Resource Management in Cloud Networking Using Economic Analysis and Pricing Models: A Survey

2.1 MB·

by  Nguyen Cong Luong

1
Resource Management in Cloud Networking Using
Economic Analysis and Pricing Models: A Survey
Nguyen Cong Luong, Ping Wang, Senior Member, IEEE, Dusit Niyato, Fellow, IEEE, Wen Yonggang, Senior
Member, IEEE, and Zhu Han, Fellow, IEEE
Abstract—This paper presents a comprehensive literature re- with each other through high speed networks. For example,
viewonapplicationsofeconomicandpricingmodelsforresource the EU-funded Scalable and Adaptive Internet Solution (Eu-
management in cloud networking. To achieve sustainable profit
funded SAIL) project [9] investigates a combination of cloud
7 advantage,costreduction,andflexibilityinprovisioningofcloud
computing infrastructure and networking capabilities, called
1 resources, resource management in cloud networking requires
0 adaptiveandrobustdesignstoaddressmanyissues,e.g.,resource cloud networking. Cloud networking actually considers the
2 allocation, bandwidth reservation, request allocation, and work- network beyond the data centers with the aim of providing
loadallocation.Economicandpricingmodelshavereceivedalot bothon-demandcomputingandnetworkresources.Withcloud
n
of attention as they can lead to desirable performance in terms
networking, the resources and services can be provisioned
a
of social welfare, fairness, truthfulness, profit, user satisfaction,
J from interconnecting distributed data centers owned by one
and resource utilization. This paper reviews applications of the

8 economicandpricingmodelstodevelopadaptivealgorithmsand or multiple providers, called cloud data center networking.
protocolsforresourcemanagementincloudnetworking.Besides, The cloud resources and services can also be integrated with

] we survey a variety of incentive mechanisms using the pricing mobile networks, i.e., mobile cloud networking. Moreover,
T
strategiesinsharingresourcesinedgecomputing.Inaddition,we
edge computing models are deployed in cloud networking to
G consider using pricing models in cloud-based Software Defined
bring the cloud resources and services close to users, and
. Wireless Networking (cloud-based SDWN). Finally, we highlight
s important challenges, open issues and future research directions thusminimizeoverallcosts,jitter,latencies,andnetworkload.
c
of applying economic and pricing models to cloud networking. The aforementioned models of cloud networking along with
[
Keywords- Cloud networking, resource management, pricing the integration of the Software-Defined Networking (SDN)

1 models, economic models.
technologyareexpectedtosupportandsatisfyalargenumber
v
of users and applications in terms of flexibility, cost and
3
6 I. INTRODUCTION availability of the services.
9 However,managingnetworkandcloudresourcestogetherin
Cloud computing is becoming the platform of choice for a
1 cloudnetworkinghasmanychallenges.Itiscrucialtohavean
0 number of applications due to the advantages of high com- integrated view of the existing physical and virtual topologies
. puting power, low service cost, high scalability, accessibility,
1 and characteristics of the resources, as well as the status of
0 andavailability.Cloudcomputingisusedasanintegralpartof all network entities. Besides, the provisioning and placement
7 society in various domains and disciplines such as education
of virtual resources must be done in the best way possible,
1 [1], commerce [2], health care services [3], transportation [4],
taking into account the available resources of both the cloud
v: andsocialnetworks[5].Cloudcomputingisexpectedtobring and networks. Moreover, the reconfigurations must often be
i huge new revenue opportunities. Recent reports have showed
X performed to resize or release the existing virtual resources
that the global revenue generated from cloud services is more
due to, e.g., the dynamic network environments (node or link
r
than $200 billion in 2016, and there will be approximately
a failure), and the variability/elasticity of resource demand. In-
3.6 billion Internet users accessing cloud services by 2018
efficient resource management negatively affects performance
(http://www.statista.com). On-demand services provided by
and cost as well as impairing system functionality.
cloud computing include, for example, Software-as-a-Service
To address the aforementioned challenges, it is vital to
(SaaS), Platform-as-a-Service (PaaS), and Infrastructure-as-a-
develop resource management approaches which guarantee
Service (IaaS). Google Docs [6], Google App Engine [7],
the scalability, efficiency, manageability, adaptability and reli-
and Amazon’s Elastic Compute Cloud (Amazon’s EC2) [8]
ability for cloud networking. Traditional approaches, e.g., the
are among the popular commercial services available in cloud
system optimization, merely focus on the system performance
computing.
metrics given system parameters and constraints rather than
The cloud computing infrastructure is typically hosted in
economicfactors,e.g.,theprofit,cost,andrevenue.Therefore,
datacenters.Therefore,thecurrentcloudservicesarebasedon
economicandpricingapproacheshavebeenrecentlyexplored,
parallel implementations in distributed data centers connected
developed,andadoptedforresourcemanagementincloudnet-
working. Compared with the system optimization approaches,
N.C.Luong,P.Wang,D.Niyato,andY.Wen,arewithSchoolofComputer
Science and Engineering, Nanyang Technological University, Singapore. E- the economic and pricing approaches provide the following
mails:[email protected],[email protected],[email protected],and advantages:
[email protected].
Z.HaniswithElectricalandComputerEngineeringandComputerScience, • Incloudnetworking,theprofitsofcloudprovidershaveto
UniversityofHouston,Houston,TX,USA.E-mail:[email protected]. bemaximizedwhilemeetingtheuserdemands.Thus,the
2
profit guarantee for all cloud providers is a primary goal. motivates us to develop the survey with the comprehensive
Pricing models based on, e.g., the profit maximization or literaturereviewontheeconomicandpricingmodelsincloud
cost minimization, have been efficiently used to achieve networking.
the goal. For convenience, the related works in this survey are
• There are various actors/stakeholders in cloud network- classified based on various models of cloud networking and
ing which belong to different entities, e.g., end-users, then their issues as shown in Table II. The models of cloud
infrastructure providers, service providers, brokers, and networking considered in this survey are cloud data center
network operators. They have different objectives, e.g., networking, mobile cloud networking, edge computing, and
the profit, revenue, cost and utility, as well as different cloud-based Video-on-Demand (VoD) systems. Furthermore,
constraints, e.g., the budget and technology. Their objec- some pricing approaches for the resource management in
tives often conflict with each other, and this makes eco- cloud-based Software Defined Wireless Networking (cloud-
nomicandpricingmodelsbecomeeffectivetoolsincloud based SDWN) are discussed. Advantages and disadvantages
networking. More specifically, through the use of negoti- of each approach are highlighted.
ation mechanisms, economic and pricing approaches can The rest of this paper is organized as follows. Section II
determineoptimalsolutionsforselfishentitiesgiventheir describes a general architecture of cloud networking. Sec-
constraints. tion III introduces the fundamentals of economic and pricing
• The demand for cloud computing and network resources models. Section IV discusses how to apply economic and
depends on many users’ attributes, e.g., the willingness pricing models for resource management in cloud data center
to pay and performance requirements. Pricing strategies networking such as bandwidth, request, and workload alloca-
which rely on the demand elasticity such as price dis- tion.Applicationsofeconomicandpricingmodelsforresource
crimination have been recently used as ideal solutions allocation in mobile cloud networking are given in Section V.
to optimize the provisioning of resources and profits of Section VI reviews economic and pricing models to address
providers. issues concerning the bandwidth allocation, task allocation,
• VideoonDemand(VoD)undoubtedlyisamongthemost and storage sharing in edge computing. Section VII considers
importantservicesincloudnetworking.Severalcommer- economic and pricing approaches for bandwidth allocation
cial video delivery services have been introduced and and Peer-to-Peer (P2P) caching in cloud-based VoD system.
becomepopular,e.g.,YouTubeandNetflix.However,the In addition, applications of economic and pricing models for
bandwidthcostoftheserviceistypicallyverysignificant. bandwidth allocation and mobile data offloading in cloud-
Pricing mechanisms, e.g., smart data pricing, have been based SDWN are given in Section VIII. We outline important
applied to regulate the user demands and maximize the challenges, open issues, and future research directions in
bandwidth utilization. Section IX. Finally, we conclude the paper in Section X. The
• Besides the high bandwidth utilization for providers, listofabbreviationsappearedinthispaperaregiveninTableI.
guaranteeingqualityofservice(QoS),e.g.,asmalldelay,
for users is very important. Pricing approaches provide
II. GENERALARCHITECTUREOFCLOUDNETWORKING
very efficient solutions for the joint optimization of both
A. Definition of cloud networking
providers and users.
• To reduce service delay for users, cloud networking has The term cloud networking is understood in a multi-
developededgecomputingmodelswhichemploydevices administrative domain scenario in which network and data
at network edges to provide closer cloud resources and center domains interact with each other through predefined
services to users. Pricing and payment strategies stimu- interfaces [20], [21]. Specifically, cloud networking extends
late users to use the edge resources rather than distant networkvirtualizationbeyondthedatacenterstoprovidecloud
data centers while still guaranteeing profits for cloud andnetworkresourcestoclients/users.Networkresourcescan
providers. bevirtualrouters,bandwidth,virtualfirewalls,oranynetwork
management software.
The definition also shows a key difference between the
Althoughthereareseveralsurveysrelatedtocloudnetwork-
cloud networking and traditional computer networks, that is
ing, they do not focus on economic and pricing approaches,
the network virtualization. By using network virtualization,
whichareemergingasapromisingtool.Forexample,asurvey
the cloud networking reduces the cost for both providers and
of applications of network virtualization for cloud computing
clients through real-time, on-demand resource and service
wasgivenin[10].ThesurveyoftechnologiesoftheNetwork-
provisioning. The resources are assigned and used by the
as-a-Service (NaaS) paradigm for supporting network-cloud
client’s needs, and the client only pays for what is used [22].
convergence was presented in [11]. There are also surveys
On the contrary, resource allocation in traditional computer
related to the architecture of SDNs, e.g., [12], [13], [14],
networks is static, and a client needs to pay for every cost
[15], and applications of edge computing [16]. There are
regardless of whether the resource has been used or not.
surveysrelatedtothepricingapproaches,e.g.,[17],[18],[19].
However, they addressed the issues in Internet or wireless
B. Architecture of cloud networking
networksonly.Tothebestofourknowledge,thereisnosurvey
specificallydiscussingtheuseofeconomicandpricingmodels The goal of a cloud networking architecture is to enable
to deal with resource management in cloud networking. This an efficient composition of cloud and network resources in a
3
TABLEI
users.Netflix(https://www.netflix.com/)isanexampleof
MAJORABBREVIATIONS
a service provider of video on demand.
Abbreviation Description • Cloud service broker: A cloud service broker (or broker
forthesakeofshortness)actsasanintermediarybetween
BBU BaseBandprocessingUnits
cloud users/end-users and cloud providers.
Cloud-RAN Cloud-RadioAccessNetwork
• End-users: The users generate resource and service re-
CAPEX CAPitalEXpenditure
questsorworkloadsthatneedtobeprocessedusingcloud
CWMSN Cloud-basedWirelessMultimediaSocialNetwork resources.
IaaS Infrastructure-as-a-Service
1) Cloud data center networking: A data center is a large
IoT InternetofThings group of networked computer servers which are capable of
MCN MobileCloudNetworking providing the remote storage, processing, or distribution of
MNO MobileNetworkOperator large amounts of data. We provide brief descriptions of the
components and resources in both intra- and inter-data center
NFV NetworkFunctionVirtualization
networking to which they will be referred in this survey.
NUM NetworkUtilityMaximization
OPEX OPerationalEXpenditure • Intra-data center networking: Intra-data center network-
ing refers to the interconnection between servers and
P2P Peer-to-Peer
storage resources through a networking system within
PaaS Platform-as-a-Service
a data center. The networking system includes virtual
RRH RemoteRadioHead switches,Top-of-Rack(ToR)switches,coreswitches,and
SaaS Software-as-a-Service non-broking switch.
SDN Software-DefinedNetworking – VirtualMachine(VM):VMisasoftwareprogramor
SDWN Software-DefinedWirelessNetworking operatingsystemwhichisabletoperformtaskssuch
as running applications and programs as a separate
SLA Service-LevelAgreement
computer [32]. Multiple VMs can exist within a
SP ServiceProvider
physical server or machine through virtualization
VCG Vickrey-Clarke-Groves
techniques. In cloud networking, a VM can be mi-
VM VirtualMachine gratedamongserverswithinadatacenterorbetween
VoD VideoonDemand data centers owned by different providers.
– Virtual switch: A virtual switch is generally a
software-based Ethernet switch function running in-
cloud environment. To achieve the goal, several architectures side a server. It can support Ethernet and/or IP
were proposed for cloud networking. They can be based on services and provide switching and routing context
intra-data center networking and inter-data center networking separation among tenants/users sharing the same
which are commonly called the cloud data center networking server.
[11],[20],[23],[24],[25].Theycanbebasedonmobilecloud – Network slicing: Network slicing allows compart-
networking [26], [27] or edge computing models [28] [29], mentalizing VMs of the same application into the
[30], [31]. same virtual networks [33] and guarantees virtual
Based on these architectures, we provide a general, unified resource isolation and virtual network performance.
architecture for cloud networking as shown in Fig. 1. The – ToR switch: A ToR switch supports Ethernet virtual
architecture has three major parts: (i) cloud data center net- LAN (VLAN) services or simple IP routing for the
working, (ii) mobile cloud networking, and (iii) edge comput- data center. The ToR switch aggregates Ethernet
ing. Their descriptions are given in what follows. Note that links from the servers. ToR switches are connected
these parts can be independent from each other. Stakeholders to one or two core switches in a data center.
or actors commonly participating in cloud networking are as – Non-blocking switch: A switch is called non-
follows. blocking if it is able to connect all ports such that
any routing request to any free output port can
• Cloud provider: A cloud provider, e.g, an IaaS cloud
be established successfully without interfering other
provider, owns and manages data centers and system
traffics.
software.
– Core switch: A core switch hosts multiple ToR
• Network provider: A network provider provides network
switches and large-scale virtual LAN services or
connectivities among data centers of cloud providers or
simple IP routing for the data center.
betweenend-usersanddatacenters.Incloudnetworking,
the network providers aim at cooperating with cloud • Inter-data center networking: Data centers can be inter-
providerstoallocatecloudnetworkresourcesandservices connected across the Wide Area Network (WAN) using
to end-users or cloud users. inter-data center networking. Some commonly referred
• Cloudtenant/clouduser:Acloudtenantcanbeaservice entitiesintheinter-datacenternetworkingareasfollows.
provider, an organization or an enterprise, which uses – Datacentergateway:Adatacentergatewayprovides
cloud resources to host applications offered to its end- connectivity among data centers and to Internet and
4
TABLEII
ATAXONOMYOFTHEAPPLICATIONSOFECONOMICANDPRICINGMODELSFORRESOURCEMANAGEMENTINCLOUDNETWORKING
(cid:97)
(cid:97)
(cid:97)(cid:97)Systemmodels Clouddata Mobilecloud Edge Cloud-based Cloud-based
(cid:97)(cid:97) centernetworking networking computing VoDsystem SDWN
(cid:97)
Designissues (cid:97)(cid:97)(cid:97) (SectionIV) (SectionV) (SectionVI) (SectionVII) (SectionVIII)
Bandwidthallocation (cid:88) (cid:88) (cid:88) (cid:88) (cid:88)
Resourceallocation (cid:88) (cid:88)
Taskallocation (cid:88)
Requestallocation (cid:88)
Workloadallocation (cid:88)
Storagesharing (cid:88)
P2Pcaching (cid:88)
Mobiledataoffloading (cid:88)
VPNcustomers.Thedatacentergatewaycanprovide users.
virtual routing and switching capabilities. A wide range of services is offered by MCN: (i) typical
– IP/MPLS network: An Internet Protocol/Multi- cloud computing atomic services, e.g., the computing, stor-
Protocol Label Switching (IP/MPLS) network is a age, and networking, (ii) support services, e.g., Monitoring
packet-switched network that employs the Internet as a Service (MaaS), (iii) virtualized network infrastructure
Protocol (TCP/IP) enhanced with the MPLS stan- services, e.g., Radio Access Network-as-a-Service (RANaaS)
dard. and Evolved Packet Core-as-a-Service (EPCaaS), (iv) new
– Resource pool: A resource pool is a collective set of virtualized applications and services, e.g., Content Delivery
resources in data centers. Networks-as-a-Service (CDNaaS), and (v) End-to-End (E2E)
– Federatedcloudnetworking:Federatedorfederation services. In particular, RANaaS allows to partially move
cloud networking refers to the cooperation among functionalities of RAN, i.e., digital processing functions, to
cloud providers to establish the federated cloud re- a data center depending on the actual needs and network
source. For the federated cloud networking, a cloud characteristics [35]. When all RAN functionalities are shifted
provider can “borrow” cloud resources from other towards the data center, and only RF functions are performed
providers if its own resources are overloaded. This atRemoteRadioHead(RRH),wehavetheconceptofCloud-
is called outsourcing. Also, a cloud provider can RAN or Centralized-RAN [36]. The RANaaS implementation
“rent out” its resources to other cloud providers if has the following major characteristics [37]:
its resources are free. This is called insourcing. • On-demand provisioning: Mobile network resources and
2) Mobile cloud networking: Mobile Cloud Networking servicesareprovisionedaccordingtothedemandelastic-
(MCN) is the EU FP7 Large-scale Integrating Project (IP) ity of mobile users.
(cordis.europa.eu/fp7/ict/future-networks). It focuses on inte- • Virtualization of RAN resources and functions: They aim
grating the cloud computing and network function virtual- at optimizing usage, management, and scalability of the
ization technologies to mobile networks [34]. MCN is able mobile network.
to provision services involving mobile network, decentralized • Resource pooling: This allows virtual operators to share
computing,andstorageasoneon-demandunifiedservice.The morededicatedresourcesandservices,andthusenabling
main characteristics of MCN are as follows [26]: more business opportunities.
• MCN improves the real-time performance of mobile • Elasticity: This characteristic enables scaling network
network functions, e.g., the baseband unit processing, resources at the data centers or controlling the number
mobility management, and QoS control, based on the of active RRHs.
high-performance cloud computing infrastructure. Thus, • Service metering: Operators provision and charge RAN
MCN enables adapting to the elasticity of the load. operation services, e.g., the usage of RRHs, on a mea-
• MCN provides an entirely new mobile cloud application surable and controllable basis.
platform as well as novel revenue streams for Telco by • Multi-tenancy: This feature ensures the security in the
orchestrating infrastructure and services across different mobile network by enabling isolation mechanisms and
domains including wireless, mobile core networks, and charging of different users.
data centers. 3) Edge computing: Edge computing is a paradigm which
• MCNhasthe3GPPLTEcompliantarchitecturetoexploit pushes the frontier of computing applications, data, and ser-
and support cloud computing. vices away from central nodes, e.g., the data centers, to the
• MCN introduces a new business actor, i.e., the MCN periphery or edges of the network [38]. Edge computing
provider, in addition to typical stakeholders, e.g., the covers a wide range of technologies including cloudlet, re-
cloud computing provider, application provider, and mote/micro/community clouds, nano data centers, volunteer
5
End-users End-users
IP/MPLS
Data
network Data
center
center
Cloudlet/
Data center edge devices
gateway
Core
Data Edge computing
switch
center ToR
switch
Storage
Cloud data
system center networking SDN
Virtual
controller
VM VM switch
Server VM Server VM
VM Virtual VM
machines
Core
ec network
iv
r
e
s a Mobile cloud
sA networking
RAN
Mobile users/
end-users
Fig.1. Ageneralarchitectureofcloudnetworking.
computing system, local cloud/fog computing, client-assisted control plane. The control and data planes are always coupled
cloud system, sensing networks, e.g., the wireless sensor net- and embedded in the same networking devices, e.g., switches
work and crowdsensing network, and distributed Peer-to-Peer and routers, to guarantee network resilience. However, such
(P2P). Edge computing has the following major advantages architecture is rigid and complex to manage and control [40],
[39]: [41]. The Software-Defined Networking (SDN) [13], [42],
[43] is an emerging networking paradigm towards simple and
• It significantly reduces the data traffic, cost, and latency
flexible network management for network operators. SDN
andimprovesQoSsincecloudresourcesandservicesare
is defined as “a network architecture where the forwarding
located close to users.
state in the data plane is managed by a remotely controlled
• It alleviates the major bottleneck and the risk of a poten-
plane decoupled from the former” [13]. In other words, SDN
tial point of failure since it does not rely on centralized
decouples the control plane from the network devices to
computing.
become an external entity, the so-called SDN controller. The
• It enhances security since data is encrypted as the data is
SDN architecture has four major features below:
moved towards the network edge.
• It provides high levels of scalability, reliability, and • The control and data planes are decoupled, and network
automation.
devices just act as forwarding elements.
4) Software-Defined Networking (SDN): A traditional net- • Forwarding decisions are flow-based instead of
workarchitectureiscomposedofthreeplanesoffunctionality, destination-based, meaning that all packets in the
i.e., data, control and management planes. The control plane same flow receive identical service policies at the
makes forwarding/routing decisions on the user traffic based forwarding devices. This allows to unify different types
on forwarding/routing tables. The data plane is responsible of network devices, e.g., routers, switches, firewalls,
for forwarding the user traffic using the decisions from the load-balancers, and traffic shapers.
6
• The control logic is moved to an external entity, i.e., the costofcloudserverusage,arethevariablecostforgenerating
SDNcontrollerortheNetworkOperatingSystem(NOS). cloud services. The advantage of the cost-based pricing is the
• The network is programmable through software applica- ease of setting the price since the price is a function of the
tions running onthe SDN controller which interactswith internal cost, i.e., the cost required to generate the service
the underlying data plane devices. [52].However,thispricingstrategydoesnotconsiderexternal
These features of SDN make the networks more pro- market factors, e.g., the pricing strategies of other providers
grammable and easily partitionable and virtualizable. In prac- and the perceive value and willingness to pay of buyers.
tice, SDN has been used to address many issues in a wide In cloud networking, the cost-based pricing has been used
range of network environments [44]. For example, it was by cloud providers for evaluating the service cost in geo-
used to address the security and resource allocation in en- diverse data center networks [53], [54]. It has been also
terprise networks [45], [46], flow control, virtual data center employed to analyze the cost saving when SDN and the Net-
embedding, and resource utilization maximization in cloud work Function Virtualization (NFV) in the cloud are enabled
networking [24], [47], [48], mobility management and load [55], [56]. However, the internal cost information in cloud
balancing in wireless access networks [49], wavelength path networking may not be easy to obtain due to the variable cost
control and QoS-aware unified control in optimal networks diversity. For example, the variable cost could depend on the
[50], and network management in home and small business geography of data centers. Further details on the cost-based
networks [51]. In particular for the cloud networking, using pricing model can be found in [57], [58].
SDN makes a number of network devices become simple 2) Differential pricing: The cost-based pricing ignores the
forwarding elements which are cheap and easy to deploy. requirements and preferences of cloud users or tenants. To
This reduces both capital and operational expenditures for maximize the profit of providers, differential pricing, also
cloudandserviceproviders.Itisalsoexpectedtosignificantly called price discrimination, can be used. Consider a cloud
improve benefits for all stakeholders, especially when SDN resource market consisting of a cloud provider with its cloud
canbecombinedwiththeeconomicandpricingmodelswhich resources,i.e.,thecomputingresourceandnetworkbandwidth.
will be discussed in the next section. Using the differential pricing, the cloud provider may charge
differentpricestodifferentcloudusersbasedontheirdemand
III. OVERVIEWANDFUNDAMENTALSOFECONOMICAND and willingness to pay. By setting higher prices for one type
PRICINGTHEORYAPPLIEDINCLOUDNETWORKING of user, the use of the differential pricing actually transfers
Economic and pricing approaches have been applied to the user surplus to the provider. Here, the user surplus is the
addressmanyissuesincloudnetworkingduetotheaforemen- difference between the total money that users are willing to
tioned benefits. In this section, we classify the economic and payandthetotalmoneythattheyactuallypay.Thus,although
pricing approaches commonly used for resource management this pricing guarantees a high revenue for the provider, it
in cloud networking as shown in Fig. 2. The classification can be unfair to cause one type of user to pay a greater
is based on how the prices are set, i.e., market-based pricing, price than another type of user. In cloud networking, the
gametheoreticandauctionbasedpricing,andNetworkUtility differential pricing has been applied for bandwidth allocation
Maximization (NUM) based pricing. among groups of users having different elasticities on cloud
resources or among cloud users having different flexibility in
resource usage as proposed in [59]. In current cloud service
A. Market-Based Pricing
markets, the differential pricing is used to set prices of the
In the following, we present the pricing models based on
cloud services based on the requirements of the users. For
economic and financial concepts which have been applied in
example, the Alibaba group (https://intl.aliyun.com/) offers
thecloudnetworking.Wefirstpresentasimplepricingmodel,
lowerpricestouserswhichrequirethecloudservicesforlong
i.e., cost-based pricing, and then describe more complex pric-
term, e.g., 1 year.
ing models including differential pricing, profit maximization
3) Profit maximization: Profit maximization is the process
pricing, and Ramsey pricing.
ofdeterminingtheoutputquantityandthecorrespondingprice
1) Cost-based pricing: Cost-based pricing is a common
which yield the highest profit for a provider [60]. We present
pricing strategy to determine the price of a service based
briefly how to find the optimal quantity and price based on
on calculating the total cost of the service and adding a
the profit maximization in the following. Assume that a cloud
percentage of the cost as a desired profit. The objective of
provider needs to determine the number of cloud resource
using the cost-based pricing is to ensure that the price makes
(i.e., the computing and network bandwidth) units, denoted
the service provider profitable, or at least the price covers
by Q and the corresponding price P for their cloud users.
the total cost of the service provider. The total cost generally
The profit of the cloud provider is π = R(P,Q) − C(Q),
consists of a fixed cost and a variable cost. The fixed cost is
where R(·,·) is the total revenue and C(·) is the total cost.
the cost that does not change when the number of sales of the
The total cost may involve a fixed cost and a variable cost.
services changes. For example, hardware costs, e.g., servers
The revenue is the amount of money that the cloud provider
and network devices. On the contrary, the variable cost varies
receivesfromsellingQresourceunitstoitsusers.Theoptimal
accordingtothenumberofsalesoftheservicesproduced.For
quantity of cloud resource units, i.e., Q∗, is determined such
example,theresourcecosts,e.g.,energyandbandwidthcosts,
that the profit is maximized, i.e., Q∗ = maxπ. The optimal
thecostofdatatransferbetweendifferentdatacenters,andthe
Q
7
Economic and pricing models for resource
management in cloud networking
Game theoretic/auction NUM-based
Market-based pricing
based pricing pricing
Cost-based Differential Profit Ramsey- Non-cooperative Stackelberg Bargaining Auction Posted-price
pricing pricing maximization based pricing game game game theory mechanism
Fig.2. Ataxonomyofeconomicandpricingmodelsincloudnetworking.
quantity allows to find the optimal price based on the demand provider’s profit. Specifically, assume that a cloud provider
curve. The demand curve is typically a linear curve to show providescloudresourceunitsintwoindependentmarkets.The
the relationship between the price of a resource unit and the cloud provider determines different prices (p ,p ) in the two
1 2
quantity of resource units that users are willing to buy. A markets. In the independent market setting, the demands of
generaldemandcurvecanbeexpressedasP =a−bQ,where the resource corresponding to two prices are (q (p ),q (p )).
1 1 2 2
a and b are proper parameters. Thus, at Q=Q∗, the optimal The marginal cost of offering one cloud resource unit in both
price is P∗ =a−bQ∗. markets is c, and the cloud provider has a fixed cost. The
Theoptimalquantityandpricecanbedeterminedusingthe objective of the cloud provider is to determine (p ,p ) to
1 2
graph as shown in Fig. 3(a) and Fig. 3(b). Fig. 3(a) shows the maximizesocialwelfaresubjecttotheconstraintthattheprofit
curves of the total cost, total revenue, and profit. The optimal of the cloud provider is not less than a threshold. Here, the
quantity Q∗ is determined at the positive peak value of the social welfare is the consumer surplus which is the area on
profit curve. Then, the optimal price P∗ is obtained from the the left of the demand curve and above the price as shown
demand curve in Fig. 3(b). However, in real markets, it is in Fig. 3(b). The threshold is a fixed profit value π∗ which is
not easy to determine the demand curve. The user demand predefined by the cloud provider. Therefore, the optimization
can be random or assumed to follow some distributions, e.g., problem of the cloud provider can be formulated as follows:
theGaussian[61].Moreover,theprofitmaximizationdoesnot 2 (cid:32)(cid:90) ∞ (cid:33)
(cid:88)
consider the market competition in determining the quantity max q (p)dp (1)
i
and price. In cloud networking, the profit maximization has (p1,p2)i=1 pi
been adopted to allocate computing and network resources to 2
(cid:88)
users [62] or to assign resource requests from users to cloud s.t. (p −c)q (p )≥π∗.
i i i
providers [63]. i=1
P (price)
The problem in (1) can be solved using the Lagrange
Total cost a
multiplier method. The relationship between the optimal price
(C) Total
revenue and the demand elasticity in market i, i∈{1,2}, is expressed
(R) Consumer as follows:
surplus p∗−c 1−λ∗ 1
P* i = , (2)
(P*,Q*) P=a - bQ p∗i λ∗ (cid:15)i
Profit (π) where λ is the Lagrange multiplier, and (cid:15)i = dqdi(ppi∗i)pq∗ii is the
coefficient of price elasticity of demand for the resource in
Q* Q (resource Q* Q (resource
units) units) market i. The price elasticity of demand gives the percentage
(a) (b) change in quantity demanded according to a fall or a rise in
its price. We get the following relationship from (2):
Fig.3. Pricingbasedonprofitmaximization. p∗−c p∗−c (cid:15)
1 / 2 = 2. (3)
p∗ p∗ (cid:15)
4) Ramsey pricing: In Ramsey pricing, different prices 1 2 1
of the same commodity are applied to different markets The expression in (3) shows that the price of a resource in
depending on the demand elasticity of the commodity [64]. a market should be relatively low when the demand elasticity
Ramseypricingissimilartothedifferentialpricing.However, in the market is high. On the contrary, in a market with
unlike the differential pricing that maximizes the profit of inelastic or less elastic demand, the provider can set a high
the provider, Ramsey pricing aims to maximize the social price because the demand for the resource does not change
welfare of users subject to a predefined threshold on the significantly according to a fall or a rise in the price. Clearly,
8
the provider desires to have a market with inelastic demand strategies since the payoffs will be worse off. However, in
since it can increase the price to earn more revenue. How- some cases, there is no Nash equilibrium at all, or there may
ever, determining demand elasticity in markets is challenging. exist multiple Nash equilibria which can make players not
Moreover, the cloud provider cannot apply this pricing in the be clear about which one to choose. Therefore, checking the
long run since users charged with a higher price will seek existenceanduniquenessoftheNashequilibriumisimportant
alternatives [65]. One application of the Ramsey pricing in when setting prices based on the non-cooperative game.
cloud networking is to regulate traffic flows of users among The non-cooperative game theory has been widely used for
data centers [66], which will be discussed in Section IV-A8. theresourcemanagementincloudnetworking.Forexample,it
hasbeenusedtomodelthebandwidthsharingamongpeersin
cloud-assisted P2P streaming systems [71] or among brokers
B. Game Theory and Auction Based Pricing
in cloudlet systems [72]. It has been adopted to maximize the
Game theory and auctions are the study of multiparticipant
profits of cloud providers as presented in [73].
decision making problems in which a choice of a participant,
2) Stackelberg game: The non-cooperative game discussed
i.e., a player, potentially affects the interests of other partici-
above assumes that players announce their pricing strategies
pants[67].Inthecontextofcloudnetworking,participantscan
simultaneously, and the players know each other’s strategies
becloudproviders,serviceproviders,cloudtenants,andusers.
at the same time. However, this may not always hold in
In the following, we briefly present game theoretic models
real markets. Therefore, sequential games can be used in
and auction mechanisms which have been widely used to
whichplayerscanannouncetheirstrategiesfollowingacertain
determineresourcepricesinthecloudnetworking.First,some
predefined order. This is the Stackelberg game [74]. In the
important terminologies are defined below [68].
Stackelberg game, the player decides its own strategic choice
• Player: A player is a participant which makes a decision after observing the strategies of other players [75]. It was
in the game.
proved that even if the players have to choose their strategies
• Payoff: A payoff, i.e., a utility, a profit, or an interest, first, their payoffs are not less than those at the Nash equilib-
reflects the desired outcome of the player.
rium[76],i.e.,duetothefirst-moveradvantage.Thefollowing
• Strategy: Player’s strategy is a set of actions/instructions providesthedefinitionandpropertiesoftheStackelberggame.
that the player can follow to achieve a desired outcome.
Assume that there are two cloud resource sellers 1 and 2
The payoff depends on not only the player’s own action,
in the market. P and P are the sets of pricing strategies
1 2
but also the actions of others.
of sellers 1 and 2, respectively. Seller 1 chooses its pricing
• Rationality:Aplayerisrationalifitsstrategyalwaysaims strategy p from set P to maximize its payoff or profit
1 1
at maximizing its own payoff.
function π (p ,p ), and seller 2 chooses its pricing strategy
1 1 2
1) Non-cooperative game: In the non-cooperative game, p from set P to maximize its payoff function π (p ,p ).
2 2 2 1 2
each player maximizes only its own payoff neither being Without loss of generality, assume that seller 2 selects its
concerned about the payoff of the other players nor about the strategybeforeseller1decidesitsselection.Seller2isnamely
socialwelfareofthenetwork[69].Inthisgame,theplayersare the leader, and seller 1 is called the follower. We have the
selfish, and they do not form coalitions or make agreements following definition [77]:
with each other. Definition 1. If there exists a mapping F : P → P such
2 1
Consider a cloud resource market in which cloud providers that, for any fixed p ∈P , π (Fp ,p )≥π (p ,p ), ∀p ∈
2 2 1 2 2 1 1 2 1
asthesellerscompeteforsellingresourcestousers.Thesellers P , and if there exists p ∈P such that π (Fp ,p )≥
1 2s2 2 2 2s2 2s2
are typically selfish, and therefore the market can be modeled π (Fp ,p ), then the pair (p ,p ) ∈ P × P , where
2 2 2 1s2 2s2 1 2
as a non-cooperative game among the sellers along with their p =Fp , is called a Stackelberg strategy pair.
1s2 2s2
pricing strategies. Assume that there are N players, and P
i Definition 1 means that the Stackelberg strategy is optimal
is a set of pricing strategies of player i, where P = P ×
1 for the leader when the follower responds to the leader with
···×PN is the Cartesian product of the individual strategy thefollower’soptimalstrategy.LetD ={(p ,p )∈P ×P :
1 1 2 1 2
sets. Let pi ∈Pi be the pricing strategy of player i. A vector p = Fp } denote the rational reaction set of seller 1 when
1 2
of strategies of N players is p = (p1,...,pN), and a vector seller 2 chooses strategy p ∈ P . Seller 1 is referred to as
2 2
of corresponding payoffs is π = (π (p),...,π (p)) ∈ RN,
1 N a rational player. Similarly, when seller 1 is the leader, let
whereπi(p)isthepayoffofplayerigiventheplayer’schosen D denote the rational reaction set of seller 2. The sets D
2 1
strategy and strategies of the others. Each player chooses its and D have significant importance which is indicated in the
2
beststrategyp∗whichmaximizesitspayoff.Asetofstrategies
i following two propositions.
p∗ = (p∗,...,p∗ ) ∈ P is the Nash equilibrium if no player
1 N Proposition 1. A strategy pair (p1s2,p2s2) is the Stackel-
can gain higher payoff by changing its own strategy when the
berg strategy with seller 2 as the leader iff (p ,p )∈D
1s2 2s2 1
strategies of the others remain the same [70], i.e.,
and
∀i,pi ∈Pi :πi(p∗i,p∗i)≥πi(pi,p∗i), (4) π2(p1s2,p2s2)≥π2(p1,p2),∀(p1,p2)∈D1. (5)
where p = (p ,...,p ,p ,...,p ) is a vector of strat- Proposition 2. A strategy pair (p ,p ) is the Nash
i 1 i−1 i+1 N 1N 2N
egy choices of all players except player i. strategy pair iff (p ,p )∈D ∩D .
1N 2N 1 2
The inequality in (4) shows the stable state of the game The expression in (5) and Proposition 2 show that
in which the players have no incentive to change their own π (p ,p )≥π (p ,p ).Inotherwords,fortheleader,
2 1s2 2s2 2 1N 2N
9
the Stackelberg strategy guarantees to achieve the payoff at • Seller:Aseller,e.g.,acloudprovider,offersitsresources
least as good as the corresponding Nash equilibrium. This is and services for sale.
because when choosing the Stackelberg strategy, the leader • Auctioneer:Anauctioneeractsasanintermediateagentto
actually imposes a solution which will be favorable to itself. conduct an auction, determine, and announce the winner.
In cloud networking, the Stackelberg game has been ap- In many cases, an auctioneer is a seller itself.
plied for allocating the cloud provider’s bandwidth to virtual • Price: A price in an auction may be a bidding price or
networks[33]andforreducingtheaccessofuserstoserversin an asking price. The bidding price is the price that the
the cloud [78]. The Stackelberg game has been also applied bidder is willing to pay for a requested resource, and the
in cloud computing. For example, it was used to maximize asking price is the price of a resource that the seller is
revenueofthecloudproviderwhilemaximizingserverclients’ willing to offer.
utilities [79] or to maximize revenue of the cloud provider There exist several studies on auctions as well as their
while guaranteeing QoS for its end-users [80]. Besides, the applications. There are a survey of the auction theory [90],
Stackelberg game has been used in Internet of Things (IoT). a survey of auction on Internet [91], or a survey of auction
For example, it was adopted to maximize the profits of approaches for resource management in wireless networks
different participants of IoT industry value chain [81] and [92]. In what follows, we discuss typical types of auctions
to improve the QoS and the network’s robustness in sensing which have been commonly applied to resource management
networks [82]. in cloud networking.
3) Bargaining game: In the bargaining game or Nash bar- (a)Conventionalauctions:Aconventionalauctionisknown
gaining game, two or more players must reach an agreement astheopen-outcryauction.Intheopen-outcryauction,bidsof
regardinghowtodistributeamonetaryamount.Considertrad- buyers are disclosed to each other during the auction. There
ing bandwidth in cloud networking between a cloud provider, are two types of the conventional auction [93].
i.e., a seller, and a cloud tenant, i.e., a buyer. A successful • English auction: The English auction is an ascending-
bargain is reached if and only if the bandwidth is allocated at bid auction, meaning that the bidding price submitted
a mutually acceptable price. Let p0 be the smallest price that by buyers increases monotonically. Specifically, buyers
s
the seller can accept for selling the bandwidth and p0 be the submit their bidding prices for the resource sequentially
b
buyer’s greatest price that the buyer is willing to pay for the or simultaneously to the auctioneer. The auction will
bandwidth. The pair (p0,p0) is called the disagreement point terminate if there is no new higher price submitted.
s b
or threat point that the seller and the buyer expect to receive The buyer with the highest price wins the resource and
if their negotiations fail to reach a settlement [83]. pays the price p∗, i.e., a hammer price, which satisfies
The strategy of the seller is to offer the selling price p∗s p0s ≤ p∗ ≤ miaxBi, where p0s is the lowest price that
to maximize its expected profit π (p ,p0), i.e., π (p∗,p0) ≥ the seller can accept to sell, and B is buyer i’s budget.
s s s s s s i
π (p ,p0), ∀p . Similarly, the strategy of the buyer is to offer Generally,p∗changesdependingonthenumberofbuyers
s s s s
the buying price p∗ to maximize its profit π (p ,p0), i.e., in the auction and may not equal p0.
b b b b s
πb(p∗b,p0b)≥πb(pb,p0b), ∀pb. If p∗b ≥p∗s, a bargain is enacted • Dutch auction: Contrary to the English auction, the
and the transaction price for trading the bandwidth can be Dutch auction is a descending-bid auction in which the
set by [84], p∗ = kp∗ +(1−k)p∗, with 0 ≤ k ≤ 1. When auctioneer or seller initially sets a high asking price for
b s
k = 1/2, the transaction price is determined by splitting the the resource and then decreases the price until one of the
difference between the buyer’s and seller’s offers. A pair of buyersacceptstheprice.Thewinningbuyerpaysthefinal
the best response offer strategies (p∗,p∗) constitutes the Nash price and receives the resource. This simple allocation
s b
bargaining solution. At this agreement point, the seller earns enables the Dutch auction to spend less time than the
(p∗−p0), and the buyer earns (p0−p∗). English auction [94].
s b
Some other scenarios in cloud networking where the bar- (b) Vickrey and Vickrey-Clarke-Groves (VCG) auctions: Vick-
gaining game has been applied are allocating requests of rey and VCG auctions are the sealed-bid auctions in which
users to data centers [85] and sharing cloud resources among buyers submit simultaneously their sealed bids to the auction-
service providers [86]. In cloud computing, the bargaining eer. Different from the open-outcry auctions, in the sealed-bid
gamehasbeenusedfornegotiatingthepriceamongthecloud auction, buyers do not know bidding strategies of each other
resource brokers and grid service providers as proposed in and cannot change their own bids during the auction.
[87] and for pricing and allocating virtual resource instances • Vickrey auction: A Vickrey auction, also known as the
forindependenttasksandworkflowtasksaspresentedin[88]. second-price sealed-bid auction, is one of the two most
4) Auction: An auction is the economic mechanism the important k-th-price sealed-bid auctions. In the Vick-
goals of which are to allocate commodities and establish rey auction, the winning buyer pays the second-highest
corresponding prices via a process known as bidding [89]. price rather than the price that it submitted [95], i.e.,
There are some common terminologies used in the auction as p∗ = max p, where p is the highest price of the
i
follows: p∈P\{pi}
winner. In other words, the winner pays a price less than
• Bidder: A bidder is a buyer which wants to purchase its expected price [96]. Therefore, the Vickrey auction
resources. In cloud networking, bidders can be end-users motivates buyers to bid truthfully, and such an auction
or cloud tenants. achievesstrategy-proofness,orincentivecompatibility,or
10
e
c e
irP cirP
Bid 1
2 Bid 2 Market
b
P
equilibrium
Supply
... curve Supply
curve
b4 Bid 4 Ask 5
P (P*,Q*)
P* Ask 4 Bid 5 P*
4
Pa
... Demand
Demand
curve
1 Ask 1 curve
Pa
Qa1 Qb2 ... Q* Resource units Q* Resource units
(a) (b)
Fig.4. (a)Discretesupplyanddemandcurvesofdoubleauction,and(b)continuoussupplyanddemandcurvesfromeconomics.
truthfulness. The truthfulness is an important property auctionistomatchasksfromsellersandbidsfrombuyers
because an auction which does not hold this property by assigning commodities from the sellers to the buyers
may be vulnerable to market manipulation and produce and payments from the buyers to the sellers accordingly.
very poor outcomes [97]. To further understand the double auction process, we
• Vickrey-Clarke-Groves(VCG)auction:TheVCGauction consider a cloud resource market in which the sellers are
is a generalization of the Vickrey auction with multiple cloud providers and the buyers are cloud tenants/users.
commodities [98]. The VCG auction allocates commodi- Upon receiving bids and asks, the auctioneer sorts the
ties in a socially optimal manner and charges the winner buyers’ bids in a non-ascending order and the sellers’
with the loss of the social value due to its getting the asks in a non-descending order. The auctioneer finds the
commodity. largest index k at which the asking price is less than the
– Assume that there is a set T of M commodities biddingprice,i.e.,pa ≤pb.Thetransactionpricep∗,i.e.,
k k
for sale T = {t ,t ,...,t }, where t is the ith a hammer price or clearing price, can be determined as
1 2 M i
commodity, and a set of N buyers, i.e., bidders, p∗ =(pa+pb)/2. The buyer receives resources, and the
k k
B ={1,2,...,N}. seller gets payment p∗. The process is repeated to match
– Let b (t ) denote a bid of bidder i for commodity the remaining buyers and sellers as well as to determine
i j
t and VM denote the social welfare value, i.e., the corresponding clearing prices.
j N
social cost, created by M commodities. In fact, the double auction has a very similar concept
Similar to the Vickrey auction, if b (t ) is the highest, to the supply and demand model [100]. As shown in
i j
bidder i wins to obtain commodity t . According to the Fig. 4(a), the asks from sellers and the bids from buyers
j
VCG auction rule, bidder i pays the price which is equal form the supply and demand curves, respectively. The
to x-axis represents the supplied resource units, and y-axis
VM −VM\{tj}, (6) represents the asking or the bidding prices. For example,
N\{i} N\{i}
seller 1 sells Q units of resources at price P (i.e.,
a1 a1
whereVNM\{i}representsthesocialwelfarevalueifbidder “Ask 1” in Fig. 4(a)), and buyer 2 bids to buy Qb2 units
i does not participate in the auction, and VNM\\{{i}tj} is ofresourcesatpricePb2(i.e.,“Bid2”inFig.4(a)),andso
the attainable social welfare value after bidder i wins on.Fig.4(a)isactuallyadiscretizedformofthestandard
commodityt .Theexpressionin(6)indicatesthatwinner supply and demand model which is shown in Fig. 4(b).
j
i needs to pay the price which is the loss in attainable Thesupplyanddemandcurvesintersectatapointwhich
welfaresufferedbytheremainingbidderssinceitgotthe is called the supply-demand equilibrium [101], i.e., the
commodity t . marketequilibriumpoint(P∗,Q∗).TheclearingpriceP∗
j
(c) Forward, reverse, and double auctions: Considering the at the equilibrium can be determined by (Pa4 +Pb4)/2
sidesofsellersandbuyers,wecanclassifyauctionsasfollows: (as shown in Fig. 4(a))).
• Forward and reverse auctions: In the forward auction, The double auction can hold important properties: indi-
there are many potential buyers and one seller. On the vidual rationality (i.e., no participant loses when joining
contrary, in the reserve auction, there are one buyer and the auction), balanced budget (i.e., the auctioneer gains
many potential sellers. money),truthfulness(i.e.,buyersandsellerssubmittruth-
• Doubleauction:Inthedoubleauction,buyersandsellers fully their bids and asks), and economic efficiency (i.e.,
simultaneously submit their bids and asks to an auc- the social welfare is the best possible).
tioneer, respectively [99]. The basic idea of the double