several dis-
vice to expect to obtain a utility of at least b=B of the maxi- tributed algorithms that efficiently overcome the dif-
mum utility that it can achieve by getting all the resources, ficulty raised by the non-unique demand functions
where b is the payment of the service and B is the total pay- of the services and converge to the ME.
ment of all the services. The proportionality property guaran-
Extended Fisher market. We systematically derive a
tees that the utility of every service at the ME is at least new convex optimization problem whose optimal
proportional to its payment/budget. Thus, it makes every solution is an exact ME in the extended Fisher mar-
service feel fair in terms of the achieved utility. ket model where buyers value the money.
In the second model, the money does have intrinsic value
Performance Evaluation. Simulations are conducted to
to the services. The services not only want to maximize their illustrate the efficacy of the proposed techniques.
revenues but also want to minimize their payments. In par- The rest of the paper is organized as follows. Section 2
ticular, each service aims to maximize the sum of its remain- describes related work. The system model and problem for-
ing budget (i.e., surplus) and the revenue from the procured mulation are given in Sections 3 and 4, respectively. The
resources, which is equivalent to maximizing the net profit centralized solution using the EG program is analyzed in
(i.e., revenue minus cost). This model is prevalent in prac- Section 5. Then, we introduce several distributed algorithms
tice. For example, several service providers (SP), each of in Section 6. The market model in which buyers aim to max-
which has a certain budget, may compete for the available imize their net profits is studied in Section 7. Simulation
resources of an edge infrastructure provider (e.g., a Telco, a results are shown in Section 8 followed by conclusions and
broker). The SPs only pay for their allocated resources and discussion of future work in Section 9.
can take back their remaining budgets. Obviously, a SP will
only buy a computing unit if the potential gain from that 2 RELATED WORK
unit outweighs the cost. It is natural for the SPs to maximize
their net profits in this case. The classical Fisher market The potential benefits and many technical aspects of EC
model does not capture this setting since the utility func- have been studied extensively in the recent literature. First,
tions of the services depend on the resource prices. the hybrid edge/fog-cloud system can be leveraged to
It is worth mentioning that, conventionally, the optimal improve the performance of emerging applications such as
dual variables associated with the supply demand con- cloud gaming and healthcare [11], [12]. In [13] , A. Mukher-
straints (i.e., the capacity constraints of the ENs) are often jee et al. present a power and latency aware cloudlet selec-
interpreted as the resource prices [32] and common tion strategy for computation offloading in a multi-cloudlet
approaches such as network utility maximization (NUM) environment. The tradeoff between power consumption
[33] can be used to compute an ME. However, these and service delay in a fog-cloud system is investigated in
approaches do not work for our models that take budget [14] where the authors formulate a workload allocation
into consideration. Indeed, the main difficulty in computing problem to minimize the system energy cost under latency
an ME in both models stems from the budget constraints constraints. A latency aware workload offloading scheme in
which contain both the dual variables (i.e., prices) and pri- a cloudlet network is formulated in [15] to minimize the
mal variables (i.e., allocation). In the second model, the pri- average response time for mobile users.
ces also appear in the objective functions of the services. In [16], M. Jia et al. explore the joint optimization of
Therefore, the ME computation problem becomes challeng- cloudlet placement and user-to-cloudlet assignment to min-
ing. Note that the pair of equilibrium prices and equilibrium imize service latency while considering load balancing. A
allocation has to not only clear the market but also simulta- unified service placement and request dispatching frame-
neously maximize the utility of every service (as elaborated work is presented in [17] to evaluate the tradeoffs between
in Section 4). the user access delay and service cost. Reference [18]
Fortunately, for a wide class of utility functions, the ME employs Stackelberg game and matching theory to study
in the first model can be found by solving a simple Eisen- the joint optimization among data service operators (DSO),
berg-Gale (EG) convex program [8], [9], [10]. However, the data service subscribers (DSS), and a set of ENs in a three-
EG program does not capture the ME in the second model. tier edge network where the DSOs can obtain computing
Interesting, by reverse-engineering the structure of the pri- resources from different ENs to serve their DSSs.
mal and dual programs in the first model, we can rigorously Another major line of research has recently focused on the
construct a novel convex optimization problem whose solu- joint allocation of communication and computational resour-
tion is an ME of the second model. This technique can also ces for task offloading in the MEC environment [19], [20],
be used to find the ME that considers other practical con- [21]. MEC allows mobile devices to offload computational
straints (e.g., operation cost of the edge servers). Our main tasks to resource-rich servers located near or at cellular BSs,
contributions include: which could potentially reduce the devices’ energy con-
sumption and task execution delay. However, these benefits

Modeling. We formulate a new market-based EC could be jeopardized if multiple users offload their tasks to
resource allocation framework and advocate the MEC servers simultaneously. In this case, a user may not
General Equilibrium theory as an effective solution only suffer severe interference but also receive a very small
method for the proposed problem. amount of EC resource, which would consequently reduce

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY KURUKSHETRA. Downloaded on April 25,2022 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 NGUYEN ET AL.: PRICE-BASED RESOURCE ALLOCATION FOR EDGE COMPUTING: A MARKET EQUILIBRIUM APPROACH 305

data rate, increase transmission delay, and cause high task

execution time on the servers. Hence, offloading decision,
allocation and scheduling of radio resources, and computa-
tional resources should be jointly considered in an integrated
framework.
Different from the existing literature, which mostly deals
with optimizing the overall system performance from a sin-
gle network operator’s point of view, we consider the EC
resource allocation problem from the game theory and mar-
ket design perspectives [8]. In particular, we study how to
allocate resources from multiple ENs to multiple services in a
fair and efficient way. We exploit the General Equilibrium
[4], a Nobel prize-winning theory, to construct an efficient
market-based resource allocation framework. Although this
concept was proposed more than 100 years ago [5], only until
1954, the existence of an ME was proved under mild condi-
tions in the seminal work of Arrow and Debreu [4]. However,
their proof based on fixed-point theorem is non-constructive
and does not give an algorithm to compute an equilibrium
[8]. Recently, theoretical computer scientists have expressed Fig. 1. An EC platform consists of geographically distributed ENs with
great interests in understanding algorithmic aspects of the various configurations. User/service requests are first aggregated at the
aggregation layer, then routed to the ENs for processing. Requests that
General Equilibrium concept. Various efficient algorithms are not handled by the EC platform will be redirected to remote clouds.
and complexity analysis for ME computation have been
accomplished over the past decade [8], [22], [23], [24], [25], Stackelberg game, which not only maximizes the revenue
[26]. Note that although the existence result has been estab- of the provider but also minimizes costs of the services.
lished, there is no general technique for computing an ME. Auction theory has been widely used to study cloud
Our proposed models are inspired by the Fisher market resource allocation [44], [45], [46]. A typical system consists
[5] which is a special case of the exchange market model in of one or several clouds and multiple users. First, the users
the General Equilibrium theory. An exchange market model submit bids, which include their desired resource bundles in
consists of a set of economic agents trading different types terms of VM types and quantities as well as the price that
of divisible goods. Each agent has an initial endowment of they are willing to pay, to an auctioneer. Then, the auctioneer
goods and a utility function representing her preferences solves a winner determination problem to identify accepted
for the different bundles of goods. Given the goods’ prices, bids. Finally, the auctioneer calculates the payment that each
every agent sells the initial endowment, and then uses the winner needs to pay to ensure truthfulness. In auction, the
revenue to buy the best bundle of goods they can afford [4], common objectives are to maximize the social welfare or
[8]. The goal of the market is to find the equilibrium prices maximize the profit of the cloud provider. Additionally,
and allocations that maximize every agent’s utility respect- only winners receive cloud resources. Furthermore, most of
ing the budget constraint, and the market clears. In the existing auction models do not consider elastic user
Fisher market model, every agent comes to the market with demands. For example, previous works often assume that
an initial endowment of money only and wants to buy cloud users are single-minded, who are interested in a spe-
goods available in the market. We cast the EC resource allo- cific bundle only and have zero value for other bundles.
cation problem as a Fisher market. We not only show Different from the existing works on cloud economics
appealing fairness properties of the equilibrium allocation, and resource allocation in general, our design objective is to
but also introduce efficient distributed algorithms to find an find a fair and efficient way to allocate resources from multi-
ME. More importantly, we systematically devise a new and ple nodes (e.g., ENs) to budget-constrained agents (i.e.,
simple convex program to capture the market in which services), which makes every agent happy with her resource
money has intrinsic value to the buyers, which is beyond allotment and ensures high edge resource utilization. The
the scope of the classical Fisher market model. proposed model also captures practical aspects, for exam-
Indeed, there is a rich literature on cloud resource alloca- ple, a service request can be served at different ENs and ser-
vice demands can be defined flexibly rather than fixed
tion and pricing [34]. In [35], [36], the authors propose differ-
bundles as in auction models.
ent profit maximization frameworks for cloud providers.
References [37], [38], [39] study how to efficiently share
resource and profit among cloud providers in a cloud federa- 3 SYSTEM MODEL
tion. Several resource procurement mechanisms are intro- Fig. 1 depicts a generic network architecture that consists of
duced in [40] to assist a user to select suitable vendors in a four layers including the traditional cloud layer, the EC
multi-cloud market. In [41], the interaction between a cloud platform, the aggregation layer, and the end-device layer.
provider and multiple services is modeled as a generalized Besides local execution and remote processing at cloud
Nash game. This model is extended to a multi-cloud multi- DCs, data and requests from end-devices (e.g., smart-
service environment in [42]. A single-cloud multi-service phones, set-top-boxes, sensors) can be handled by the EC
resource provision and pricing problem with flat, on-demand, platform. Note that some data and computing need to be
and on-spot VM instances is formulated in [43] as a done in the local to keep data privacy. A request typically

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY KURUKSHETRA. Downloaded on April 25,2022 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 306 IEEE TRANSACTIONS ON CLOUD COMPUTING, VOL. 9, NO. 1, JANUARY-MARCH 2021

first goes to a Point of Aggregation (PoA) (e.g., switches/ TABLE 1

routers, BSs, APs), then it will be routed to an EN for proc- Notations
essing. In the EC environment, various sources (e.g., smart-
Notation Meaning
phones, PCs, servers in a lab, under-utilized small/medium
data centers in schools/hospitals/malls/enterprises, BSs, PoA, SP Point of Aggregation, service provider
telecom central offices) can act as ENs. Indeed, service/con- EN, EC Edge node, edge computing, data center
EG, ME Eisenberg-Gale, market equilibrium
tent/application providers like Google, Netflix, and Face- CES Constant Elasticity of Substitution
book can proactively install their content and services onto MBB, EF Maximum bang-per-buck, envy-free index
ENs to serve better their customers. Additionally, enter- PropDyn Proportional Response Dynamics
prises, factories, organizations (e.g., hospitals, universities, PropBR Proportional Sharing Best Response (BR)
museums), commercial buildings (shopping malls, hotels, i, j Service index and EN index
airports), and other third parties (e.g., sensor networks) can M, N Number of ENs and number of services
M, N Set of ENs and set of services
also outsource their services and computation to the intelli-
Di ðpÞ Set of ENs giving service i MBB at prices p
gent edge network. Bi Budget of service i
We consider a system encompassing various services ai;j Revenue of service i from unit resource of EN j
and a set of geographically distributed ENs with different cj Resource capacity of EN j
configurations and limited computing capacities. Each ser- dni;j Network delay between service i and EN j
vice has a budget for resource procurement and wants to dpi;j Processing delay of service i at EN j
offload as many requests as possible to the edge network. Timax Maximum delay tolerance of service i
The value of an EN to a service is measured in terms of the xi;j Resource amount of EN j allocated to service i
maximum revenue that it can generate by using the EN’s xi Vector of resources allocated to service i
resource. An EN may have different values to different serv- pj Price of one computing unit of EN j
ices. Since some ENs (e.g., ones with powerful servers) can p, X Resource price vector, resource allocation matrix
be over-demanded while some others are under-demanded, ui ðxi Þ Revenue function of service i
it is desirable to harmonize the interests of the services so Ui ðxi ; pÞ Utility function of service i
PRi Proportionality ratio of service i
that each service is happy with its allotment while ensuring
high resource utilization. An intuitive solution is to assign
prices to ENs and let each service choose its favorite considered as a separate EN. While the computing units in
resource bundle. We assume that there is a platform lying each EN are homogeneous, different ENs can have different
between the services and the ENs. Based on the information types of computing units. Let xi;j be the number of comput-
collected from the ENs (e.g., computing capacity) and the ing units of EN j allocated to service
services (e.g., budgets, preferences), the platform computes
i. The vector of resour-
ces allocated to service i is xi ¼ xi;1 ; xi;2 ; . . . ; xi;M . Finally,
an ME solution including resource prices and allocation, define Bi as the budget of service i. Table 1 summarizes
which not only maximizes the satisfaction of every service important notations used in the paper.
but also fully allocates the ENs’ resources. Our goal is to compute an ME including an equilibrium
In the first model, each service seeks solely to maximize price vector p ¼ ðp1 ; p2 ; :::; pM Þ, where pj is price of EN j,
its revenue under the budget constraint, without concerning and a resource allocation matrix X , in which the element at
about the money surplus after purchasing resources. This the ith row and jth column is xi;j . The utility Ui ðxi ; pÞ of ser-
can be the case where the services and ENs belong to the
vice i is defined as a function of the amount of resources xi
same entity, and each service is assigned a virtual budget
that it receives and the P resource prices p. The capacity con-
representing the service’s priority. In the second model, the
straint of ENs renders: N i¼1 xi;j  cj ; 8j 2 M. Without loss
remaining money does have intrinsic value to the services.
of generality, we normalize the capacity of every EN to be 1
In this case, each service aims to maximize its net profit. For
(i.e., cj ¼ 1; 8j) and scale related parameters (e.g., price,
example, this can be the case where services and ENs are
resource allocation) accordingly. This normalization is just
owned by different entities, and each SP (e.g., Google, Face-
to
PNsimplify expressions and equations. Hence, we have:
book, enterprises) has a certain budget for buying resources
from an infrastructure provider (e.g., a Telco). For simplic- i¼1 xi;j  1; 8j; xi;j  0; 8j:

ity, we assume that the values of ENs to the services are Each service is a player in our market game. Given a
fixed. Our model can be extended to capture time-varying price vector p, service i aims to P maximize its utility Ui ðxi ; pÞ
valuation in a multi-period model by considering each pair subject to the budget constraint j xi;j pj  Bi .
of an EN and a time slot as an independent EN. Definition 4.1. An ME solution (p ; X ) needs to satisfy two
following conditions:
4 PROBLEM FORMULATION

Condition 1: Given the equilibrium resource price
4.1 EC Resource Allocation Problem vector p ¼ ðp1 ; p2 ; . . . ; pM Þ, xi is an optimal resource
Let M, N , M, and N be the sets of ENs and services, and the bundle of service i, for all i, i.e., we have
numbers of ENs and services, respectively. Denote i as the
service index and j as the EN index. We assume that each xi ¼ ðxi;1 ; . . . ; xi;M Þ 2 arg
P max Ui ðxi ; p Þ (1)
xi 0; p x Bi
EN j has cj homogeneous computing units (e.g., servers) j j i;j
[18]. If an EN has several types of computing units, we can
always divide the EN into several clusters, each of which
Condition P
2: All the resources are fully allocated, i.e.,
contains only homogeneous units. Then, each cluster can be we have: i xi;j ¼ 1; 8j.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY KURUKSHETRA. Downloaded on April 25,2022 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 NGUYEN ET AL.: PRICE-BASED RESOURCE ALLOCATION FOR EDGE COMPUTING: A MARKET EQUILIBRIUM APPROACH 307

The first condition can be interpreted as the user satisfaction We model the processing delay at ENs using the widely
condition while the second condition is often called the market used M/G/1 queues and assume that the workload is
clearing condition in Economics [6]. The first condition ensures evenly shared among computing units [18], [27], [28], [41].
that the equilibrium allocation xi maximizes the utility of ser- The average response time dpi;j of EN j for processing service
vice i at the equilibrium prices p considering the user budget i can be computed as follows:
constraint. The second condition maximizes the resource uti- 1
lization of the ENs. It also means the ENs’ resources are fully dpi;j ¼ 
; 8i; j; (3)
sold in the market, which consequently maximizes the profit mi;j  xi;j
i;j
of every EN since the equilibrium prices are non-negative.
where mi;j be the service rate of one computing unit of EN j
The services are players competing for the limited EC resour-
for handling service i, and i;j is the request arrival rate (i.e.,
ces, while the platform tries to satisfy the market clearing
number of requests per time unit) of service i to EN j. For
condition. Prices are used to coordinate the market. i;j
queue stability, we have xi;j < mi;j ; 8i; j: Otherwise, the
Let ui ðxi Þ be the gain/profit/revenue of service i can
achieve from the procured resources. We consider two mod- queuing delay will be infinite as requests accumulated.
els. In the first model (basic model), every service i wants to From (3), we have
1
maximize Ui ðxi ; pÞ ¼ ui ðxi Þ and does not care about how i;j
 Timax  dni;j
much it has to pay as long as the total payment is under its mi;j  xi;j
  (4)
budget. Here, utility of a service is its revenue. In the second 1
model, instead of revenue, the services aim to maximize their ) i;j  xi;j mi;j  max :
Ti  di;j
net profits (i.e., revenue minus cost).
P The service utility in
this model is Ui ðxi ; pÞ ¼ ui ðxi Þ  j pj xi;j ; 8i. We focus on Therefore, if dni;j < Timax , the maximum number of requests
the first model throughout the paper. The second model is that service i can process at EN j is
examined in Section 7.
n  1  o
4.2 Service Utility Model max ¼ max xi;j mi;j  ; 0
i;j
Timax  di;j (5)
In practice, the services may use different criteria to define ¼ xi;j qi;j ; 8i; j
ui ðxi Þ. Our framework takes ui ðxi Þ as an input to compute
an ME solution. How each service evaluates the ENs is not where qi;j ¼ maxfðmi;j  T max1d Þ; 0g. Define a successful
i i;j
our focus. While the proposed model is generic, we consider request as the request whose total delay is smaller or equal
linear functions for the ease of exploring the framework. Let to the maximum delay tolerance. Let ri be the benefit of suc-
ai;j be the gain of serviceP i from one resource unit of EN j. cessfully serving one request of service i [18]. Then, given
Then, we have: ui ðxi Þ ¼ j ai;j xi;j ; 8i. Extensions to more xi;j computing units, the revenue of service i is
general functions will be discussed throughout the paper.
In the following, we present an example of how ai;j can be ui;j ðxi;j Þ ¼ ri qi;j xi;j ¼ ai;j xi;j ; 8i; j; (6)
computed. We consider only delay-sensitive services, which
are also a main target application of EC. For simplicity, we with ai;j ¼ ri qi;j . Thus, we have
assume that the transmission bandwidth is sufficiently large X
M X
M
and the data size of a request is small (e.g., Apple Siri, Goo- ui ðxi Þ ¼ ui;j ¼ ai;j xi;j ; 8i; (7)
gle Voice Search, Google Maps, AR, and Translation). Thus, j¼1 j¼1

the data transmission delay (i.e., size/bandwidth) is

in which ai;j can be computed beforehand. Note that we
assumed to be negligible and we consider only propagation
implicitly assume the request pool of a service is unlimited.
delay and processing delay [27], [41].
We will discuss later how some assumptions can be relaxed.
The total delay of a request of service i from the time a
user sends the request to the time she receives a response Definition 4.2. A function uð:Þ is homogeneous of degree d,
includes the round-trip delay dUEPoA
i between the user and where d is a constant, if uðaxÞ ¼ ad uðxÞ; 8 a > 0 [8].
a PoA of the service, the round-trip network delay dni;j
between the PoA and an EN j hosting the service, and the From (7), it is easy to verify that ui ðxi Þ is a linear function
processing delay at the EN dpi;j . Note that an EN can be that is homogeneous of degree 1.
located in the same place with a PoA (e.g., a BS). In reality, Remark. The value of an EN to a service can be defined
dUEPoA
i is quite small, and we assume it is fixed similar to flexibly. For example, a service may give higher values to
[15]. In other words, we study the system only from the ENs in a populated area or ENs with high reliability. A
aggregation level to the EC platform. For simplicity, we assume suitable weight can be added to ai;j . In the proposed
that each service is located at one PoA (e.g., an IoT gateway, model, each service informs the platform its budget and
a BS, a building). If a service has several PoAs, we need to how much it values different ENs. Based on these infor-
take sum over all the PoAs to get the total number of mation, the platform computes suitable resource alloca-
requests of the service handled by the EC platform. Denote tion satisfying given design objectives. How each service
Timax as the maximum tolerable delay of service i, we have utilizes its allocated resources in the operation stage is
dpi;j þ dni;j  Timax ; 8i; j: (2) not the focus of this work. The key concern of our work is
how to harmonize the interests of different services that
Obviously, the maximum number of requests max
i;j that EN j may have different preferences towards the ENs. Also,
can process is zero if dni;j  Timax . we consider only delay-sensitive services to illustrate one

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY KURUKSHETRA. Downloaded on April 25,2022 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 308 IEEE TRANSACTIONS ON CLOUD COMPUTING, VOL. 9, NO. 1, JANUARY-MARCH 2021

way to model the service utility function. It can be justi- The KKT conditions give
fied by the fact that non-delay-sensitive services can be
handled effectively by cloud DCs and the precious edge @L Bi
¼  hi ¼ 0; 8i (13)
resources can be reserved for important low-latency serv- @ui ui
ices. Nevertheless, our model is generic enough to handle
other service types as long as we can define the utility of @L ai;j
¼ Bi  pj þ ni;j ¼ 0; 8i; j (14)
a service as a suitable function of its allocated EC resour- @xi;j ui
ces. Finally, although we consider computing resources X  X 
only, the proposed framework can apply to a system in ui ¼ ai;j xi;j ; 8i; pj 1  xi;j ¼ 0; 8j (15)
which each service evaluates an EN based on a combina- j i
tion of different resource types of the EN, such as com- ni;j xi;j ¼ 0; 8i; j; pj  0; 8j; ni;j  0; 8i; j: (16)
puting, storage, and bandwidth.
We can infer the following
5 CENTRALIZED SOLUTION
ui ai;j
In the first model, 8i; j :  (17)
P each service i aims to maximize Bi pj
Ui ðxi ; pÞP¼ ui ðxi Þ ¼ j ai;j xi;j subject to the budget con-
ui ai;j
straint j pj xi;j  Bi , 8i. If p is a price vector, the ratio 8i; j : if xi;j > 0 ) ni;j ¼ 0 ) ¼ (18)
ai;j =pj is defined as the bang-per-buck of EN j to service i, Bi pj
which indicates the utility gained by service i through one X X
8j : pj > 0 ) xi;j ¼ 1; xi;j < 1 ) pj ¼ 0: (19)
unit of money spent on EN j (assuming 0/0 = 0). The maxi- i i
mum bang-per-buck (MBB) of service i over the set of ENs is
ai ¼ maxj fai;j =pj g [22]. The demand set Di ðpÞ of service i The dual variable pj in the EG program can be interpreted as
includes all ENs giving it the MBB value, i.e., Di ðpÞ ¼ the price of EN j. Conditions (17) and (18) imply that xi;j > 0
fj : ai;j =pj ¼ ai g; 8i. Intuitively, to maximize its utility, each if and only if j 2 Di ðpÞ, i.e., each service buys resources only
service will spend full budget to buy resources from only from ENs giving it the MBB. This also maximizes ui ðxi Þ.
ENs giving it the MBB. Therefore, a pair ðX; pÞ is an ME if: i) Note that ui =Bi ¼ ai ; 8i. The following theorem captures
given prices p, service i will exhaust its budget to buy some properties of the equilibrium as well as the relationship
resources only from ENs in Di ðpÞ; and ii) the market clears between the EG program and the ME solution.
at prices p. In the following, we will show that the ME in the
Theorem 5.1. The optimal solution to the EG convex program (8),
first model can be inferred from the optimal solution of a
(9), (10), and (11) is an ME. Specifically, the Lagrangian dual
convex optimization problem. Also, we will describe some
variables corresponding to the ENs’ capacity constraints (10) are
properties of the equilibrium. Specifically, for the case of
the equilibrium prices. At the equilibrium, the resource allocation
buyers with linear utilities, the ME can be found by solving
not only maximizes the utility but also exhausts the budget of
the EG convex program given below [8], [9]:
every service. Furthermore, each service purchases resources
X
N
only from ENs giving its MBB. Additionally, the optimal utili-
maximize Bi ln ui (8) ties of the services as well as equilibrium prices are unique.
X ;u
i¼1
subject to Proof. Let X  and ui be the optimal solution to the EG pro-
X
M
gram. Then, X and ui need to satisfy the KKT conditions
ui ¼ ai;j xi;j ; 8i (9) (13), (14), (15), (16), (17), (18), and (19). Denote h , p , and
j¼1
n as the optimal dual variables. From (14), we have
X
N ai;j
xi;j  1; 8j (10) Bi ¼ pj  ni;j ; 8i; j: (20)
ui
i¼1

Multiplying both sides of (20) by xi;j and adding the

xi;j  0; 8i; j: (11) resulting equalities, we get
This problem always has an interior feasible solution by
Bi X X
simply setting xi;j ¼  > 0, for all i and j, where  is suffi-  ai;j xi;j ¼ ðpj  ni;j Þxi;j ; 8i; j: (21)
ui j
ciently small such that all constraints (10) and (11) are satis- j
fied with strict inequality. Hence, Slaters condition holds P
Since ni;j xi;j ¼ 0; 8i;
P j,  and ui  ¼ j ai;j xi;j ; 8i, equa-
and the the Karush–Kuhn–Tucker (KKT) conditions are nec- 
essary and sufficient for optimality [32]. Denote hi , pj , and
tion (21) implies j pj xi;j ¼ Bi ; 8i. Thus, the optimal

ni;j as the dual variables associated with constraints (9), (10), solution to the EG program (8), (9), (10), and (11) fully
and (11), respectively. We have the Lagrangian exhausts the budget of every service. Furthermore, as
X X X shown above, at the optimality, each service buys resour-
Lðu; X; h; p; nÞ ¼ Bi ln ui þ pj ð1  xi;j Þ ces only from ENs giving its MBB value. In other words,
i j i the optimal solution to the EG program maximizes the
X X  XX (12)
þ hi ai;j xi;j  ui þ ni;j xi;j : utility of every service subject to the budget constraint
i j i j because every service uses all of its money to purchase its
MBB resources. This can be inferred from (17) and (18).

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY KURUKSHETRA. Downloaded on April 25,2022 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 NGUYEN ET AL.: PRICE-BASED RESOURCE ALLOCATION FOR EDGE COMPUTING: A MARKET EQUILIBRIUM APPROACH 309

We now consider the market clearing condition. From X X ai;j X  1 X

(19), we can observe that resources of ENs with positive x0i;j pj  x0i;j  xi;j ai;j ¼ xi;j pj : (22)
ai ai
price pj are fully allocated. For ENs with zero prices, their j j j j
resources can be allocated arbitrarily without affecting
The second inequality is due to ui ðx0i Þ  ui ðxi Þ; 8i. Thus
the optimal utility of service since the price is zero [8].
Thus, the market clears. Since (X  , p ) satisfies both con- X
x0i;j pj  Bi ; 8i: (23)
ditions of an ME, the optimal solution to the EG program j
is an ME. P
Finally, since the objective function (8) is strictly concave Since ui ðx0i Þ > ui ðxi Þ for some i, j x0i;j pj  Bi for some
in ui for all i, the optimal utilities are unique. The unique- i. Adding both sides of (23) over all buyers renders
ness of equilibrium prices can be inferred from (18). u
t X XX X X X
Bi < x0i;j pj ¼ xi;j pj  pj ; (24)
From (20), if pj ¼ 0; then ni;j ¼ 0 and ai;j ¼ 0; 8i; j, i i j i j j
which means an EN has price of zero only when it is not
0
P
wanted by all services. We can remove this EN from our because i xi;j  1; 8j (i.e., the capacity constraints of
system. In the following, we consider only the case where ENs). However, (24) means the total prices of all the ENs
pj > 0; 8j. Also, it can be shown that Theorem 5.1 is not is greater than the total budget of all buyers, which can-
only applied to linear utilities, but also true for a wider class not occur. Thus, the equilibrium allocation X  is Pareto-
of homogeneous concave utility functions [10]. Please refer optimal.
to Appendix D in our technical report [47] for more details. - Envy-freeness: To prove that X is envy-free, we need
Next, we study the properties of the equilibrium alloca- to show: Bi0 ui ðxi Þ  Bi ui ðxi0 Þ; 8i; i0 2 N . Let bi;j be the
tion. First, from (8), (9), (10), and (11), it can be easily veri- total money that service i spends on EN j. We have
fied that the equilibrium allocation is scale-free. It means
X X bi;j
that it does not matter if service i reports ai ¼ ðai;1 ; . . . ; ai;M Þ Bi0 ui ðxi Þ ¼ Bi0 ai;j xi;j ¼ Bi0 ai;j
or ei ai for some constant ei , the allocation that it receives is j j
pj
the same. Also, if a service divides its budget into two parts X ai;j X
and acts as two different services with the same original ¼ Bi0 bi;j ¼ B0i ai bi;j
j
pj j
utility function, then the total allocation it obtains from X
the new ME is equal to the original equilibrium allocation. ¼ B0i ai Bi ¼ Bi ai bi0 ;j (25)
j
Furthermore, the equilibrium allocation is not only Pareto-
optimal but also possesses many appealing fairness pro- X ai;j X bi0 ;j
 Bi bi0 ;j ¼ Bi ai;j
perties such as envy-freeness, sharing incentive, and j
pj j
pj
proportionality. X
An allocation X is Pareto-optimal if there does not exist ¼ Bi ai;j xi0 ;j ¼ Bi ui ðxi0 Þ; 8i; j:
j
any allocation X0 such that ui ðx0i Þ  ui ðxi Þ for all i, with
strict inequality holds for at least one i. When budgets are Note that the equalities in the second line of (25) can be
equal, an envy-free allocation X implies ui ðxi Þ  ui ðxi0 Þ for inferred from the fact that each buyer only buys resour-
all i and i0 2 N [7]. Since the budgets can be different, we ces from ENs in its demand set Di while the first inequal-
need to extend this classical definition. An allocation X is a
ity in the fourth line holds because ai  pi;jj P
; 8i; j.
envy-free if ui ðxi Þ  ui ðxi0 BBi0 Þ; 8i; i0 2 N . Let x^ be the allo- 
i
- Proportionality: From Theorem 5.1, i xi;j ¼ 1; 8j.
cation where each service receives resource from every EN Thus, for linear utilities and the envy-free property, we
proportional to its budget, i.e., x^i;j ¼ PB0 i ; 8i; j. The shar- have
Bi0
i
ing-incentive property implies ui ðxi Þ  ui ð^ xi Þ; 8i: Finally, X 
 X

define the proportionality ratio (PR) of service i as: ui ðCÞ ¼ ui xi ¼ ui xi þ ui xi0
i i0 6¼i
PRi ðxi Þ ¼ uuiiðx iÞ
ðCÞ , in which ui ðCÞ is the utility of service i P (26)

 X Bi0
 Bi0

i0
when it receives all the resources from the market (i.e.,  ui xi þ ui xi ¼ ui xi :
Bi Bi
C ¼ ð1; . . . ; 1Þ; C 2 RM ). If PRi ðxi Þ  PBiB , we say that the i0 6¼i
i0 i0
allocation X satisfies the proportionality property. Hence, ui ðxi Þ  PBiB ui ðCÞ; 8i.
i0 i0

Theorem 5.2. At equilibrium, the allocation is Pareto-optimal - Sharing-incentive: At the ME (X  ; p ), no service

and envy-free. It also satisfies the sharing-incentive and propor- spends more than its budget. We have
tionality properties. XX X X X X
xi;j pj  Bi ) pj xi;j  Bi : (27)
Proof. i j i j i i
- Pareto Optimality: We show this by contradiction. Assu- P P
me allocation X is not Pareto-optimal. Then, there exists Thus, j pj  Pi Bi . Consequently,
P resource bundle x^i
an allocation X0 such that ui ðx0i Þ  ui ðxi Þ for Pall i, and costs service i: j x^i;j pj ¼ j PBiB pj  Bi ; 8i: So, ser-
i0 i0
ui ðx0i Þ > ui ðxi Þ for some i. Note that ui ðxi Þ ¼ j ai;j xi;j . vice i can afford to buy bundle x^i at prices p . However,
Consider any feasible allocation X 0 . Recall the MBB of out of all feasible bundles that are affordable to service i,
a
buyer i is ai ¼ maxj pi;jj . We have its favorite one is xi . It means ui ðxi Þ  ui ð^
xi Þ; 8i: u
t

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY KURUKSHETRA. Downloaded on April 25,2022 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 310 IEEE TRANSACTIONS ON CLOUD COMPUTING, VOL. 9, NO. 1, JANUARY-MARCH 2021

Relaxing the coupling constraints, the partial Lagrangian is

X X
X 
LðX; pÞ ¼ Bi ln ui ðxi Þ þ pj 1  xi;j
i j i
X X  X (29)
¼ Bi ln ui ðxi Þ  pj xi;j þ pj :
i j j
Thus, given a price vector p, each service solves
X
maximize Bi ln ui ðxi Þ  pj xi;j : (30)
xi 0
j
To overcome the difficulty raised by the non-uniqueness of
the optimal demand of the services with linear utilities, we
Fig. 2. Market equilibrium with linear utilities.
propose to approximate the linear utility function by a Con-
stant Elasticity of Substitution (CES) function, which is
6 DECENTRALIZED SOLUTION widely used in Economics and Computer Science [6], [8]. PMA
CES function 1
has the following form: u CES
i ðx i Þ ¼ ð j¼1
A common approach for implementing distributed algo- ðai;j xi;j Þr Þr ; r < 1; r 6¼ 0: Indeed, the linear utility function
rithm is to let the platform iteratively compute prices of is a special case of the CES function family as r ! 1. We can
the ENs and broadcast the updated prices to the services. approximate the original linear utility function by a CES
Then, each service finds its optimal demand bundle and function where r ¼ 1   with  is arbitrarily small. As  ! 0,
sends the updated demand to the platform. This price- uCES ! ui . Clearly, a CES function is strictly concave and
i
based strategy can be implemented in a tatonnment style homogeneous [6]. Hence, the EG program and Theorem 5.1
or using the dual decomposition method [33]. Unfortu- also apply to CES functions [8], [10]. Additionally, we can
nately, linear utilities may result in non-unique optimal observe that maximizingPa CES function above is equivalent
demand bundles because multiple ENs may give the to maximizing ui ðxi Þ ¼ j ðai;j xi;j Þr . Since a CES function is
same MBB to a buyer. Hence, the algorithm cannot termi- strictly concave, the optimal demand bundle of a service is
nate without aggregated demand coordination from the unique. Consider the following optimization problem
platform. Consider an example with two services and
X
three ENs. The system parameters are: B1 = $1, B2 = $4, maximize ui ðxi Þ subject to pj xi;j  Bi : (31)
a1 ¼ ð1; 10; 4Þ, and a2 ¼ ð4; 8; 8Þ. xi 0
j
Fig. 2a presents the ME from the centralized EG pro-
gram. The value associated with each edge between a Proposition 6.1. Given a positive price vector p and a CES
service and an EN indicates the amount of resource that approximation function, each service i can either solve Problem
the service buys from the EN. For example, in Fig. 2a, we (30) or Problem (31). Both the problems have the same closed
have: x1;1 ¼ 0; x1;2 ¼ 0:5, and x1;3 ¼ 0: The equilibrium form solution as follows:
price vector is p ¼ ð1; 2; 2Þ. The demand sets are: D1 ¼ f2g  r 1r1
ai;j Bi
and D2 ¼ f1; 2; 3g. Given the equilibrium prices, the set of xi;j ¼ r : (32)
optimal (i.e., utility-maximizing) resource bundles of ser- pj PM ai;j 1r
j¼1 pj
vice 2 is infinite. Hence, even if a distributed algorithm
reaches the exact equilibrium prices at some iteration, it
Proof. Refer to our technical report [47]. u
t
may not stop since the total demand reported by the
buyers may not equal to the total supply. For instance, in Thus, based on the dual decomposition method where
Fig. 2b, although the platform announces the exact equi- each service solves the sub-problem (30), we have the fol-
librium prices, service 2 may choose to buy all resources lowing distributed algorithm with CES function approxima-
from EN2 and EN3. Then, the algorithm may never termi- tion (Algorithm 1). With a sufficiently small step size, it is
nate. In the following, we present two distributed algo- guaranteed to terminate and converge to an (approximate)
rithms to find the ME. global optimal solution [32], [33]. Our simulation results
confirm that Algorithm 1 produces a solution arbitrarily
6.1 Dual Decomposition with Function close to the optimal one from the centralized EG program.
Approximation
Using Lagrangian relaxation [32], [33], we can decompose 6.2 Proportional Response Dynamics Strategy
the EG convex program into sub-problems, each of which In this section, we present the Proportional Response
can be solved by a service. We observe that the EG pro- Dynamics (PropDyn) algorithm proposed by the P2P commu-
gram (8), (9), (10), and (11) can be written equivalently as nity. This distributed algorithm is very simple to implement
follows. and has been proved to converge to an ME [29]. Basically, in
every iteration t, each service updates its bids proportional to
X
N the utilities it receives from the previous iteration. Specifi-
maximize Bi ln ui ðxi Þ u ðt1Þ
cally, bi;j ðtÞ ¼ Bi ui;ji ðt1Þ ; 8i; j; t: Since the ENs’ capacities
X
i¼1
(28) are normalized, the price P of an EN equals to the total bids
X
N
sent to it, i.e., pj ðtÞ ¼ i bi;j ðtÞ. By bidding bi;j ðt  1Þ to EN j,
subject to xi;j  1; 8j; xi;j  0; 8i; j:
i¼1
service i obtains an amount of resource xi;j ðt  1Þ ¼
bi;j ðt  1Þ=pj , and gains a utility ui;j ðt  1Þ ¼ ai;j xi;j ðt  1Þ.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY KURUKSHETRA. Downloaded on April 25,2022 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 NGUYEN ET AL.: PRICE-BASED RESOURCE ALLOCATION FOR EDGE COMPUTING: A MARKET EQUILIBRIUM APPROACH 311

P
Finally, ui ðt  1Þ ¼ j ui;j ðt  1Þ is the total utility of service i strategy converges, it converges to an NE. As mentioned in
at iteration t  1. The salient feature of this algorithm is that it [30], the algorithm normally converges after a few rounds.
can be implemented efficiently in a distributed manner. In Interestingly, our simulation shows that buyers do not
particular, each EN only needs to know the total bid that it gain significantly by playing BR. Indeed, most of buyers
receives to compute the price while each buyer only needs to achieve lower utilities in the PropBR scheme compared to
know its own information and learns its utilities achieved in the PropDyn scheme. Furthermore, to play BR dynamics,
the previous iteration to compute its new bids. The algorithm each buyer has to know total bids of others and the actual
terminates when the price deviation of every EN is suffi- capacity of every EN [30]. In PropDyn, buyers only need to
ciently small [29]. The major difference between this novel know their own information. Therefore, in a proportional
algorithm and traditional distributed algorithms is that in sharing system, buyers may not have incentives to play BR.
each iteration, every service computes its new bids as men-
tioned above instead of its optimal demand bundle. 7 NET PROFIT MAXIMIZATION
Different from the basic model, in the second model, the serv-
Algorithm 1. Function Approximation Algorithm
ices try to optimize their net profits (i.e., revenue minus
1: Initialization: iteration t = 0, set initial prices of ENs cost) insteadP of revenue. Specifically, the net profit of service
pð0Þ ¼ p0 , and set step size að0Þ and tolerance g to be small. i is vi ðxi Þ ¼ j ðai;j  pj Þxi;j ; 8i: Given prices p, the objective
2: repeat of service i is to maximize Ui ðxi ; pÞ ¼ vi ðxi Þ subject to:
3: At iteration t, the platform broadcasts prices p(t) to the P
j xi;j pj  Bi ; 8i and xi;j  0; 8i; j. Indeed, maximizing
P
buyers.
the net profitP vi ðxi Þ is equivalent to maximizing
P j ðai;j 
4: Each buyer computes its optimal demand xi ðtÞ using (32)
pj Þxi;j þ Bi ¼ j ai;j xi;j þ si , where si ¼ Bi  j pj xi;j is the
and sends it to the platform.
surplus money of service i after purchasing xi . Inspired by
5: The platform updates the prices
PN 
pj ðtþ 1Þ ¼ max fpj ðtÞ
the EG program for the basic model, we would like to con-
 þ aðtÞ 1  i¼1 xi;j ðtÞ ; 0g; 8j struct a similar convex program to capture the ME in this
6: until pj ðt þ 1Þ  pj ðtÞ < g; 8j, or the number of iterations t
is too large. new model.
7: Output: equilibrium prices p and optimal allocation X  . Note that without budget consideration, this game-theo-
retic problem can be solved efficiently by writing down a
social welfare maximization problem (i.e., maximizing sum
Algorithm 2. Best Response Dynamics Algorithm [30] of utilities of all the services), then use the dual decomposi-
1: Sort ENs according to the decreasing order of bi;j
a
i;j
. tion method [33] to decompose it into sub-problems, each of
Output a sorted list Li ¼ fi1 ; i2 ; . . . ; iM g: which is solved by one service. Each sub-problem is exactly
2: Find the largest k such that a net profit maximization problem of a service. Unfortu-
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ai;i bi;ik Pk nately, this strategy fails when we consider budget since the
Pk pkffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ðBi þ j¼1 bi;ij Þ  bi;ik  0 social welfare maximization problem cannot be decom-
ai;ij bi;ij
j¼1
3: Set bil ¼ 0 for l > k, and for 1  l  k, set posed due to the coupling budget constraints.
pffiffiffiffiffiffiffiffiffiffiffiffiffiffi   Our derivation of the new convex optimization problem
ai;i bi;il Pk
bil ¼ Pk plffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Bi þ j¼1 bi;ij  bi;il is based on reverse-engineering the basic model.
ai;ij bi;ij
j¼1

Proposition 7.1. The equilibrium prices in the basic model can

To illustrate the effectiveness of the PropDyn mechanism be found by solving the following convex problem.
as well as the ME concept, we compare it with the Propor-
tional Sharing Best Response (BR) mechanism (PropBR) pro- X
M X
N
minimize pj  Bi lnðhi Þ
posed in [30], which aims to find a Nash Equilibrium (NE). p;h
j¼1 i¼1 (33)
In a non-cooperative game, a NE is a stable state of a system
subject to pj  ai;j hi ; 8i; j; pj  0; 8j:
where no player can gain by a unilateral change of strategy
if the strategies of the others are fixed [8]. Both [29] and [30]
study a proportional sharing system where the resource of Proof. We can obtain this convex problem by using Lagrang-
every node is shared proportionally to the services accord- ian and Fenchel conjugate function [32] to construct the
bi;j
ing to their bids. Specifically, we have xi;j ¼ bi;j þb i;j
; 8i; j, dual problem of the original EG program. Indeed, hi and
where bi;j is the total bid of all the services except i. In both pj are the dual variables associated with (9) and (10). See
mechanisms, the actions of the services are the bids (bi;j ) our technical report for the full proof [47]. u
t
submitted to the ENs. However, instead of updating its bids P
Clearly, to maximize vi ðxi Þ ¼ j ðai;j  pj Þxi;j , service i
following the rule in PropDyn, each service in the PropBR
will never buy resource from EN j if ai;j < pj . In other
mechanism selfishly maximizes its utility given strategies
words, service i would only buy resources from ENs in the
taken by other services [30]. p p
set Ai ¼ j : ai;jj  1 : From (33), we have hi  ai;jj ; 8i. From
Algorithm 2 is the BR algorithm that buyer i will execute
given the total bid bi;j of other buyers. The whole algorithm these observations, we conjecture that the following pror-
is implemented in rounds. In each round, each buyer in gram captures the equilibrium prices in our second market
turns runs Algorithm 2 and updates its bid vector bi to the model (i.e., net profit maximization).
platform. The platform broadcasts new bids to all buyers in X
M X
N
the system. A round completes when all buyers have minimize pj  Bi lnðhi Þ: (34)
p;h
updated their bids. Obviously, whenever this BR dynamics j¼1 i¼1

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY KURUKSHETRA. Downloaded on April 25,2022 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 312 IEEE TRANSACTIONS ON CLOUD COMPUTING, VOL. 9, NO. 1, JANUARY-MARCH 2021

Fig. 4. Performance comparison.

the area. We generate a total of 100 ENs and 1000 locations.

We assume that each service is located at one location. For
the sake of clarity in analysis, in the base case, we consider a
Fig. 3. Valuations of ENs to the buyers. small system with 8 ENs and 4 services (i.e., M = 8 and
N = 4), which are selected randomly in the set of 100 ENs and
subject to 1000 services. The network delay between a service and an
EN is assumed to be proportional to the distance between
pj  ai;j hi ; 8i; j; hi  1; 8i; hi  0; 8i; pj  0; 8j:
them. The maximum tolerable delay of the services follows a
uniform distribution over the interval [15, 25]. The service
Theorem 7.2. The solution of the following convex program is
rate mi;j is generated randomly from 80 to 240 requests per
exactly an ME of the new market model.
time unit. The service price is from 2 to 3 per 100000 requests.
N 
X  The number of computing units in the ENs ranges from
maximize Bi ln ui  si 10 to 20. From these parameters, we can compute ai;j of the
X ;u;s
i¼1 services as in (6). The net profit maximization model is con-
X
M sidered in Section 8.5. In the base case, we assume that the
subject to ui  ai;j xi;j þ si ; 8i (35) services have equal budget. Fig. 3 depicts the valuations of
j¼1
the ENs to the buyers in the base case. The base case is used in
X
N
all the simulations unless mentioned otherwise.
xi;j  1; 8j; xi;j  0; 8i; j; si  0; 8i:
i¼1
8.2 Performance Comparison
At the equilibrium, the total of money spent and surplus In the first model captured by the EG program, the absolute
money of every service equals to its budget. Additionally, the value of the budget only affects the equilibrium prices by a
optimal utility of every service is unique and greater or equal scaling factor (e.g., all the prices increase twice as the budget
to its budget. For any buyer who has surplus money, her utility of every service is double) and does not affect the allocation
equals her budget. and utilities of the services. The budget is normalized such
that the total budget of all services is one. The prices act as a
Proof. See our technical report [47]. u
t
means to allocate resources only.
The convex problem (35) is indeed the dual program of We consider five schemes, including: the proposed ME,
problem (34). We can interpret problem (35) as follows. the proportional sharing (Prop.), the social welfare maximi-
First, the utility of a service is the sum of its revenue and its zation with equal weights (SW1), social welfare maximiza-
surplus money. The first part of the objective function is the tion with different weights (SW2), and the maxmin fairness
weighted sum of logarithmic utilities of the services similar (MaxMin) schemes. In the proportional sharing, each buyer
to that of the EG program. However, since the surplus i receives PBiB portion of resource of every EN. In the social
money does not contribute (i.e., not visible) to the market, i i
welfare maximization schemes, budget P is not considered,
we should subtract this amount from the aggregated utility and the objective is to maximize i wi ui ðxi Þ subject to the
function, i.e., the objective function. Finally, similar to the capacity constraints of the ENs. wi is the weighting factor of
EG program, although budget constraints are not included service i. In SW1, all weights are equal. In SW2, the weight
in (35), the optimal solution satisfies these constraints. It is of each service is its budget. Finally, without budget consid-
worth noting that, somewhat surprisingly, although our eration, the MaxMin scheme aims to maximize mini ui ðxi Þ
reverse-engineering approach is specialized for linear reve- under ENs’ capacity constraints.
nue functions only, the convex program (35) works also for Figs. 4a and 6b present performance comparison among
a wider class of homogeneous concave revenue functions. these schemes under both equal budget and different bud-
This proof relies on the fact that if ui ðxi Þ is concave and get settings. We can observe that the ME scheme balances
homogeneous of degree one, then ui ðxi Þ þ si is also concave well the tradeoff between system efficiency and fairness.
homogeneous of degree one. Please refer to Appendix E in First, the ME scheme considerably outperforms the Prop.
[47] for proof sketch. scheme, which confirms the sharing-incentive property of the
ME solution. The MaxMin scheme produces a fair allocation
8 NUMERICAL RESULTS among the buyers but the total utility of the buyers is much
8.1 Simulation Settings lower compared to other schemes. Noticeably, although the
We consider a square area with dimensions of 10 km  10 km. total utility is largest, both schemes SW1 and SW2 produce
The locations of ENs and services are generated randomly in undesirable allocations since some buyers (e.g., buyer 1) are

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY KURUKSHETRA. Downloaded on April 25,2022 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 NGUYEN ET AL.: PRICE-BASED RESOURCE ALLOCATION FOR EDGE COMPUTING: A MARKET EQUILIBRIUM APPROACH 313

Fig. 5. Utility efficiency comparison (N = 4).

Fig. 8. Impact of budget ratio on the buyers’ utilities.

Fig. 6. Envy-freeness comparison (N = 4).

not allocated anything and have zero utility in these

schemes. Figs. 6a and 6b compare envy-freeness indices
ðxi Þ=Bi
(EF ¼ mini;j uuiiðx j Þ=Bj
[30]) of the different schemes. An allo-
cation is envy-free if EF equals to one. The Prop. scheme is
obviously envy-free by definition. These figures also confirm
that the ME is envy-free. Furthermore, the proposed ME
scheme significantly outperforms the social welfare maximi-
zation and MaxMin schemes in terms of envy-free fairness.
Finally, we can show that the ME satisfies the proportional- Fig. 9. Impact of budget ratio on the equilibrium prices.
ity fairness property. Due to the space limitation, we leave
this result in our technical report [47]. EN7, and EN8 can satisfy the delay requirement of service 1
as seen in Fig. 3, the prices of EN7 and EN8 change consid-
8.3 Sensitivity Analysis erably as budget of service 1 varies. Also, because EN5 and
First, we examine the impact of budget on the equilibrium EN6 are less valuable to the buyers, their equilibrium prices
allocation by varying the budget ratio among the buyers. are significantly lower than the prices of other ENs while
Figs. 7a, 7b, and 7b show impact of budget on the equilib- the prices of EN2 and EN8 are highest because they have
rium allocation as we vary the budget ratio between serv- high values to all the buyers. These observations imply the
ices 1 and 2. We observe that buyer 1 is allocated more proposed method is effective in pricing.
resources as her budget increases, which also increases her The impact of the number of players (i.e., number of ENs
utility. The allocation and utility of buyer 2 decrease as her and number of services) on the ME is illustrated in Figs. 10a
budget decreases. Fig. 8 further supports this observation and 10b. The buyers have the same budget in this case. We
where r is the budget ratio between services 1 and 2. show the utilities of buyers 1, 2, 3, and 4 in these figures. As
Hence, we can conclude that the proposed algorithm is expected, as the number of buyers increases, the utility of
effective to capture service priority in terms of budget in the individual buyer decreases since the same set of ENs has to
allocation decision. be shared among more services. On the other hand, the ser-
Fig. 9 shows the dependence of the equilibrium prices on vice utility increases significantly as the number of ENs
the budget ratio of the buyers. For example, since only EN2, increases.

Fig. 7. Impact of budget ratio on the equilibrium allocation.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY KURUKSHETRA. Downloaded on April 25,2022 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 314 IEEE TRANSACTIONS ON CLOUD COMPUTING, VOL. 9, NO. 1, JANUARY-MARCH 2021

Fig. 10. Impact of M and N on the players’ utilities.

Fig. 13. Convergence of CES approximation.

Fig. 11. Convergence of EN prices and bids.

Fig. 14. CES approximation utility comparison.
8.4 Analysis of Distributed Algorithms
8.4.1 Proportional Dynamics Allocation 8.4.2 Function Approximation Algorithm
The proportional dynamics mechanism (PropDyn) has low The convergence properties of the CES approximation
complexity and can be implemented in a distributed man- scheme as well as its performance are reported. Thanks to
ner. The convergence properties of this algorithm in the the closed form expression of the optimal demand, the algo-
base case with 8 ENs and 4 services is shown in Figs. 11a rithm runs very fast even with high number of iterations.
and 11b. As we can see, the prices and the bids converge As expected, the number of iterations depends strongly on
after a few tens of iterations. The running time of the algo- the step size and the initial prices. The convergence of
rithm is in order of milliseconds. Figs. 12a, 12b, and 12c EN6’s price (p6 ) is shown in Fig. 13a. The number of itera-
compare the buyers’ utilities in the PropDyn and PropBR tions decreases as the initial prices are close to the final ME
schemes. We select a particular instance with a set of 10 prices, which are unknown. The number of iterations
buyers and 20 ENs from the generated system data. Note decreases as the step size increases, but we cannot increase
that we have run simulation with numerous instances and the step size g too much to ensure convergence. Fig. 13b
obtain similar trends. The utility of each buyer in the partic- presents the price traces of different ENs until convergence
ular instance is presented in Fig. 12a. As we can see, the util-
with a ¼ 0:001 and p0 ¼ 0:2.
ity values are higher for most of the buyers in the PropDyn
In Fig. 14a, we study the performance of the approxima-
scheme compared to the those in the PropBR scheme.
tion scheme by comparing utility of the buyers under the
In 12b, we add a random variable to each ai;j and run the
centralized convex program (EG), the approximation CES
schemes 100 times and take the average results. In 12c, we
generate ai;j randomly in the range between 0.01 and 0.09. utility (CES), and the approximation linear utility (Approx.).
As we can observe, the buyers’ utilities tend to be higher in In the Approx. scheme, the utility of buyer i is xi;j ai;j where
the PropDyn scheme in comparison with the PropBR xi;j is the solution of the optimization problem with CES
scheme. Furthermore, the PropBR requires buyers to know approximation utilities. As we can observe, the values of
more system information to play their BR actions in each the utilities are very similar, which confirms that the pro-
round. The numerical results show that it brings almost no posed approximation scheme performs well. In this figure,
benefit to the buyers (no utility gain in most cases) to play we set r to be 0.99. Finally, the equilibrium prices with dif-
PropBR scheme. Hence, we can infer that the buyers should ferent values of r is shown in Fig. 14b. It is easy to see that
just follow the PropDyn scheme and obtain an ME allocation. the prices are almost equal for different values of r.

Fig. 12. Utility comparison between PropBR and PropDyn.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY KURUKSHETRA. Downloaded on April 25,2022 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 NGUYEN ET AL.: PRICE-BASED RESOURCE ALLOCATION FOR EDGE COMPUTING: A MARKET EQUILIBRIUM APPROACH 315

Fig. 15. ME in the net profit maximization model.

8.5 Net Profit Maximization Model 9 CONCLUSION AND FUTURE WORKS

We now evaluate the second model where the services aim
In this work, we consider the resource allocation for an EC
to maximize their net profits. We use the same system set-
system which consists geographically distributed heteroge-
ting with 4 services and 8 ENs in the base case as before.
neous ENs with different configurations and a collection of
From the objective function of the buyer, we know that a
services with different desires and buying power. Our main
buyer will buy resource from an EN only when the price
contribution is to suggest the famous concept of General
of the EN is less than or equal to its utility gain from the
Equilibrium in Economics as an effective solution for the
EN. In the revenue maximization case as in the basic model,
underlying EC resource allocation problem. The proposed
the equilibrium prices increase linearly at the same rate as
the budget. However, as can be seen in Fig. 15a, this prop- solution produces an ME that not only Pareto-efficient but
erty does not hold in the net profit maximization model. also possesses many attractive fairness properties. The
Budget scale is the scaling factor by which we multiply the potential of this approach are well beyond EC applications.
original budget. The figure shows that the equilibrium pri- For example, it can be used to share storage space in edge
ces increase then become saturated after certain values of caches to different service providers. We can also utilize the
the budgets. At these (saturated) prices, buying resources proposed framework to share resources (e.g., communica-
from the ENs or not does not change the utility for a buyer tion, wireless channels) to different users or groups of users
(i.e., pj ¼ ai;j ). When the budget is large enough, the utili- (instead of services and service providers). Furthermore, the
ties of the buyers become equal to their budgets. It means proposed model can extend to the multi-resource scenario
procuring resources or not does not bring any additional where each buyer needs a combination of different resource
benefit to the buyers. These results are shown in Figs. 15b types (e.g., storage, bandwidth, and compute) to run its ser-
and 15c. vice. We will formally report these cases (e.g., network slic-
Figs. 16a and 16b present equilibrium prices and opti- ing, NFV chaining applications) in our future work.
mal utilities, respectively, as the budget varies. Rev.max The proposed framework could serve as a first step to
corresponds to the first model (i.e., revenue maximization) understand new business models and unlock the enormous
with scale equal to 1. As we can observe, for the same bud- potential of the future EC ecosystem. There are several
get (i.e., scale = 1), equilibrium prices in the second model future research directions. For example, we will investigate
are smaller than equilibrium prices in the first model the ME concept in the case when several edge networks
because in the net profit maximization model, a service cooperate with each other to form an edge/fog federation.
only buys resource from an EN that gives it positive gain. Investigating the impacts of the strategic behavior on the
Also, the service utilities at the equilibrium in the second efficiency of the ME is another interesting topic. Note that
model is greater than those in the first model due to lower N. Chen et al. [24] have shown that the gains of buyers for
equilibrium prices and budget surplus is considered in the strategic behavior in Fisher markets are small. Additionally,
second model. Finally, in the second model, the equilib- in this work, we implicitly assume the demand of every ser-
rium prices and optimal utilities increase as the budget vice is unlimited. It can be verified that we can add the max-
increases. As explained above, equilibrium prices become imum number of requests constraints to the EG program to
saturated in the second model at certain points. Hence, the capture the limited demand case, and the solution of this
equilibrium prices increase very little as the budget scale modified problem is indeed an ME. However, although the
increases from 1 to 1.5. optimal utilities of the services in this case are unique, there
can have infinite number of equilibrium prices. We are
investigating this problem in our ongoing work. Also, inte-
grating the operation cost of ENs into the proposed ME
framework is a subject of our future work. Finally, how to
compute market equilibria with more complex utility func-
tions that capture practical aspects such as task moving
expenses among ENs and data privacy is an interesting
future research direction. It is also interesting to test the per-
formance of the proposed approach on real datasets of an
Fig. 16. Impact of budget on equilibrium prices and utilities. EC system when EC is widely deployed.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY KURUKSHETRA. Downloaded on April 25,2022 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 316 IEEE TRANSACTIONS ON CLOUD COMPUTING, VOL. 9, NO. 1, JANUARY-MARCH 2021

ACKNOWLEDGMENTS [23] V. V. Vazirani and M. Yannakakis, “Market equilibrium under

separable, piecewise-linear, concave utilities,” J. ACM, vol. 58,
This research was supported by the Natural Sciences and no. 3, May 2011, Art. no. 10.
[24] N. Chen, X. Deng, B. Tang, and H. Zhang, “Incentives for strategic
Engineering Research Council of Canada and by a Vanier behavior in Fisher market games,” in Proc. Conf. Artif. Intell.,
Graduate Scholarship. Feb. 2016, pp. 453–459.
[25] X. Chen, D. Paparas, and M. Yannakakis, “The complexity of
non-monotone markets” J. ACM, vol. 64, no. 3, Jun. 2017,
REFERENCES Art. no. 20.
[1] M. Chiang and T. Zhang, “Fog and IoT: An overview of research [26] J. Garg, R. Mehta, M. Sohoni, and V. V. Vazirani, “A complemen-
opportunities,” IEEE Internet Things J., vol. 3, no. 6, pp. 854–864, tary pivot algorithm for market equilibrium under separable,
Dec. 2016. piecewise-linear concave utilities,” SIAM J. Comput., vol. 44, no. 6,
[2] M. Satyanarayanan, “The emergence of edge computing,” Com- pp. 1820–1847, 2015.
put., vol. 50, no. 1, pp. 30–39, Jan. 2017. [27] Z. Liu, M. Lin, A. Wierman, S. Low, and L. L. H. Andrew,
[3] W. Shi, J. Cao, Q. Zhang, Y. Li, and L. Xu, “Edge computing: “Greening geographical load balancing,” IEEE/ACM Trans. Netw.,
vision and challenges,” IEEE Internet Things J., vol. 3, no. 5, vol. 23, no. 2, pp. 657–671, Apr. 2015.
pp. 637–646, Oct. 2016. [28] L. Tang and H. Chen, “Joint pricing and capacity planning in the
[4] K. J. Arrow and G. Debreu, “Existence of equilibrium for a com- IaaS cloud market,” IEEE Trans. Cloud Comput., vol. 5, no. 1,
petitive economy,” Econometrica, vol. 22, no. 3, pp. 265–290, 1954. pp. 57–70, Jan.-Mar. 2017.
[5] W. C. Brainard and H. E. Scarf, “How to compute equilibrium pri- [29] F. Wu and L. Zhang, “Proportional response dynamics leads to
ces in 1891,” Cowles Foundation Discussion Papers 1272, Cowles market equilibrium,” in Proc. 39th Annu. ACM Symp. Theory Com-
Foundation for Research in Economics, Yale University, 2000, put., 2007, pp. 354–363.
https://ideas.repec.org/p/cwl/cwldpp/1272.html [30] M. Feldman, K. Lai, and L. Zhang, “The proportional-share alloca-
[6] A. Mas-Colell, M. D. Whinston, and J. R. Green, Microeconomic tion market for computational resources,” IEEE Trans. Parallel Dis-
Theory, 1st ed. New York, NY, USA: Oxford Univ. Press, 1995. trib. Syst., vol. 20, no. 8, pp. 1075–1088, Aug. 2009.
[7] H. Moulin, “Fair division and collective welfare,” Cambridge, [31] J. F. Nash, “The bargaining problem,” Econometrica, vol. 18,
MA, USA: MIT Press, 2004. pp. 155–162, 1950.
[8] N. Nisan, T. Roughgarden, E. Tardos, and V. Vazirani, [32] S. Boyd and L. Vandenberghe, “Convex Optimization,”
“Algorithmic Game Theory,” Cambridge, U.K.: Cambridge Univ. Cambridge, U.K.: Cambridge Univ. Press, 2004.
Press, 2007. [33] D. P. Palomar and M. Chiang, “A tutorial on decomposition meth-
[9] E. Eisenberg and D. Gale, “Consensus of subjective probabilities: ods for network utility maximization,” IEEE J. Sel. Areas Commun.,
The pari-mutuel method,” Ann. Math. Statistics, vol. 30, pp. 165– vol. 24, no. 8, pp. 1439–1451, Aug. 2006.
168, 1959. [34] N. C. Luong, P. Wang, D. Niyato, Y. Wen, and Z. Han, “Resource
[10] E. Eisenberg, “Aggregation of utility functions,” Manage. Sci., management in cloud networking using economic analysis and
vol. 7, pp. 337–350, 1961. pricing models: a survey,” IEEE Commun. Surv. Tut., vol. 19, no. 2,
[11] Y. Lin and H. Shen, “CloudFog: leveraging fog to extend cloud pp. 954–1001, Apr.–Jun. 2017.
gaming for thin-client MMOG with high quality of service,” IEEE [35] H. Xu and B. Li, “Dynamic cloud pricing for revenue maxi-
Trans. Parallel Distrib. Syst., vol. 28, no. 2, pp. 431–445, Feb. 2017. mization,” IEEE Trans. Cloud Comput., vol. 1, no. 2, pp. 158–171,
[12] L. Gu, D. Zeng, S. Guo, A. Barnawi, and Y. Xiang, “Cost efficient Jul.–Dec. 2013.
resource management in fog computing supported medical [36] A. N. Toosi, K. Vanmechelen, K. Ramamohanarao, and R. Buyya,
cyber-physical system,” IEEE Trans. Emerg. Topics Comput., vol. 5, “Revenue maximization with optimal capacity control in infra-
no. 1, pp. 108–119, Jan.-Mar. 2017. structure as a service cloud markets,” IEEE Trans. Cloud Comput.,
[13] A. Mukherjee, D. De, and D.G. Roy, “A power and latency vol. 3, no. 3, pp. 261–274, Jul.–Sep. 2015.
aware cloudlet selection strategy for multi-cloudlet environ- [37] L. Mashayekhy, M. M. Nejad, and D. Grosu, “Cloud federations in
ment,” IEEE Trans. Cloud Comput., https://ieeexplore.ieee.org/ the sky: formation game and mechanism,” IEEE Trans. Cloud Com-
document/7506034/ put., vol. 3, no. 1, pp. 14–27, Jan.-Mar. 2015.
[14] R. Deng, R. Lu, C. Lai, T. H. Luan, and H. Liang, “Optimal work- [38] M. Hadji and D. Zeghlache, “Mathematical programming
load allocation in fog-cloud computing toward balanced delay approach for revenue maximization in cloud federations,” IEEE
and power consumption,” IEEE Internet Things J., vol. 3, no. 6, Trans. Cloud Comput., vol. 5, no. 1, pp. 99–111, Jan.-Mar. 2017.
pp. 1171–1181, Dec. 2016. [39] I. Petri, J. Diaz-Montes , M. Zou, T. Beach, O. Rana, and M. Parashar,
[15] X. Sun and N. Ansari, “Latency aware workload offloading in the “Market models for federated clouds,” IEEE Trans. Cloud Comput.,
cloudlet network,” IEEE Commun. Lett., vol. 21, no. 7, pp. 1481– vol. 3, no. 3, pp. 398–410, Jul.–Sep. 2015.
1484, Jul. 2017. [40] A. Prasad and S. Rao, “A mechanism design approach to resource
[16] M. Jia, J. Cao, and W. Liang, “Optimal cloudlet placement and procurement in cloud computing,” IEEE Trans. Comput., vol. 63,
user to cloudlet allocation in wireless metropolitan area no. 1, pp. 17–30, Jan. 2014.
networks,” IEEE Trans. Cloud Comput., vol. 5, no. 4, pp. 725–737, [41] D. Ardagna, B. Panicucci, and M. Passacantando, “Generalized Nash
Oct.–Dec. 2017. equilibria for the service provisioning problem in cloud systems,”
[17] L. Yang, J. Cao, G. Liang, and X. Han, “Cost aware service place- IEEE Trans. Serv. Comput., vol. 6, no. 4, pp. 429–442, Oct.-Dec. 2013.
ment and load dispatching in mobile cloud systems,” IEEE Trans. [42] D. Ardagna, M. Ciavotta, and M. Passacantando, “Generalized
Comput., vol. 65, no. 5, pp. 1440–1452, May 2016. Nash equilibria for the service provisioning problem in multi-
[18] H. Zhang, Y. Xiao, S. Bu, D. Niyato, F.R. Yu, and Z. Han, “Computing cloud systems,” IEEE Trans. Serv. Comput., vol. 10, no. 3, pp. 381–
resource allocation in three-Tier IoT fog networks: A joint optimiza- 395, May-Jun. 2017.
tion approach combining stackelberg game and matching,” IEEE [43] V. Cardellini, V. Di Valerio , and F. Lo Presti, “Game-theoretic
Internet Things J., vol. 4, no. 5, pp. 1204–1215, Oct. 2017. resource pricing and provisioning strategies in cloud systems,”
[19] S. Sardellitti, G. Scutari, and S. Barbarossa, “Joint optimization of IEEE Trans. Serv. Comput.
radio and computational resources for multicell mobile-edge [44] L. Mashayekhy, M. M. Nejad, and D. Grosu, “Physical machine
computing,” IEEE Trans. Signal Inf. Process. Netw., vol. 1, no. 2, resource management in clouds: a mechanism design approach,”
pp. 89–103, Jun. 2015. IEEE Trans. Cloud Comput., vol. 3, no. 3, pp. 247–260, Jul.–Sep. 2015.
[20] X. Lyu, H. Tian, C. Sengul, and P. Zhang, “Multiuser joint task off- [45] X. Wang, X. Wang, H. Che, K. Li, M. Huang, and C. Gao, “An
loading and resource optimization in proximate clouds,” IEEE intelligent economic approach for dynamic resource allocation in
Trans. Veh. Technol., vol. 66, no. 4, pp. 3435–3447, Apr. 2017. cloud services,” IEEE Trans. Cloud Comput., vol. 3, no. 3, pp. 275–
[21] X. Chen, “Decentralized computation offloading game for mobile 289, Jul.–Sep. 2015.
cloud computing,” IEEE Trans. Parallel Distrib. Syst., vol. 26, no. 4, [46] S. Chichin, Q. B. Vo, and R. Kowalczyk, “Towards efficient and
pp. 974–983, Apr. 2015. truthful market mechanisms for double-sided cloud markets,”
[22] N. R. Devanur, C. H. Papadimitriou, A. Saberi, and V. V. Vazirani, IEEE Trans. Serv. Comput., vol. 10, no. 1, pp. 37–51, Jan.-Feb. 2017.
“Market equilibrium via a primal–dual algorithm for a convex [47] Technical report. Available (2018. May). [Online]: https://arxiv.
program,” J. ACM, vol. 55, no. 5, Nov. 2008, Art. no. 22. org/abs/1805.02982

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY KURUKSHETRA. Downloaded on April 25,2022 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 NGUYEN ET AL.: PRICE-BASED RESOURCE ALLOCATION FOR EDGE COMPUTING: A MARKET EQUILIBRIUM APPROACH 317

Duong Tung Nguyen received the BSc degree Vijay K. Bhargava (S’70-M’74-SM’82-F’92-LF’13)
in electronics and telecommunications from the received the BASc, MASc, and PhD degrees from
Hanoi University of Science and Technology, Queens University, Kingston, in 1970, 1972, and
Hanoi, Vietnam, in 2011, and the MSc degree in 1974, respectively. He is currently a professor
telecommunications from the Institut National de with the Department of Electrical and Computer

du Que
la Recherche Scientifique, Universite
bec, Engineering, The University of British Columbia,
Montre
al, QC, Canada, in 2014. Currently, he is Vancouver, where he served as the Department
working toward the PhD degree in the Depart- Head from 2003 to 2008. Previously, he was with
ment of Electrical and Computer Engineering, the Indian Institute of Science from 1974 to 1975,
University of British Columbia, Canada. His the University of Waterloo in 1976, Concordia Uni-
research focuses on networking and economics. versity from 1976 to 1984, and the University of
He is a student member of the IEEE. Victoria from 1984 to 2003. He has held visiting appointments with the
Ecole Polytechnique de Montreal, the NTT Research Laboratory, the
Tokyo Institute of Technology, the University of Indonesia, The Hong Kong
Long Bao Le (S’04-M’07-SM’12) received the University of Science and Technology, Tohoku University, and Friedrich
BEng degree from Ho Chi Minh City University of Alexander University, Germany. He is currently a honorary professor with
Technology, Vietnam, in 1999, the MEng degree the University of Electronic Science and Technology of China, Chengdu,
from Asian Institute of Technology, Pathumthani, and a gandhi distinguished professor with IIT Bombay. He is currently with
Thailand, in 2002, and the PhD degree from the the Institute for Scientific Information Highly Cited list. He served as the
University of Manitoba, Winnipeg, MB, Canada, Founder and president of Binary Communications Inc., from 1983 to 2000.
in 2007. He was a post-doctoral researcher with He has co-authored/co-editor of seven books the latest of which is
the University of Waterloo from 2007 to 2008, Wireless-Powered Communication Networks (Cambridge University
and the Massachusetts Institute of Technology Press, 2016). He is a fellow of The Royal Society of Canada, The Canadian
from 2008 to 2010. Since 2010, he has been with Academy of Engineering and the Engineering Institute of Canada. He is a
the Institut National de la Recherche Scientifique foreign fellow of the National Academy of Engineering, India, and has
(INRS), Universite
du Que
bec, Montre
al, QC, Canada, where he is cur- served as a distinguished visiting fellow of the Royal Academy of Engineer-
rently an associate professor. He co-authored the books Radio ing, United Kingdom. He has received numerous awards for his teaching,
Resource Management in Multi-Tier Cellular Wireless Networks (Wiley, research, and service to the IEEE. The latest awards are the Killam Prize in
2013) and Radio Resource Management in Wireless Networks: An Engi- Engineering awarded by the Canada Council for the Arts and the Humboldt
neering Approach (Cambridge University Press, 2017). His current Research Prize awarded by the Alexander von Humboldt Foundation of
research interests include smartgrids, cognitive radio, radio resource Germany. A long-time Volunteer of the IEEE, he has served as a director of
management, network control and optimization, and emerging enabling Region 7, in 1992 and 1993, a vice president of Regional Activities Board-
technologies for 5G wireless systems. He is a member of the Editorial RAB (now MGA), from 1994 to 1995, a president of the Information Theory
Board of the IEEE Transactions on Wireless Communications and the Society in 2000, a president of the IEEE Communications Society, in 2012
IEEE Communications Surveys and Tutorials. He has served as a Tech- and 2013, and he is currently the director for Division III, in 2018and 2019.
nical Program Committee chair/co-chair for the several IEEE conferen- He has served as an editor of the IEEE Transactions on Communications
ces including the IEEE WCNC, the IEEE VTC, and the IEEE PIMRC. He and as the editor-in-chief of the IEEE Transactions on Wireless Communi-
is a senior member of the IEEE. cations. He is a life fellow of the IEEE.

" For more information on this or any other computing topic,

please visit our Digital Library at www.computer.org/csdl.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY KURUKSHETRA. Downloaded on April 25,2022 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.