Towards Optimal Deployment of Cloud-AssistedVideo Distribution Services

Jian He†, Di Wu†, Yupeng Zeng†, Xiaojun Hei‡, Yonggang Wen§

†Department of Computer Science, Sun Yat-Sen University, China‡Dept. of Electronics & Information Engineering, Huazhong University of Science & Technology, China

§School of Computer Engineering, Nanyang Technological University, Singapore{hejian9@mail2, wudi27@mail, zengyp@mail2}.sysu.edu.cn, heixj@hust.edu.cn, ygwen@ntu.edu.sg

Abstract—For Internet video services, the high fluctuation ofuser demands in geographically distributed regions results inlow resource utilizations of traditional CDN systems. Due tothe capability of rapid and elastic resource provisioning, cloudcomputing emerges as a new paradigm to reshape the model ofvideo distribution over the Internet, in which resources (suchas bandwidth, storage and etc.) can be rented on demand fromcloud data centers to meet volatile user demands. However, it ischallenging for a video service provider (VSP) to optimally deployits distribution infrastructure over multiple geo-distributed clouddata centers. A VSP needs to minimize the operational costinduced by the rentals of cloud resources without sacrificing userexperience in all regions. The geographical diversity of cloudresource prices further makes the problem complicated. In thispaper, we investigate the optimal deployment problem of cloud-assisted video distribution services and explore the best tradeoffbetween the operational cost and the user experience. We aim atpaving the way for building the next-generation video cloud. Tothis objective, we first formulate the deployment problem into amin-cost network flow problem, which takes both the operationalcost and the user experience into account. Then we apply theNash bargaining solution to solve the joint optimization problemefficiently and derive the optimal bandwidth provisioning strategyand optimal video placement strategy. In addition, we extend thealgorithms to the online case and consider the scenario whenpeers participate into video distribution. Finally, we conductextensive simulations to evaluate our algorithms in the realisticsettings. Our results show that our proposed algorithms canachieve a good balance among multiple objectives and effectivelyoptimize both operational cost and user experience.

I. INTRODUCTION

In the past decade, Internet video has become a very popularapplication over the Internet and accounts for the majorityof network traffic [1]. Internet video service providers (e.g.,YouTube, Hulu) generally resort to CDN (Content Distri-bution Network) to conduct large-scale video distribution.However, CDN solutions are inadequate for the emergingvideo traffic growth. First, due to their semi-static resourceprovisioning mechanism, the resource utilization of existingCDNs is extremely low (normally ranging between 5% -10%)[2], which directly translates into a high operational cost.Especially, significant oversubscription has been observed forflash-crowd situations [3] when a lot of users are trying toconsume the same video content. Second, the emerging user-generated video contents (e.g., Youtube, Ku6, etc) are long-tail nature [4], defying the operational principle of serving the

popular contents by CDNs. Therefore, new paradigms shouldbe developed in order to provide the capabilities of scaling upand down the provisioned resources in a dynamic manner andimprove the resource utilization ratio.

The emergence of cloud computing [5] opens a new doorfor designing the next-generation video distribution platform.Cloud-based services are more cost-effective, highly scalableand reliable. Recently, there have been quite a few researchworks (e.g., [6], [7], [8], [9], [10], [11], [12], [13]) onexploiting the cloud platform for media content distribution.Compared with conventional approaches, a unique feature ofcloud-assisted video distribution is the capability of rapidand elastic resource provisioning. Cloud resources, such asCPU, memory, storage and bandwidth, can be automaticallyallocated in a fine granularity to meet the demand from endusers timely. A video service provider can rent the distributioninfrastructure from the cloud service provider (CSP) in an on-demand manner, and avoid the long-term hardware investmentand low utilization due to resource over-provisioning.

However, for bandwidth-intensive services such as videodistribution, a number of challenging issues need to be ad-dressed before real deployment on cloud platforms. First, asusers are spreading over multiple regions, to improve user QoE(Quality of Experience) at different regions, a video serviceprovider should deploy its distribution service in multiple geo-graphically distributed data centers, which are possibly ownedby multiple CSPs with different pricing strategies. Thus, for aVSP, a set of critical questions need to be clearly answered:how should a video service provider minimize its operationalcost? how much resources (e.g., bandwidth) should be provi-sioned at each location? how should user requests be directedto different geo-distributed data centers so as to maximizeoverall user experience? Second, the popularities of videos tobe distributed are dynamic and evolutionary over time. Thus,the deployment of cloud resources is also a dynamic process.It means that a video service provider should adjust resourceprovisioning at different regions proactively and place videocontents according to the changes of user demands. We needto design online algorithms instead of offline algorithms tohandle demand fluctuation.

In this paper, we investigate the optimal deployment ofcloud-assisted video distribution services. To this end, we

propose a set of practical algorithms to address the abovechallenges and validate their effectiveness via extensive sim-ulations. Our main contributions in this paper can be summa-rized as follows:

• We explicitly formulate the problem of cloud deploymentusing network flow theory. The optimal deployment prob-lem can be transformed into the problem of finding themin-cost flow in a network flow model. We identify theconflict between optimizing bandwidth provisioning costand viewing latency and show the related Pareto curve.

• We investigate the joint optimization problem of clouddeployment and apply the concept of Nash bargainingsolution (NBS) to solve it efficiently. The obtained resultbears the properties of both optimality and fairness.

• We design an optimal bandwidth provisioning algorithmwith dual decomposition for one-shot optimization, whichcan be easily extended to the continuous online case.We also derive the associated optimal video contentplacement strategy.

• We further study the case when peer bandwidth resourcesat the edge can be exploited. We consider a locality-awarepeer-assisted design and derive the corresponding optimaldeployment strategy.

• We conduct extensive trace-driven simulations to validatethe effectiveness of our proposed algorithms in the real-istic settings. The experimental results indicate that ouralgorithms can achieve a good balance among multipleobjectives and optimize the operational cost and the userexperience simultaneously.

To the best of our knowledge, our work is the first to applyNash bargaining solution to optimize the bandwidth cost andthe latency cost jointly for cloud-assisted video distributionservices, which can achieve both optimality and fairness. Inaddition, our algorithm can be easily extended to incorporatethe peer contribution at the edge.

The remainder of this paper is organized as follows. SectionII reviews previous work in the area of Internet video distri-bution and cloud service deployment. Section III presents theformulation of the optimal cloud deployment problem usingnetwork flow theory. Section IV considers the optimizationof the bandwidth cost and the latency cost jointly when de-ploying cloud-assisted video services, and proposes an onlineNash bargaining algorithm to solve such a joint optimizationproblem. We also further explore the scenario when peersalso participate into video distribution in the local region.Section V proves the effectiveness of our proposed algorithmsby experimental evaluation. Finally, Section VI concludes thepaper and discusses the future work.

II. RELATED WORK

Internet video distribution has received a great amountof attention in the past decade. CDN (e.g., Akamai[14],Limelight[15]) and P2P (e.g., Coolstreaming[16], PPLive[17],MBoard[18]) are two widely adopted technologies for buildinglarge-scale video distribution platforms. However, CDN is tooexpensive for small video service providers, while P2P is

hard to guarantee viewing quality of end users and becomesinefficient when handling unexpected flash crowd[19]. XunleiKankan[20] and LiveSky[21] resort to a hybrid CDN-P2Pdesign to reduce cost without sacrificing user experience.However, there is still a big gap for designing an ideal cost-effective, highly scalable and reliable solution.

The emergence of cloud computing provides a promisingapproach to deliver QoS-guaranteed multimedia services eco-nomically [6]. Virtualization techniques can be exploited forresource allocation in a fine granularity to provide contentdistribution and multimedia processing service. Initial attemptshave been made by academic researchers (e.g., [7], [8]) andindustrial practitioners (e.g., [22], [23]) to migrate VoD appli-cations to the cloud. Niu et al. [24], [9] further investigatedthe pricing strategies for VoD service providers and automaticbandwdith allocation to satisfy user demands. In terms oflive media streaming, Wang et al. [10] presented a genericframework called CALMS for the migration of live streamingservices to the cloud, which adaptively leases and adjusts cloudresources to meet dynamic user demands. CloudStream [11]considered realtime transcoding of videos in different qualitiesover cloud and delivered video streams via a cloud-based SVCproxy. Chen et al. [12] investigated the problem of placingWeb server replicas in storage clouds to provide cost-effectiveCDNs. A few recent research work [13], [25], [26], [27] alsoinvestigated the distribution of long-tailed and highly dynamicsocial media contents over the cloud, and studied how topartition the social contents and conduct resource provisioningmore efficiently.

Our work differs from previous work in the followingaspects: First, instead of simply defining user latency as aconstraint, we focus on the optimization of provisioning costand user experience jointly, considering that VSPs in differentscales have different optimization preferences. In addition,we consider non-linear cost functions in our problem, whichare more generic but incur higher complexity in problemsolving. Second, we consider the deployment policy when peerresources can be exploited. The proposed algorithm can beeasily extended to meet dynamic user demands. Our workis also different from conventional research work on trafficengineering (e.g., [28], [29], etc), which focus on balancingnetwork traffic from the perspective of ISPs. Instead, ouralgorithms are more tailored to optimize the deployment ofcloud servers from the perspective of video service providers.

III. PROBLEM FORMULATION

A. Cloud-Assisted Video Distribution Services

In the cloud computing paradigm, a cloud serviceprovider(CSP) can own multiple geographically distributeddata centers and manage all the hardware/software resourcesin each data center for resource pooling. To provide videodistribution services, a video service provider(VSP) will rentresources (including CPU, bandwidth, storage) from one ormore CSPs, possibly with different resource pricing plans.The allocation of cloud resources is generally in the unit ofvirtual machine with predefined configurations. By adjusting

the number of virtual machines rented at each data center,a VSP can scale up and down the provisioned resources toquickly handle demand fluctuation in different regions. Fromthe perspecive of a VSP, it is challenging to determine howto provision resources at each location dynamically. Eachvideo request generated by end users will be routed to agiven data center according to certain routing policies. Suchrequest routing policies should be carefully designed so as towell balance the operational cost and the user experience. Forexample, if always routing a user request to a local data center,it may lead to a much higher bandwidth cost in spite that suchstrategy can provide good user experience.

Let us first consider a typical cloud-assisted video distribu-tion scenario. Suppose there is a VSP that needs to deliverN videos to its users spreading over K(K ≥ 1) geographicalregions. The video service provider is relying on cloud servershosted in M data centers to provide video services. Fig.1 provides an overview of a simplified cloud-assisted videodistribution platform.

...

DC1 DC2 DCM

CSP1CSP2

...

VSP

Region 1 Region 2 Region K

Fig. 1. A simplified cloud-assisted video distribution platform

To optimize the user experience, end users prefer to receivevideo streams from cloud servers with low latency (e.g.,cloud servers deployed in the same region). In case that localbandwidth supply is insufficient, a user can also receive videostreams from cloud servers at other regions but at the costof a higher latency. For a VSP, its objective is to optimizethe user experience at all regions and minimize its operationalcost at the same time. To this end, a VSP should determinehow much resources (e.g., bandwidth, storage, CPU) shouldbe purchased from different CSPs and how to deploy theseresources at different regions for better viewing quality.

B. Network Flow Model

Next, we start to study the deployment problem from asimple case and formulate the bandwidth provisioning probleminto a min-cost network flow problem. Consider a scenarioin which user demand for each video generated from everyregion is known a priori, the bandwidth provisioning problemcan be transformed into a network flow model illustrated inFig. 2. Note that the above assumption is only made forthe modeling purpose. In the later section, we will relax theassumption that user demand should be known a priori and

utilize prediction methods to predict future user demand forpractical implementation.

S T

...

...

...

...

...

...

...

DC1

DC2

R1

R2

RK

DC1/V1

DC1/VN

DC2/V1

DC2/VN

DCM/V1

DCM/VN

R1/V1

R1/VN

R2/V1

R2/VN

RK/V1

RK/VN

1d

2d

Kd

1

1d

Nd1

1

2d

Nd2

1

Kd

N

Kd

1

1b

Nb1

1

2b

Nb2

1

Mb

N

Mb

1b

2b

Mb

DCM

...

Fig. 2. Network flow model for cloud deployment problem

In Fig. 2, node S and T are two virtual nodes that representthe source and destination nodes, which are used to indicatethe total bandwidth supply from distributed data centers andthe total bandwidth demand generated by all users. The set ofdata centers involved in video distribution are represented bynodes {DCi, i = 1, · · · ,M}. The set of user regions are repre-sented by nodes {Rk, k = 1, · · · ,K}. To further illustrate thedemand and supply for each video in detail, we introduce a setof auxiliary nodes {DCi/Vj , i = 1, · · · ,M, j = 1, · · · , N}and {Rk/Vj , k = 1, · · · ,K, j = 1, · · · , N}. The flow bibetween S and DCi refers to the cloud bandwidth provisionedat the i-th data center, and bji between DCi and DCi/Vjrefers to the amount of bandwidth allocated to distributevideo j among bandwidth bi, which satisfies bi =

∑Nj=1 b

ji .

Similarly, dk denotes the total user demand generated fromthe k-th region and djk between nodes Rk and Rk/Vj denotesthe demand of a video j from the k-th region and satisfiesdk =

∑Nj=1 d

jk. The flow bjik between nodes DCi/Vj and

Rk/Vj represents the amount of bandwidth supplied by the i-th data center to the k-th region for distributing video j, whichsatisfies bi =

∑Nj=1

∑Kk=1 b

jik.

Each edge in the graph is also associated with a cost. Foredges between S and DCi(i = 1, · · · ,M), the associated costis mainly bandwidth cost. The unit price of bandwidth canbe different for data centers at different locations. Accordingto the pricing strategy of typical CSPs [29], we define thebandwidth cost function ψi(x) associated with a data centerDCi as a non-decreasing concave function, where x is theamount of the bandwidth flow supplied by DCi. The concavityproperty implies that the more bandwidth a customer buysfrom the CSP, the cheaper the unit price is. Such pricingpolicy can incentivize cloud customers to buy more bandwidth.For a VSP, one of its objective is to minimize its totalbandwidth cost, namely, BC =

∑Mi=1 ψi(bi). To take the

user experience into account, we also associate the edgesbetween data centers and regions with a cost, which representsthe distribution latency from a data center to users in aregion. Users prefer to receive video streams from data centersdeployed in the same region to better their viewing quality. To

Notation DescriptionM the number of geographically distributed data centersN the number of videos delivered by a VSPK the number of regions that users are spreading overψi(·) the bandwidth cost function associated with the i-th data centerωik(·) the latency cost function between the i-th data center and the k-th regionbjik the bandwidth supply from the i-th data center to the k-th region for distributing video jbi the total bandwidth supplied by the i-th data centerdjk the demand of video j generated by users in the k-th regiondk the total demand generated by viewers in the k-th region1ji an indicator to represent whether video j is stored in the i-th data centernk the number of viewers in the k-th region

TABLE INOTATION

this end, for each edge between nodes DCi/Vj and Rk/Vj , theassociated latency cost is assumed to be determined by a non-decreasing convex function ωik(·). Intuitively, ωik(·) stands forthe latency between a data center DCi and users in a regionRk, which is independent of distributed video Vj . In addition,ωik(·) is a function of the amount of the bandwidth flow froma data center DCi to users in region Rk. The intuition behindsuch definition is that the latency between a data center and auser region includes the propagation delay, the queuing delay,etc. The propagation delay is determined by the geographicaldistance, while the queuing delay is largely determined bythe amount of the bandwidth flow according to the queuingtheory. In spite that the propagation delay is independent of theamount of the bandwidth flow, the queuing delay will increaseas a non-decreasing convex function of the amount of thebandwidth flow. The convexity property of the latency costfunction follows the assumption made in [28] considering theincreasing the queuing delay when the amount of the flowincreases. Other edges are associated with a cost 0. Notationsused in this paper are listed in Table I.

In addition to bandwidth cost, the provisioning strategyshould also minimize overall latency cost at the same time.By incorporating both bandwidth cost and latency cost intothe network flow model, we can transform the deploymentproblem into a min-cost network flow problem, in which theoperational cost and the user experience are both taken intoaccount. Before delving into the joint optimization of bothcosts, we first consider the optimization of bandwidth costand latency cost individually.

Note that, different from computation-centric cloud ap-plications (e.g., MapReduce [30]), video distribution is abandwidth-centric service, in which bandwidth cost accountsfor the majority of deployment cost compared with storageand computation cost. Given the small fraction of storageand computation cost in the deployment of video distributionservices on the cloud, we focus more on optimizing bandwidthprovisioning in this paper. Anyway, our framework model canbe easily extended to take other types of resource costs intoaccount.

C. Bandwidth Cost-only Optimization

In the bandwidth cost-only optimization problem, denotedas P1, the objective of a VSP is to minimize the total

bandwidth cost (denoted by BC) while satisfying the totaluser demands in all regions. All the latency costs ωik(·) in thenetwork flow model are set to zero. It applies to the case whena VSP is sensitive to the operational cost but regardless of theuser experience. Based on network flow theory, the problemP1 can be formulated as follows:

P1: minimize BC =M∑i=1

ψi(K∑

k=1

N∑j=1

bjik)

s.t.M∑i=1

bjik ≥ djk, ∀k, j

bjik ≥ 0, ∀i, j, k (1)

By applying the nonlinear optimization theory [31], we canobtain the following theorem:

Theorem III.1 For the bandwidth cost-only optimizationproblem P1, the optimal bandwidth provisioning strategy is toonly let the i∗-th data center distribute all the videos, wherei∗ = argmini{ψi(D), i = 1, · · · ,M}, and D =

∑Kk=1 dk.

Proof: The concave property of bandwidth cost functioncan guarantee that ψi∗(x1) + ψi∗(x2) ≤ ψi∗(D) where x1 +x2 = D. Because of ψi(x) ≥ ψi∗(x) for any x ≥ 0 and i ̸= i∗,we have ψi1(x1) + ψi2(x2) ≤ ψi∗(x1) + ψi∗ ≤ ψi∗(D), forany other two data centers i1, i2 and any x1, x2 satisfying x1+x2 = D. Therefore, the strategy of letting the i∗-th data centerdistribute all the videos will incur the minimum bandwidthcost compared with other strategies.

The intuition behind the above theorem is that, if not takinglatency cost into account, a VSP prefers to only provisionbandwidth at the data center with the lowest bandwidth costand use one data center to serve video requests from all re-gions. For the implementation, as all bandwidth cost functionsψi(·) can be known beforehand, the problem P1 can be easilysolved by a central coordinator and we can find the data centerwith the lowest cost. Later, when receiving a video request, thecoordinator returns the address of that data center to the user.However, the above solution is absolutely not a good strategyas the user experience directly impacts the market share of avideo service provider.

D. Latency Cost-only Optimization

For the latency cost-only optimization, a VSP aims at opti-mizing the user experience at all regions without consideringthe bandwidth cost. In this case, we set all the bandwidthcost ψi(·) associated with edges in the network flow model tozero. Then the latency cost-only optimization problem P2 canbe formulated as follows:

P2: minimize LC =M∑i=1

K∑k=1

ωik(N∑j=1

bjik)

s.t.M∑i=1

bjik ≥ djk, ∀j, k

bjik ≥ 0, ∀i, j, k (2)

Accordingly, by solving the problem P2, we can obtain thefollowing theorem.

Theorem III.2 For the latency cost-only optimization prob-lem P2, the optimal bandwidth provisioning strategy B ={bjik,∀i, k, j} is given by

bjik =

{djk, i = i∗(k)0, i ̸= i∗(k)

where i∗(k) = argmini{ωik(dk), i = 1, · · · ,M}.

Proof: Because the remote latency cost is larger than thelocal latency cost, that is, ωik(x) ≥ ωi′k(x) for i = i∗(k),i′ ̸=i∗(k) and x ≥ 0. Then, we have ωi1k(x1) + ωi2k(x2) ≤ωi∗(k)(dk) for any i1, i2 and x1 + x2 = dk. Therefore, thebandwidth provisioning strategy B incurs minimum latencycost.

Intuitively, when only optimizing latency cost, the optimalprovisioning strategy adopted by a VSP should allow eachviewer to receive its full video stream from the data centerwith the lowest latency. In an extreme case, suppose that thereis one data center deployed at each region and the latency toa local data center is smaller than that to any remote datacenter, the optimal strategy to minimize overall user latency isto localize all the video traffic and only let local data centersto serve users in the same region.

The provisioning strategies derived from P1 and P2 can beconsidered as two extreme cases. In reality, the VSP needsto take both bandwidth cost and latency cost into account.However, these two objectives will be conflicting with eachother. For example, when we try to reduce latency cost byprovisioning more bandwidth at the local data center with ahigher bandwidth price, the total bandwidth cost of a VSPwill increase. Therefore, for a multi-objective optimizationproblem, we aim at finding a set of Pareto points, in whichno single objective can be improved without decreasing an-other. Among all the Pareto points, the optimal Pareto pointguarantees both Pareto optimality and fairness.

IV. JOINT OPTIMIZATION FOR CLOUD-ASSISTED VIDEODISTRIBUTION

In this section, we consider to optimize the bandwidth costand the latency cost jointly. To expose the tradeoff between t-wo conflicting objectives, we first define the objective functionas the simple weighted sum of the bandwidth cost BC andthe latency cost LC, where α is a non-negative scalar valuethat indicates the relative weight of two objectives.

P3: minimize BC + α · LC

s.t.M∑i=1

bjik ≥ djk, ∀j, k

bjik ≥ 0,∀i, j, k (3)

From the constraints of optimization problem P1, P2 andP3, we can find that their feasible solutions are the same.Without loss of generality, we define the feasible solution setas B and optimization variables B = {bjik, ∀i, k, j} ∈ B.

LC

BC

0

1

1

Pareto Curve

Feasible Solution

Set

Fig. 3. Pareto curve for the joint optimization problem of cloud-assisted videodistribution

By varying the value of α in the interval [0,∞) andnormalizing BC and LC to the interval [0, 1], we can plot atypical Pareto curve for the joint optimization problem P3 inFig. 3. It can be clearly observed that the optimal operationalpoints on the Pareto curve bear the Pareto property. For a givenα, our objective in the joint optimization problem is to findthe optimal operational point on the Pareto curve.

Actually, α can be removed from the objective function ofP3 if we multiply the unit latency cost by a factor of α in theproblem formulation. Then the objective function of ProblemP3 can be further simplified as BC+LC. In the next sections,we will use BC+LC to replace the original objective functionof P3 for simplicity.

A. Nash Bargaining Solution for Joint Optimization Problem

To solve the above joint optimization problem with con-flicting objectives, we adopt Nash bargaining solution [32]as a theory tool, which can ensure optimality and fairnesssimultaneously. Optimality refers to Pareto optimality, which

implies that the solution achieves the maximum reduction ofthe total cost starting from an initial point not on the Paretocurve (i.e., the disagreement point in the Nash bargainingprocess). While fairness guarantees that the tradeoff betweentwo conflicting objectives are balanced, which means that costminimization should benefit both objectives simultaneously.

In the bargaining process, we can imagine that there aretwo independent components with conflicting objectives whoare trying to cooperate with each other to achieve an optimaland fair operation point. We can find the optimal point ofBC + LC on the Pareto curve derived from the problem P3by solving the following optimization problem P4 (see below).The optimal solution obtained in P4 by the Nash bargainingprocess corresponds to the optimal solution in P3 since anyfurther reduction of the total cost is not possible. In that case,we can achieve the minimum total cost.

P4: maximize |BC −BC0| · |LC0 − LC|s.t. B ∈ B (4)

where (BC0, LC0) is the disagreement point that representsthe starting point of the negotiation. The problem P4 aimsat obtaining the optimal Pareto point that lies on the Paretocurve generated by solving P3. Denote BC0 as the minimumbandwidth cost , and LC0 as the maximum latency costunder the current demand respectively. From the conflictingproperty of the bandwidth cost and the latency cost, theminimum bandwidth cost also implies the maximum latencycost. Therefore, we can guarantee that the point (BC0, LC0)also lies in the feasible solution set and can be chosen as thedisagreement point. Therefore, the optimization problem canbe rewritten as follows:

maximize (BC −BC0) · (LC0 − LC) (5)

The maximization of (BC −BC0)(LC0 − LC) is a trans-formation of the original optimization problem. By utilizingthe Nash bargaining solution to obtain the optimal solution of(BC −BC0)(LC0 − LC), we are able to obtain the optimalsolution of the original optimization problem P3. From theconflicting property of BC and LC, if we decrease the valueof BC, then the value of LC will be increased. The Nashbargaining solution can achieve the optimal point (BC,LC)that maximizes the value of (BC −BC0)(LC0 − LC).

Define the total bandwidth supply from the i-th data centerto the k-th region as b̂ik =

∑Nj=1 b

jik and b̂ = {b̂ik, ∀k, i}.

Denote the bandwidth cost of the i-th data center as BCi =ψi(

∑Kk=1 b̂ik) and the latency cost associated with the i-th data

center as LCi =∑K

k=1 ωik(b̂ik). By applying the optimiza-tion decomposition theory[33], we can derive a complexity-efficient algorithm. The optimization problem (5) can berewritten as the following optimization problem since the logfunction does not change the optimal solution of the originalproblem:

maximize log(BC −BC0) + log(LC0 − LC)s.t.

∑Mi=1 b̂ik ≥ dk

b̂ik ≥ 0, ∀i, k (6)

Then, we apply dual decomposition to the problem (6), andget the following Lagrangian:

L(b̂,λ) = log(M∑i=1

ψi(K∑

k=1

b̂ik)−BC0)

+log(LC0 −M∑i=1

K∑k=1

ωik(b̂ik))

+K∑

k=1

λk · (M∑i=1

b̂ik − dk) (7)

where λk ≥ 0 is the Lagrange multiplier associated with thelinear bandwidth supply constraint.

Before stepping into further analysis of dual decomposition,we should analyze the property of the Lagrangian (7). Fromthe concavity of the bandwidth cost function and the convexityof the latency cost function, we can know that (BC − BC0)and (LC0−LC) are both concave functions. The compositionwith the log function does not change their concave property.Therefore, the Lagrangian is also a concave function. Themaximal value of the Lagrangian for a given λk is:

b̂⋆(λ) = argmaxb̂≥0

[L(b̂,λ)] (8)

The result of (8) is unique due to the concavity of theLagrangian. The dual problem of (6) is defined as follows:

minimize g(λ) = L(b̂⋆(λ),λ)

s.t. λ ≥ 0 (9)

The dual problem (9) can be solved by subgradient methodwith the following updates:

λk(t+ 1) = [λk(t)− β(M∑i=1

b̂⋆ik − dk)]+ (10)

where β is a sufficiently small positive step-size. Nash bar-gaining solution can solve the joint optimization problemP4 efficiently. Through dual decomposition, we can obtainoptimal results by utilizing subgradient methods. A centralizedalgorithm of optimal bandwidth provisioning is provided inAlgorithm 1.

In the centralized algorithm, the main complexity lies in thecomputation of the optimal result of (8). The updates of λk canbe disseminated in parallel. In the practical implementation,the components to optimize the bandwidth cost and the latencycost can achieve the best tradeoff by iteratively negotiating thevalue of λk.

If the number of videos is huge, it will be impractical toplace all videos at every data center. Based on the optimal

Algorithm 1 NBS-based Optimal Bandwidth ProvisioningAlgorithmInput:

M,N,K;Demand dk;Bandwidth cost function ψi;Latency cost function ωik;

Output:The optimal bandwidth provisioning b̂⋆ik.

1: Initialization step: set λk ≥ 0.2: Compute the initial disagreement point (BC0, LC0) based

on input.3: Compute the optimal value of (8), then obtain optimal

solutions b̂⋆(λ(t))4: Updates λk according to (10).5: Set t← t+1 and go to step 3(until satisfying termination

criterion).6: Return: b̂ik

bandwidth provisioning algorithm, we can derive the corre-sponding video placement strategy to minimize storage cost.The video placement algorithm is provided in Algorithm 2,which is inherently a greedy algorithm. In the algorithm, avideo with higher popularity is first placed at a data center withthe largest amount of available bandwidth for provisioning sothat the number of replicas can be minimized (i.e., the storagespace can be minimized).

Fig. 4 provides an example to illustrate how the greedyvideo placement algorithm works. In our example, supposethat there are 4 data centers and 4 videos to be distributed.Videos are sorted by their bandwidth demand from all regions.The decision on video placement is executed in multiplerounds. The left side of Fig. 4 shows the available bandwidthsupply of each data center, while the right side shows theunsatisfied bandwidth demand of each video. In the example,it takes 6 rounds to make the video placement decision. Ineach round, one video with the largest unsatisfied bandwidthdemand will be placed on the data center with the largestavailable bandwidth supply. The selected data center tries tosatisfy one video’s bandwidth demand as much as possible ineach round. It is possible to take multiple rounds to satisfythe bandwidth demand of one video, e.g., video 1. In the lastround, the demands of all videos have been satisfied and thenumber of videos placed at each data center has also beendetermined. In this example, both DC3 and DC4 should storetwo videos, while the other two data centers need to store onlyone video. The total number of video replicas is 6.

In the case that prices of unit storage space are homoge-neous across data centers, we have the following theorem:

Theorem IV.1 The greedy video placement algorithm shownin Algorithm 2 can minimize the overall storage cost.

Proof: Since the prices of unit storage space are homogeneousamong data centers, the problem of minimizing storage cost

Algorithm 2 Greedy Video Placement AlgorithmInput:

Optimal bandwidth provisioning results b̂⋆ik(i.e., outputfrom Algorithm 1)

Output:Video placement strategy 1j

i

1: Initialize 1ji = 0, ∀i, j

2: Sort data centers by their total bandwidth supply bi =∑Kk=1 b̂

⋆ik in the descending order, and generate a list of

data centers Ld.3: Sort videos by their total demand d̂j =

∑Kk=1 d

jk, and

generate a video list Lv .4: Choose the head of the list Lv to be stored in Ld(1).5: Set 1Lv(1)

Ld(1)= 1, j = Lv(1), and i = Ld(1).

6: for each region k do7: if the left demand of video j from region k djk > b̂⋆ik

then8: Update the left available supply b̂⋆ik = b̂⋆ik − d

jk, and

the left demand djk = 0.9: else

10: Update the left demand djk = djk − b̂⋆ik, and the leftavailable supply b̂⋆ik = 0.

11: end if12: end for13: if the left demand of video Lv(1) from each region is zero

then14: Delete the head of Lv .15: end if16: Sort the list Lv by updated total demand in decreasing

order.17: if all the bandwidth supply of Ld(1) have been used then18: Delete the head of Ld.19: else20: Sort the list Ld by supply in the descending order.21: end if22: if Ld or Lv is empty then23: Return 1j

i

24: else25: Back to step 4.26: end if

can be directly transformed to the problem of minimizingthe number of video replicas. We only need to prove thatthe number of video replicas generated by Algorithm 2 isminimum. For a data center, the video placement problemis actually a Knapsack problem with each video incurringthe same storage cost. In each round, the algorithm greedilychooses a video with the minimal unit cost of bandwidthsupply. The selection of a data center with the maximum leftavailable supply and a video with the maximum left unsatisfieddemand can guarantee the minimum property of the unit costof bandwidth supply. Denote the unit cost per unit bandwidthsupply in each round i as δi, then the total storage cost willbe 1

(∑

i1δi

)·(∑

j bj). The minimum δi guarantees the minimum

Initial State

Round 1

Round 2

Round 3

Round 4

Round 5

Round 6

Video Demand

Video 1 Video 2 Video 3 Video 4

Bandwidth Supply

DC 1 DC 2 DC 3 DC 4

Fig. 4. Illustration of our proposed video placement algorithm

total storage cost.

B. Online Algorithm for Joint OptimizationIn the above section, we only consider the optimization of

the joint cost function in the offline case, in which user demandin the future time slots can be known a priori by the algorithm.In this section, we extend our algorithm to a continuous onlinecase, in which the algorithm has no knowledge about thefuture but use prediction methods to predict user demand inthe future time slots. For each time slot, a VSP can adjustits deployment strategy at different data centers according tothe predicted demand. To predict the demand of videos, weutilize prediction techniques proposed in previous work [34],[35], [9]. The design of a new prediction method is beyondthe scope of this paper. In the online case, our objective isto minimize the joint cost function in each time slot and theprobability of under-provisioning. At the beginning of eachtime slot, we first predict the demand of videos in all regionsand then utilize Algorithm 1 to obtain the optimal bandwidthprovisioning and video placement strategies.

Assume that the total demand from each region can bepredicted. Let d̄k denote the random variable which representsthe actual total demand from the k-th region. The mean andvariance of d̄k are µk and σ2

k respectively. For convenience, letD = [d̄1, · · · , d̄K ], µ = [µ1, · · · , µK ] and σ = [σ1, · · · , σK ].Assume that all random variables d̄k follow Gaussian distri-butions d̄k ∼ N(µk, σk) (Note that the method can be easilyextended to other distributions).

We can utilize the ARIMA and GARCH models [36] tocalculate µk and σk based on the observed history of userdemands. For example, suppose the next time slot to be t anddenote the observed history of user demand from region k ashk = {dk(0), dk(1), · · · , dk(t − 1)}, where dk(τ), 0 ≤ τ ≤t− 1 is the observed user demand from region k by t. Then,

µk in the time slot t can be calculated based on hk by usingthe ARIMA model, and σk in the time slot t can be calculatedbased on hk by using the GARCH model. Similar techniquesare also used in [34] for the calculation of µk and σk.

To mitigate the probability of under-provisioning, the pre-dicted demand input dk should satisfy that the actual demandmust be met with high probability:

P (dk < d̄k) ≤ ϵ

where ϵ is a small constant. Because d̄k follows a Gaussiandistribution, we have

dk ≥ E[d̄k] + θ√var[d̄k] = µk + θσk (11)

where θ = F−1(1− ϵ).In the online case, the predicted demand is represented by

dk = µk+θσk. At the beginning of each time slot, we calculate(µk, σk) first so as to obtain the predicted demand dk. The on-line bandwidth provisioning strategy is illustrated in Algorithm3. Note that, the prediction accuracy affects the performance ofonline algorithms. However, existing prediction techniques canguarantee that the prediction accuracy is very high for normaldemand patterns (e.g., diurnal pattern of user demands). Inour later experiments, it is found that the impact is not verysignificant even for a lower prediction accuracy.

C. Optimization for Video Distribution with Peer Contribution

Even in the context of cloud-assisted video distribution,P2P technique is still useful to reduce the provisioning costand increase system scalability. If taking resources contributedby end users into account, the deployment strategy of a VSPshould be adjusted accordingly. To reduce distribution latency,we consider a locality-aware P2P distribution scenario, in

Algorithm 3 Online NBS-based Bandwidth Provisioning Al-gorithmInput:

M,N,K;µk and σk for random variable d̄k;Bandwidth cost function ψi;Latency cost function ωik;

Output:The optimal bandwidth provisioning b̂⋆ik.

1: Initialization step: set λk ≥ 0.2: Compute the predicted demand dk = µk + θσk.3: Compute the initial disagreement point (BC0, LC0) based

on predicted demand dk.4: Compute the optimal value of (7), then obtain optimal

solutions b̂⋆(λ(t))5: Updates λk according to (10).6: Set t← t+1 and go to step 4(until satisfying termination

criterion).7: Return: b̂ik

which P2P traffic is confined to the same region where aviewer resides.

Although peer contribution can reduce the amount of provi-sioned bandwidth, the efficiency of peer exchange is impactedby the number of peers in the same distribution swarm. Anefficiency function η(x) is defined to represent the fraction ofdemand that can be satisfied by peer contribution, where x isthe number of peers in the swarm. η(x) is assumed to be anon-decreasing function, which implies that the efficiency ofpeer exchange can be improved with more peers in the swarm.η(0) = 0, η(x) ∈ [0, 1] when x > 0. Suppose that there arenk users in the k-th region, if considering peer contribution,the optimization problem P4 is transformed to the followingproblem P5:

P5: maximize log(BC −BC0) + log(LC0 − LC)

s.t.M∑i=1

b̂ik ≥ dk(1− η(nk))

b̂ik ≥ 0, ∀i, k (12)

PredictorPredicted Demand

Calculator

Determine Peer

ContributionAlgorithm 1

Greedy Video

Placement CalculatorCollect Video

Viewing Information

kk,

kd))(1( kk nd

ikb

ji1

Actual Viewing Information

Viewing History

Algorithm 2

Algorithm 3 without Peer ContributionAlgorithm 3 with

Peer Contribution

kd

kd

))(1( kk nd

Fig. 5. Overview of the relationship among all the algorithmic components

As to the number of future users in each region, it can bederived by the predicted demand directly. The Lagrangian (7)is redefined as follows:

L(b̂,λ) = log(M∑i=1

ψi(K∑

k=1

b̂ik)−BC0)

+log(LC0 −M∑i=1

K∑k=1

ωik(b̂ik))

+

K∑k=1

λk · (M∑i=1

b̂ik − dk(1− η(nk)))

(13)

The subgradient update (10) is transformed to:

λk(t+ 1) = [λk(t)− β(M∑i=1

b̂⋆ik − dk(1− η(nk)))]+ (14)

The online NBS-based bandwidth provisioning strategy(shown in Algorithm 3) can still be utilized to solve theproblem P5. In the practical implementation, a tracker servercan be deployed at each region to maintain peer information(similar approaches have been adopted by [7], [10]). A usercan obtain video services from two resources: (1) cloud datacenters; or (2) other peers in the same region. To better utilizethe resources contributed by peers, a user will first query thetracker to obtain a peer list and seek to obtain video data fromother peers in the same region; only when the request can notbe served by local peers will it be routed to a cloud data center.

In the end, to have a big picture of our proposed algorithms,we illustrate the relationship among different algorithmiccomponents in Fig. 5. At the beginning of each time slot,the predictor makes prediction of video demands based on therecent viewing history according to Eqn (11). The predicteddemand serves as the input to Algorithm 1 to determine theoptimal bandwidth provisioning strategy. Next, Algorithm 2 isused to obtain the optimal video placement strategy. If takingpeer contribution into account, user requests are first directedto peers with available bandwidth. Only when a user requestcannot be served by peers will it be routed to a data center by ascheduler according to the derived optimal strategies. We havea module to determine peer contribution based on the swarmsizes and adjust the predicted demand accordingly. Finally,video viewing behaviors in the current slot are archived forfuture demand prediction.

V. EXPERIMENTAL EVALUATION

In this section, we conduct a series of simulation experi-ments to evaluate the effectiveness of our proposed algorithms.To be more realistic, we first use the datasets obtain fromPPLive to drive our simulation and compare with other alter-native strategies. More details about simulation experiments isavailable in our technical report [37].

A. Dataset Description

The PPLive VoD dataset consists of over 25 million server-side log records on Nov 7, 2010. Each log record containsinformation including the viewer ID, viewing starting time,viewing duration time, streaming rate, as well as the viewedvideo’s ID. We divide the whole day into 48 time slots, eachof which lasts for 30 minutes. The evolution of the numberof viewers in each time slot is illustrated in Fig. 6. Similar toprevious measurement results [17], we can see that the peakdemand happens at the night of that day.

1 5 10 15 20 25 30 35 40 45 480

2

4

6

8

10

12x 10

5

Time

Num

ber

of P

eers

Number of peers in each time slot

Fig. 6. Number of PPLive VoD viewers in each time slot

To obtain geographical distribution of viewers, we exploitMaxMind GeoIP [38] database to map users’ IP addressesto six geographic regions, denoted by R1, R2, · · · , R6. Thepercentages of users in each region are varying with time (asshown in Fig. 7) and a majority of users are from regions R1and R2, which are located in China. Note that the region R6is defined as a virtual region that includes viewers whose IPaddresses cannot be mapped successfully or from regions withtoo few viewers.

1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 48

10−3

10−2

10−1

100

Time

Per

cent

age

R1R2R3R4R5R6

Fig. 7. Time-varying geographical distribution of viewers

We set the locations of data centers based on the statisticsreported by [39]. The bandwidth prices at each location areretrieved from the websites of major ISPs or cloud providers(e.g., China Telecom [40], Amazon EC2 [41], etc.) in thecorresponding region. It is found that their pricing strategiesusually bear concave property. Table II shows the prices to

rent 100Mbps bandwidth per day in each region. Due to thespace limit, we haven’t listed the entire pricing plan of eachregion in our paper.

We assume that propagation latency is proportional to thegeographical distance between two regions. Similar to [26], wedefine propagation latency as 0.02∗dist(km)+5, where distis the geographical distance between any two regions and canbe obtained from Google Map [42]. The geographical distancebetween the region R6 and other regions are assumed to be1500km. As to the efficiency of peer exchange, the efficiencyfunction η(x) is defined as a non-decreasing concave functionwhere x is the number of viewers in a given region. In ourexperiments, η(x) is defined as min{⌊ x

300⌋ ·0.1, 1}. The valueof η(x) is within the range [0, 1] in any region.

For comparison, we also consider three other provisioningstrategies, which are listed as below:

• Centralized Strategy, in which all video requests areserved by a single data center, which has the lowestbandwidth cost;

• Local-only Strategy, in which video requests are onlyserved by local data centers with the lowest latency;

• Random Strategy, in which each video request is servedby a random data center regardless of its location.

B. PPLive Trace-driven Simulation

We first compare the normalized total cost BC + LC ofour proposed algorithms with other provisioning strategies inFig. 8(a). Due to the demand dynamics, the curves fluctuatewith time. It can be observed that NBS strategy can reducearound 80% of the total cost compared with the centralizedand random strategy. The reduction is about 50% comparedwith the local-only strategy. By considering peer contribution,we can further reduce the incurred total cost.

To analyze our proposed algorithms in detail, we study theincurred the bandwidth cost and the latency cost separately. Inorder to facilitate the comparison among different strategies,we introduce a new metric called normalized cost ratio, whichis defined as the cost incurred by the current strategy normal-ized by the sum of costs incurred by all strategies (includingthe current one). Fig. 8(b) illustrates the normalized bandwidthcost ratio under different bandwidth provisioning strategies. Asexpected, the local-only strategy incurs the largest bandwidthcost, and the optimal bandwidth cost can be achieved by thecentralized strategy. The NBS strategy can reduce around 10%of the bandwidth cost compared with the local-only strategy,and the NBS-P2P strategy can further reduce 5% ∼ 15%bandwidth cost.

Fig. 8(c) illustrates the normalized latency cost ratio underdifferent bandwidth provisioning strategies. Local-only strate-gy achieves the minimal latency cost. Our proposed NBS-P2Pstrategy can achieve comparable latency cost as that of thelocal-only strategy. NBS-based strategies can reduce latencycost significantly compared with centralized strategy and ran-dom strategy. More specially, the NBS strategy can reducearound 60% of latency cost compared with the centralized

Region R1 R2 R3 R4 R5 R6Prices($/100Mbps · day) 35.85 57.38 66.92 40.01 43.23 53.85

TABLE IIPRICES TO RENT 100MBPS BANDWIDTH PER DAY IN EACH REGION

1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 480

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

Time

Nor

mal

ized

Tot

al C

ost

Centralized StrategyLocal−only StrategyRandom StrategyNBS StrategyNBS−P2P Strategy

(a) Normalized total cost

1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 480.1

0.12

0.14

0.16

0.18

0.2

0.22

0.24

0.26

Time

Ban

dwid

th C

ost R

atio

Centralized StrategyLocal−only StrategyNBS StrategyRandom StrategyNBS−P2P Strategy

(b) Normalized bandwidth cost ratio

1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 480

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

Time

Late

ncy

Cos

t Rat

io

Centralized StrategyNBS StrategyRandom StrategyNBS−P2P StrategyLocal−only Strategy

(c) Normalized latency cost ratio

1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 480

0.020.040.060.08

0.10.120.140.160.18

0.20.220.240.26

Time

Tot

al C

ost R

atio

1.2

1.3

1.4

1.5

Onl

ine

Tot

al C

ost/O

fflin

e T

otal

Cos

tOffline AlgorithmOnline AlgorithmOnline/Offline Ratio

(d) Competitiveness of online algorithm

Fig. 8. Comparison of different bandwidth provisioning strategies (using PPLive trace-driven simulation)

strategy, and the NBS-P2P strategy can reduce around 65% oflatency cost compared with the NBS strategy.

We also evaluate the efficiency of our proposed onlinealgorithm by comparing it with the offline algorithm. Fig. 8(d)shows the efficiency ratio is within the range (1.25, 1.5).

C. Synthetic Trace-driven Experiments

The main purpose to conduct synthetic trace-driven ex-periments is to validate the applicability of our algorithmsin different scenarios, as the PPLive trace only representsone kind of user demand. In the section, we adopt moregeneral cost functions and run a set of synthetic trace-drivenexperiments in a larger scale.

In our experiments, we consider a synthetic scenario inwhich a video service provider relies on 30 geo-distributeddata centers on the cloud platform to distribute videos to usersspreading over 50 regions. The bandwidth cost function asso-ciated with the i-th data center is defined as ψi(x) = ρi · xγ ,where ρi is a scalar and γ ∈ (0, 1) is a factor to guarantee theconcavity of ψi(x). The latency cost function between a datacenter and a region is defined as ω(x) = µ · xν , where µ is ascalar to identify different cost functions for every data centerand region pair and ν is used to ensure the convex property ofω(x). For a local data center and region pair, the factor µ willbe less than that of a remote pair. To simplify our experiments,

ρi is defined as a random variable that is uniformly distributedin the range of [1,20], denoted by ρi ∼ U(1, 20) 1. We alsoassume that γ = 0.5, µ ∼ U(1, 5), ν = 1.5 in the experiments.However, it should be noted that any concave bandwidth costfunction and convex latency cost function are applicable in theexperiments. As to the efficiency of peer exchange, we definethe function η(x) in a similar way as that in the trace-drivenexperiments.

We follow a similar approach as [9] to generate syntheticvideo demands in each region. All regions are divided intothree classes, which represent user population in differentscales. The number of regions in each class and their cor-responding parameters are given as follows:

• Class I: 20 regions, with mean µi ∼ U(100, 500) andvariance σi ∼ U(10, 50);

• Class II: 20 regions, with mean µi ∼ U(500, 1000) andvariance σi ∼ U(10, 100);

• Class III: 10 regions, with mean µi ∼ U(1000, 1500)and variance σi ∼ U(10, 150);

Similar to the above section, we evaluate the efficacy ofdifferent bandwidth provisioning algorithms by comparing thetotal cost BC + LC. Fig. 9(a) illustrates the normalized totalcost resulted from different bandwidth provisioning strategies.

1U(x, y) means a uniform distribution in the range [x, y].

If not taking peer contribution into account, the NBS strategyachieves the optimality in all time slots. Due to the convexproperty of latency cost function, local-only strategy canobtain sub-optimal results compared with centralized strategyand random strategy. The NBS strategy can reduce about 49%of the total cost compared with local-only strategy. We alsoobserve that the NBS-P2P strategy can reduce about 51%of the total cost compared with the NBS strategy due tobandwidth resources contributed by peers.

We also analyze the percentage of local and remote band-width supply to understand the optimality of our proposedNBS-based algorithm. Fig. 9(b) shows the percentage oflocal bandwidth supply by all data centers in each time slotcorresponding to two NBS-based algorithms. Here, NBS refersto Algorithm 3, while NBS-P2P refers to Algorithm 3 consid-ering peer contribution. The ratio of local bandwidth supply ofthe NBS strategy is within the range (0.248, 0.256). By takingpeer contribution into account, the ratio of local bandwidthsupply can be further reduced to the range (0.242, 0.252).

To analyze the bandwidth and latency cost resulted fromdifferent bandwidth provisioning strategies, we calculate thebandwidth cost and the delay cost separately. Fig. 9(c) il-lustrates the normalized bandwidth cost ratio of differentprovisioning strategies. The centralized strategy incurs theleast bandwidth cost, which well confirms the correctnessof Theorem III.1. On the contrary, the local-only strategyperforms the worst in terms of bandwidth cost reduction.We observe that the NBS strategy can reduce around 10%of the bandwidth cost compared with the local-only strategy.Furthermore, the NBS-P2P strategy can reduce 5% ∼ 20% ofthe bandwidth cost compared with the NBS strategy.

Fig. 9(d) shows the normalized latency cost ratio of differ-ent provisioning strategies. Local-only strategy performs thebest in terms of latency cost reduction, which confirms thecorrectness of Theorem III.2. For latency cost, the reductionof two NBS-based strategies is close to the best strategy.Compared with the worst strategy, i.e., centralized strategy,the NBS strategy can reduce more than 70% of latency cost.Furthermore, the NBS-P2P strategy can reduce around 50%of latency cost compared with the NBS strategy.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 200.42

0.44

0.46

0.48

0.5

0.52

0.54

0.56

0.58

Time

Tot

al C

ost R

atio

1.14

1.16

1.18

1.2

1.22

1.24

1.26

1.28

1.3

Onl

ine

Tot

al C

ost/O

fflin

e T

otal

Cos

t

Offline StrategyOnline StrategyOnline/Offline Ratio

Fig. 10. The efficiency of online algorithm with prediction

Fig. 10 illustrates the competitiveness of our online NBS-based bandwidth provisioning algorithm compared with theoffline algorithm. In the offline algorithm, the demands can be

known beforehand at the beginning of each time slot; whilein the online algorithm, we need to predict the demand anduse the predicted demand as the input. We also present theratio between the costs incurred by online strategy and offlinestrategy in Fig. 10. In most cases, the cost of online algorithmis only 1.15 ∼ 1.3 times that of offline algorithm.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 201.8

2

2.2

2.4

2.6

2.8

Time

Rat

io o

f the

Num

ber

of V

ideo

Rep

licas

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 2010

12

14

16

18

Time

Rat

io o

f the

Num

ber

of V

ideo

Rep

licas

NBS Placement/Centralized PlacementNBS−P2P Placement/Centralized Placement

Full Placement/NBS PlacementFull Placement/NBS−P2P Placement

Fig. 11. Efficiency of Video Placement Strategy

To check the efficiency of our video placement algorithm,we also compare our algorithm (see Algorithm 2) with twoother video placement strategies: (1) Full placement, whichplaces all videos in each data center; (2) Centralized place-ment, which only places all videos in a single data center. Wecalculate the number of video replicas resulted from differentstrategies accordingly. Fig. 11 illustrates the ratio betweenthe number of replicas resulted from different algorithms. Forthe NBS strategy, its competitiveness is 1.8 ∼ 2.4 comparedwith centralized placement strategy, while the full placementstrategy incurs around 15 times the number of replicas resultedfrom NBS strategy. In our experiment, we find that the numberof videos stored in a data center will not exceed 20% of thetotal number of videos.

VI. CONCLUSION

The on-demand self-service feature of cloud computing en-ables a video service provider to provision its cloud resourcesadaptively for video distribution. In this paper, we study theoptimal deployment of cloud resources in multiple geograph-ically distributed data centers to improve the user experienceand minimize the operational cost simultaneously. We builda network flow model to formulate the cloud deploymentproblem and derive corresponding optimal bandwidth provi-sioning and video placement strategies with Nash bargainingsolutions. We also examine the case when peer contribution isconsidered. Our work in this paper provides useful guidelinesfor different video service providers to provision their serviceseffectively. In the future, we plan to investigate the deploymentover a hybrid cloud infrastructure with both public clouds andprivate clouds, and study how to better select data centers tooptimize video applications with different QoS requirements.

ACKNOWLEDGMENT

The authors appreciate the kindness of PPTV company forsharing PPLive traces.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 200

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

Time

Nor

mal

ized

Tot

al C

ost

Local−only StrategyCentralized StrategyRandom StrategyNBS StrategyNBS−P2P Strategy

(a) Normalized total cost

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 200.24

0.26

0.242

0.244

0.246

0.248

0.25

0.252

0.254

0.256

0.258

Time

Per

cent

age

of L

ocal

Ban

dwid

th S

uppl

y

NBSNBS−P2P

(b) Percentage of local bandwidth supply in NBS-based strate-gies

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 200.08

0.1

0.12

0.14

0.16

0.18

0.2

0.22

0.24

0.26

Time

Ban

dwid

th C

ost R

atio

Centralized StrategyNBS StrategyRandom StrategyLocal−only StrategyNBS−P2P Strategy

(c) Normalized bandwidth cost ratio

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

0.1

1

0.02

Time

Late

ncy

Cos

t Rat

io

Random StrategyNBS StrategyCentralized StrategyNBS−P2P StrategyLocal−only Strategy

(d) Normalized latency cost ratio

Fig. 9. Comparison of different bandwidth provisioning strategies (using synthetic trace-driven simulation)

REFERENCES

[1] “Cisco visual networking index: Forecast and methodology, 2010-2015,”http://www.cisco.com/en/US/solutions/collateral/ns341/ns525/ns537/ns705/ns827/white paper c11-481360 ns827 Networking SolutionsWhite Paper.html.

[2] M. Freedman, “Experiences with CoralCDN: A five-year operationalview,” in USENIX NSDI, 2010, pp. 7–7.

[3] P. Wendell and M. Freedman, “Going viral: Flash crowds in an openCDN,” in ACM IMC, 2011.

[4] M. Cha, H. Kwak, P. Rodriguez, Y. yeol Ahn, and S. Moon, “I Tube, YouTube, Everybody Tubes: Analyzing the Worlds Largest User GeneratedContent Video System,” in Proceedings of the 5th ACM/USENIX InternetMeasurement Conference (IMC07), 2007.

[5] M. Armbrust, A. Fox, R. Griffith, A. D. Joseph, R. H. Katz,A. Konwinski, G. Lee, D. A. Patterson, A. Rabkin, I. Stoica,and M. Zaharia, “Above the clouds: A Berkeley view of cloudcomputing,” EECS Department, University of California, Berkeley,Tech. Rep. UCB/EECS-2009-28, Feb 2009. [Online]. Available:http://www.eecs.berkeley.edu/Pubs/TechRpts/2009/EECS-2009-28.html

[6] W. Zhu, C. Luo, J. Wang, and S. Li, “Multimedia cloud computing:Directions and applications,” Signal Processing Magazine, vol. 28, no. 3,pp. 59–69, 2011.

[7] Y. Wu, C. Wu, B. Li, X. Qiu, and F. Lau, “CloudMedia: When cloud ondemand meets video on demand,” in IEEE ICDCS, 2011, pp. 268–277.

[8] H. Li, L. Zhong, J. Liu, B. Li, and K. Xu, “Cost-effective partialmigration of VoD services to content clouds,” in IEEE InternationalConference on Cloud Computing (IEEE CLOUD), 2011, pp. 203–210.

[9] D. Niu, H. Xu, B. Li, and S. Zhao, “Quality-assured cloud bandwidthauto-scaling for video-on-demand applications,” in IEEE INFOCOM,vol. 12, 2012.

[10] F. Wang, J. Liu, and M. Chen, “CALMS: Cloud-assisted live mediastreaming for globalized demands with time/region diversities,” in IEEEINFOCOM, 2012.

[11] Z. Huang, C. Mei, L. Li, and T. Woo, “CloudStream: delivering high-quality streaming videos through a cloud-based SVC proxy,” in IEEEINFOCOM mini conference, 2011.

[12] F. Chen, K. Guo, J. Lin, and T. Porta, “Intra-cloud lightning: BuildingCDNs in the cloud,” in IEEE INFOCOM, 2012.

[13] Y. Jin, Y. Wen, G. Shi, G. Wang, and A. Vasilakos, “CoDaaS: Anexperimental cloud-centric content delivery platform for user-generatedcontents,” in Proc. of IEEE International Conference on Computing,Networking and Communications, 2012.

[14] “Akamai,” http://www.akamai.com/.[15] “Limelight,” http://www.limelightnetworks.com/.[16] X. Zhang, J. Liu, B. Li, and Y. Yum, “CoolStreaming/DONet: a data-

driven overlay network for peer-to-peer live media streaming,” in IEEEINFOCOM, vol. 3. IEEE, 2005, pp. 2102–2111.

[17] X. Hei, C. Liang, J. Liang, Y. Liu, and K. Ross, “A measurement studyof a large-scale P2P IPTV system,” IEEE Transactions on Multimedia,vol. 9, no. 8, pp. 1672–1687, 2007.

[18] H. Shen, L. Zhao, J. Stokes, and J. Li, “Toward P2P-Based MultimediaSharing in User Generated Contents,” IEEE Transactions on Paralleland Distributed Systems, vol. 23, pp. 966–975, 2012.

[19] F. Liu, B. Li, L. Zhong, B. Li, H. Jin, and X. Liao, “Flash Crowd inP2P Live Streaming Systems: Fundamental Characteristics and DesignImplications,” IEEE Transactions on Parallel and Distributed Systems,vol. 23, pp. 1227–1239, 2012.

[20] “Xunlei Kankan,” http://www.xunlei.com/.[21] H. Yin, X. Liu, T. Zhan, V. Sekar, F. Qiu, C. Lin, H. Zhang, and B. Li,

“LiveSky: Enhancing CDN with P2P,” ACM Transactions on MultimediaComputing, Communications, and Applications (TOMCCAP), vol. 6,no. 3, p. 16, 2010.

[22] V. Adhikari, Y. Guo, F. Hao, M. Varvello, V. Hilt, M. Steiner, andZ. Zhang, “Unreeling Netflix: Understanding and improving multi-CDNmovie delivery,” in IEEE INFOCOM, 2012.

[23] V. Aggarwal, X. Chen, V. Gopalakrishnan, R. Jana, K. Ramakrishnan,and V. Vaishampayan, “Exploiting virtualization for delivering cloud-based IPTV services,” in IEEE Conference on Computer Communica-tions Workshops (INFOCOM WKSHPS). IEEE, 2011, pp. 637–641.

[24] D. Niu, C. Feng, and B. Li, “A theory of cloud bandwidth pricing forvideo-on-demand providers,” in IEEE INFOCOM, 2012.

[25] J. Pujol, V. Erramilli, G. Siganos, X. Yang, N. Laoutaris, P. Chhabra,and P. Rodriguez, “The little engine (s) that could: Scaling online

social networks,” in ACM SIGCOMM Computer Communication Review,vol. 40, no. 4, 2010, pp. 375–386.

[26] Y. Wu, C. Wu, B. Li, L. Zhang, Z. Li, and F. Lau, “Scaling social mediaapplications into geo-distributed clouds,” in IEEE INFOCOM, 2012.

[27] X. Cheng and J. Liu, “Load-balanced migration of social media tocontent clouds,” in ACM NOSSDAV, 2011, pp. 51–56.

[28] W. Jiang, R. Zhang-Shen, J. Rexford, and M. Chiang, “Cooperativecontent distribution and traffic engineering in an ISP network,” in ACMSIGMETRICS, 2009.

[29] D. Goldenberg, L. Qiu, H. Xie, Y. Yang, and Y. Zhang, “Optimizingcost and performance for multihoming,” in ACM SIGCOMM ComputerCommunication Review, vol. 34, no. 4, 2004, pp. 79–92.

[30] J. Dean and S. Ghemawat, “MapReduce: Simplified data processing onlarge clusters,” Communications of the ACM, vol. 51, no. 1, pp. 107–113,2008.

[31] D. Bertsekas, Convex optimization theory. Athena Scientific, 2009.[32] K. Binmore, A. Rubinstein, and A. Wolinsky, “The nash bargaining

solution in economic modelling,” The RAND Journal of Economics, pp.176–188, 1986.

[33] D. Palomar and M. Chiang, “A tutorial on decomposition methodsfor network utility maximization,” IEEE Journal on Selected Areas inCommunications, vol. 24, no. 8, pp. 1439–1451, 2006.

[34] D. Niu, Z. Liu, B. Li, and S. Zhao, “Demand forecast and performanceprediction in peer-assisted on-demand streaming systems,” in IEEEINFOCOM, 2011, pp. 421–425.

[35] G. Gursun, M. Crovella, and I. Matta, “Describing and forecasting videoaccess patterns,” in IEEE INFOCOM. IEEE, 2011, pp. 16–20.

[36] P. J. Brockwell and R. A. Davis, Introduction to Time Series andForecasting. Springer, 2002.

[37] J. He, D. Wu, Y. Zeng, X. Hei, and Y. Wen, “Towards optimaldeployment of cloud-assisted video distribution services,” CS-2012-0616, Sun Yat-sen University, Tech. Rep., 2012. [Online]. Available:http://sist.sysu.edu.cn/∼dwu/cloud-tr.pdf

[38] “Maxmind GeoIP,” http://www.maxmind.com/app/country.[39] G. Zhang, X. Zhao, X. Hei, and W. Cheng, “A comparison measurement

study of hybrid CDN-P2P VoD systems,” EIE Department, HuazhongUniversity of Science and Technology, Tech. Rep., 2012.

[40] “China Telecom homepage,” http://www.chinatelecom.com.cn/.[41] “Amazon EC2 pricing,” http://aws.amazon.com/ec2/pricing/.[42] “Google map homepage,” http://maps.google.com/.