Abstract—As an effective and efficient way to provide computing resources and services to customers on demand, cloud computinghas become more and more popular. From cloud service providers’ perspective, profit is one of the most important considerations, andit is mainly determined by the configuration of a cloud service platform under given market demand. However, a single long-termrenting scheme is usually adopted to configure a cloud platform, which cannot guarantee the service quality but leads to seriousresource waste. In this paper, a double resource renting scheme is designed firstly in which short-term renting and long-term rentingare combined aiming at the existing issues. This double renting scheme can effectively guarantee the quality of service of all requestsand reduce the resource waste greatly. Secondly, a service system is considered as an M/M/m+D queuing model and the performanceindicators that affect the profit of our double renting scheme are analyzed, e.g., the average charge, the ratio of requests that needtemporary servers, and so forth. Thirdly, a profit maximization problem is formulated for the double renting scheme and the optimizedconfiguration of a cloud platform is obtained by solving the profit maximization problem. Finally, a series of calculations are conductedto compare the profit of our proposed scheme with that of the single renting scheme. The results show that our scheme can not onlyguarantee the service quality of all requests, but also obtain more profit than the latter.Index Terms—Cloud computing, guaranteed service quality, multiserver system, profit maximization, queuing model, service-levelagreement, waiting time.F1 INTRODUCTIONAS an effective and efficient way to consolidate computingresources and computing services, clouding computinghas become more and more popular [1]. Cloud computingcentralizes management of resources and services,and delivers hosted services over the Internet. The hardware,software, databases, information, and all resources areconcentrated and provided to consumers on-demand [2].Cloud computing turns information technology into ordinarycommodities and utilities by the the pay-per-use pricingmodel [3, 4, 5]. In a cloud computing environment, thereare always three tiers, i.e., infrastructure providers, servicesproviders, and customers (see Fig. 1 and its elaboration inSection 3.1). An infrastructure provider maintains the basichardware and software facilities. A service provider rentsresources from the infrastructure providers and providesservices to customers. A customer submits its request to aservice provider and pays for it based on the amount andthe quality of the provided service [6]. In this paper, weaim at researching the multiserver configuration of a serviceprovider such that its profit is maximized.Like all business, the profit of a service provider in cloudcomputing is related to two parts, which are the cost andthe revenue. For a service provider, the cost is the renting_ The authors are with the College of Information Science and Engineering,Hunan University, and National Supercomputing Center in Changsha,Hunan, China, 410082.E-mail: jingmei1988@163.com, lkl@hnu.edu.cn, oyaj@hnu.edu.cn,lik@newpaltz.edu._ Keqin Li is also with the Department of Computer Science, State Universityof New York, New Paltz, New York 12561, USA._ Kenli Li is the author for correspondence.Manuscript received ****, 2015; revised ****, 2015.cost paid to the infrastructure providers plus the electricitycost caused by energy consumption, and the revenue is theservice charge to customers. In general, a service providerrents a certain number of servers from the infrastructureproviders and builds different multiserver systems for differentapplication domains. Each multiserver system is toexecute a special type of service requests and applications.Hence, the renting cost is proportional to the number ofservers in a multiserver system [2]. The power consumptionof a multiserver system is linearly proportional to the numberof servers and the server utilization, and to the square ofexecution speed [7, 8]. The revenue of a service provider isrelated to the amount of service and the quality of service.To summarize, the profit of a service provider is mainlydetermined by the configuration of its service platform.To configure a cloud service platform, a service providerusually adopts a single renting scheme. That’s to say, theservers in the service system are all long-term rented. Becauseof the limited number of servers, some of the incomingservice requests cannot be processed immediately. Sothey are first inserted into a queue until they can handledby any available server. However, the waiting time of theservice requests cannot be too long. In order to satisfyquality-of-service requirements, the waiting time of eachincoming service request should be limited within a certainrange, which is determined by a service-level agreement(SLA). If the quality of service is guaranteed, the serviceis fully charged, otherwise, the service provider serves therequest for free as a penalty of low quality. To obtain higherrevenue, a service provider should rent more servers fromthe infrastructure providers or scale up the server executionspeed to ensure that more service requests are processedwith high service quality. However, doing this would lead to0018-9340 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. Seehttp://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI10.1109/TC.2015.2401021, IEEE Transactions on ComputersTRANSACTIONS ON COMPUTERS, VOL. *, NO. *, * 2015 2sharp increase of the renting cost or the electricity cost. Suchincreased cost may counterweight the gain from penaltyreduction. In conclusion, the single renting scheme is not agood scheme for service providers. In this paper, we proposea novel renting scheme for service providers, which notonly can satisfy quality-of-service requirements, but also canobtain more profit. Our contributions in this paper can besummarized as follows._ A novel double renting scheme is proposed forservice providers. It combines long-term rentingwith short-term renting, which can not only satisfyquality-of-service requirements under the varyingsystem workload, but also reduce the resource wastegreatly._ A multiserver system adopted in our paper is modeledas an M/M/m+D queuing model and the performanceindicators are analyzed such as the averageservice charge, the ratio of requests that need shorttermservers, and so forth._ The optimal configuration problem of serviceproviders for profit maximization is formulated andtwo kinds of optimal solutions, i.e., the ideal solutionsand the actual solutions, are obtained respectively._ A series of comparisons are given to verify the performanceof our scheme. The results show that theproposed Double-Quality-Guaranteed (DQG) rentingscheme can achieve more profit than the comparedSingle-Quality-Unguaranteed (SQU) rentingscheme in the premise of guaranteeing the servicequality completely.The rest of the paper is organized as follows. Section 2reviews the related work on profit aware problem in cloudcomputing. Section 3 presents the used models, includingthe three-tier cloud computing model, the multiserver systemmodel, the revenue and cost models. Section 4 proposesour DQG renting scheme and formulates the profitoptimization problem. Section 5 introduces the methods offinding the optimal solutions for the profit optimizationproblem in two scenarios. Section 6 demonstrates the performanceof the proposed scheme through comparison with thetraditional SQU renting scheme. Finally, Section 7 concludesthe work.2 RELATED WORKIn this section, we review recent works relevant to the profitof cloud service providers. Profit of service providers isrelated with many factors such as the price, the marketdemand, the system configuration, the customer satisfactionand so forth. Service providers naturally wish to set a higherprice to get a higher profit margin; but doing so woulddecrease the customer satisfaction, which leads to a risk ofdiscouraging demand in the future. Hence, selecting a reasonablepricing strategy is important for service providers.The pricing strategies are divided into two categories,i.e., static pricing and dynamic pricing. Static pricing meansthat the price of a service request is fixed and knownin advance, and it does not change with the conditions.With dynamic pricing a service provider delays the pricingdecision until after the customer demand is revealed, so thatthe service provider can adjust prices accordingly [9]. Staticpricing is the dominant strategy which is widely used inreal world and in research [2, 10, 11]. Ghamkhari et al. [11]adopted a flat-rate pricing strategy and set a fixed price forall requests, but Odlyzko in [12] argued that the predominantflat-rate pricing encourages waste and is incompatiblewith service differentiation. Another kind of static pricingstrategies are usage-based pricing. For example, the priceof a service request is proportional to the service time andtask execution requirement (measured by the number ofinstructions to be executed) in [10] and [2], respectively.Usage-based pricing reveals that one can use resources moreefficiently [13, 14].Dynamic pricing emerges as an attractive alternativeto better cope with unpredictable customer demand [15].Mac´ıas et al. [16] used a genetic algorithm to iterativelyoptimize the pricing policy. Amazon EC2 [17, 18] has introduceda ”spot pricing” feature, where the spot price fora virtual instance is dynamically updated to match supplyand demand. However, consumers dislike prices to change,especially if they perceive the changes to be ”unfair” [19, 20].After comparison, we select the usage-based pricing strategyin this paper since it agrees with the concept of cloudcomputing mostly.The second factor affecting the profit of service providersis customer satisfaction which is determined by the qualityof service and the charge. In order to improve the customersatisfaction level, there is a service-level agreement (SLA)between a service provider and the customers. The SLAadopts a price compensation mechanism for the customerswith low service quality. The mechanism is to guaranteethe service quality and the customer satisfaction so thatmore customers are attracted. In previous research, differentSLAs are adopted. Ghamkhari et al. [11] adopted a stepwisecharge function with two stages. If a service request ishandled before its deadline, it is normally charged; butif a service request is not handled before its deadline, itis dropped and the provider pays for it due to penalty.In [2, 10, 21], charge is decreased continuously with theincreasing waiting time until the charge is free. In thispaper, we use a two-step charge function, where the servicerequests served with high quality are normally charged,otherwise, are served for free.Since profit is an important concern to cloud serviceproviders, many works have been done on how to boosttheir profit. A large body of works have recently focusedon reducing the energy cost to increase profit of serviceproviders [22, 23, 24, 25], and the idle server turning offstrategy and dynamic CPU clock frequency scaling are adoptedto reduce energy cost. However, only reducing energycost cannot obtain profit maximization. Many researchersinvestigated the trade-off between minimizing cost andmaximizing revenue to optimize profit. Both [11] and [26]adjusted the number of switched on servers periodicallyusing different strategies and different profit maximizationmodels were built to get the number of switched on servers.However, these works did not consider the cost of resourceconfiguration.Chiang and Ouyang [27] considered a cloud serversystem as an M/M/R/K queuing system where all service0018-9340 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. Seehttp://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI10.1109/TC.2015.2401021, IEEE Transactions on ComputersTRANSACTIONS ON COMPUTERS, VOL. *, NO. *, * 2015 3requests that exceed its maximum capacity are rejected. Aprofit maximization function is defined to find an optimalcombination of the server size R and the queue capacity Ksuch that the profit is maximized. However, this strategyhas further implications other than just losing the revenuefrom some services, because it also implies loss of reputationand therefore loss of future customers [3]. In [2], Cao etal. treated a cloud service platform as an M/M/m model,and the problem of optimal multiserver configuration forprofit maximization was formulated and solved. This workis the most relevant work to ours, but it adopts a singlerenting scheme to configure a multiserver system, whichcannot adapt to the varying market demand and leads tolow service quality and great resource waste. To overcomethis weakness, another resource management strategy isused in [28, 29, 30, 31], which is cloud federation. Usingfederation, different providers running services thathave complementary resource requirements over time canmutually collaborate to share their respective resources inorder to fulfill each one’s demand [30]. However, providersshould make an intelligent decision about utilization ofthe federation (either as a contributor or as a consumerof resources) depending on different conditions that theymight face, which is a complicated problem.In this paper, to overcome the shortcomings mentionedabove, a double renting scheme is designed to configurea cloud service platform, which can guarantee the servicequality of all requests and reduce the resource waste greatly.Moreover, a profit maximization problem is formulated andsolved to get the optimal multiserver configuration whichcan product more profit than the optimal configurationin [2].3 THE MODELSIn this section, we first describe the three-tier cloud computingstructure. Then, we introduce the related models used inthis paper, including a multiserver system model, a revenuemodel, and a cost model.3.1 A Cloud System ModelThe cloud structure (see Fig. 1) consists of three typicalparties, i.e., infrastructure providers, service providers andcustomers. This three-tier structure is used commonly inexisting literatures [2, 6, 10].Fig. 1: The three-tier cloud structure.In the three-tier structure, an infrastructure provider thebasic hardware and software facilities. A service providerrents resources from infrastructure providers and preparesa set of services in the form of virtual machine (VM). Infrastructureproviders provide two kinds of resource rentingschemes, e.g., long-term renting and short-term renting. Ingeneral, the rental price of long-term renting is much cheaperthan that of short-term renting. A customer submits aservice request to a service provider which delivers serviceson demand. The customer receives the desired result fromthe service provider with certain service-level agreement,and pays for the service based on the amount of the serviceand the service quality. Service providers pay infrastructureproviders for renting their physical resources, and chargecustomers for processing their service requests, which generatescost and revenue, respectively. The profit is generatedfrom the gap between the revenue and the cost.3.2 A Multiserver ModelIn this paper, we consider the cloud service platform as amultiserver system with a service request queue. Fig. 2 givesthe schematic diagram of cloud computing [32].Fig. 2: The schematic diagram of cloud computing.In an actual cloud computing platform such as AmazonEC2, IBM blue cloud, and private clouds, there are manywork nodes managed by the cloud managers such as Eucalyptus,OpenNebula, and Nimbus. The clouds provideresources for jobs in the form of virtual machine (VM). Inaddition, the users submit their jobs to the cloud in whicha job queuing system such as SGE, PBS, or Condor is used.All jobs are scheduled by the job scheduler and assignedto different VMs in a centralized way. Hence, we can considerit as a service request queue. For example, Condor isa specialized workload management system for computeintensivejobs and it provides a job queueing mechanism,scheduling policy, priority scheme, resource monitoring,and resource management. Users submit their jobs to Condor,and Condor places them into a queue, chooses whenand where to run them based upon a policy [33, 34]. Hence,it is reasonable to abstract a cloud service platform as a multiservermodel with a service request queue, and the modelis widely adopted in existing literature [2, 11, 35, 36, 37].In the three-tier structure, a cloud service provider servescustomers’ service requests by using a multiserver system0018-9340 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. Seehttp://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI10.1109/TC.2015.2401021, IEEE Transactions on ComputersTRANSACTIONS ON COMPUTERS, VOL. *, NO. *, * 2015 4Fig. 3: The multiserver system model, where servicerequests are first placed in a queue before they areprocessed by any servers.which is rented from an infrastructure provider. Assumethat the multiserver system consists of m long-term rentedidentical servers, and it can be scaled up by temporarilyrenting short-term servers from infrastructure providers.The servers in the system have identical execution speed s(Unit: billion instructions per second). In this paper, a multiserversystem excluding the short-term servers is modeledas an M/M/m queuing system as follows (see Fig. 3). Thereis a Poisson stream of service requests with arrival rate λ,i.e., the interarrival times are independent and identicallydistributed (i.i.d.) exponential random variables with mean1/λ. A multiserver system maintains a queue with infinitecapacity. When the incoming service requests cannot be processedimmediately after they arrive, they are firstly placedin the queue until they can be handled by any availableserver. The first-come-first-served (FCFS) queuing disciplineis adopted. The task execution requirements (measured bythe number of instructions) are independent and identicallydistributed exponential random variables r with mean r(Unit: billion instructions). Therefore, the execution times oftasks on the multiserver system are also i.i.d. exponentialrandom variables x = r/s with mean x = r/s (Unit:second). The average service rate of each server is calculatedas μ = 1/x = s/r, and the system utilization is defined asρ = λ/mμ = λ/m _ r/s.Because the fixed computing capacity of the servicesystem is limited, some requests would wait for a long timebefore they are served. According to the queuing theory, wehave the following theorem about the waiting time in anM/M/m queuing system.Theorem 3.1. The cumulative distribution function (cdf) ofthe waiting time W of a service request isFW(t) = 1 πm1 ρemμ(1ρ)t, (1)whereπm =(mρ)mm![mΣ1k=0(mρ)kk!+(mρ)mm!(1 ρ)]1.Proof 3.1. We have known that the probability distributionfunction (pdf) of the waiting time W of a service requestisfW(t) = (1 Pq)u(t) + mμπme(1ρ)mμt,where Pq = πm/(1 ρ) and u(t) is a unit impulsefunction [2, 38]. Then, FW(t) can be obtained by straightforwardcalculation.3.3 Revenue ModelingThe revenue model is determined by the pricing strategyand the server-level agreement (SLA). In this paper, theusage-based pricing strategy is adopted, since cloud computingprovides services to customers and charges them ondemand. The SLA is a negotiation between service providersand customers on the service quality and the price. Becauseof the limited servers, the service requests that cannot behandled immediately after entering the system must wait inthe queue until any server is available. However, to satisfythe quality-of-service requirements, the waiting time of eachservice request should be limited within a certain rangewhich is determined by the SLA. The SLA is widely used bymany types of businesses, and it adopts a price compensationmechanism to guarantee service quality and customersatisfaction. For example, China Post gives a service timecommitment for domestic express mails. It promises thatif a domestic express mail does not arrive within a deadline,the mailing charge will be refunded. The SLA is alsoadopted by many real world cloud service providers suchas Rackspace [39], Joyent [40], Microsoft Azure [41], and soon. Taking Joyent as an example, the customers order SmartMachines, Smart Appliances, and/or Virtual Machines fromJoyent, and if the availability of a customer’s services isless than 100%, Joyent will credit the customer 5% of themonthly fee for each 30 minutes of downtime up to 100% ofthe customer’s monthly fee for the affected server. The onlydifference is that its performance metric is availability andours is waiting time.In this paper, the service level is reflected by the waitingtime of requests. Hence, we define D as the maximumwaiting time here that the service requests can tolerate, inother words, D is their deadline. The service charge of eachtask is related to the amount of a service and the servicelevelagreement. We define the service charge function fora service request with execution requirement r and waitingtime W in Eq. (2),R(r,W) ={ar, 0 _ W _ D;0, W > D.(2)where a is a constant, which indicates the price per onebillion instructions (Unit: cents per one billion instructions).When a service request starts its execution before waitinga fixed time D (Unit: second), a service provider considersthat the service request is processed with high quality-ofserviceand charges a customer ar. If the waiting time of aservice request exceeds deadline D, a service provider mustserve it for free. Similar revenue models have been used inmany existing research such as [2, 11, 42].According to Theorem 1, it is easy to know that theprobability that the waiting time of a service request exceedsits deadline D isP(W _ D) = 1 FW(D) =πm1 ρemμ(1ρ)D. (3)3.4 Cost ModelingThe cost of a service provider consists of two major parts,i.e., the rental cost of physical resources and the utilitycost of energy consumption. Many existing research suchas [11, 43, 44] only consider the power consumption cost.As a major difference between their models and ours, the0018-9340 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. Seehttp://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI10.1109/TC.2015.2401021, IEEE Transactions on ComputersTRANSACTIONS ON COMPUTERS, VOL. *, NO. *, * 2015 5resource rental cost is considered in this paper as well, sinceit is a major part which affects the profit of service providers.A similar cost model is adopted in [2]. The resources canbe rented in two ways, long-term renting and short-termrenting, and the rental price of long-term renting is muchcheaper than that of short-term renting. This is reasonableand common in the real life. In this paper, we assume thatthe long-term rental price of one server for unit of time isβ (Unit: cents per second) and the short-term rental priceof one server for unit of time is γ (Unit: cents per second),where β < γ.The cost of energy consumption is determined by theelectricity price and the amount of energy consumption. Inthis paper, we adopt the following dynamic power model,which is adopted in the literature such as [2, 7, 45, 46]:Pd = NswCLV 2f, (4)where Nsw is the average gate switching factor at each clockcycle, CL is the loading capacitance, V is the supply voltage,and f is the clock frequency [45]. In the ideal case, therelationship between the clock frequency f and the supplyvoltage V is V / fϕ for some constant ϕ > 0 [46]. Theserver execution speed s is linearly proportional to the clockfrequency f, namely, s / f. Hence, the power consumptionis Pd / NswCLs2ϕ+1. For ease of discussion, we assumethat Pd = bNswCLs2ϕ+1 = ξsα where ξ = bNswCL andα = 2ϕ + 1. In this paper, we set NswCL = 7.0, b = 1.3456and ϕ = 0.5. Hence, α = 2.0 and ξ = 9.4192. The value ofpower consumption calculated by Pd = ξsα is close to thevalue of the Intel Pentium M processor [47]. It is reasonablethat a server still consumes some amount of static power [8],denoted as P_ (Unit:Watt), when it is idle. For a busy server,the average amount of energy consumption per unit of timeis P = ξsα + P_ (Unit: Watt). Assume that the price ofenergy is δ (Unit: cents per Watt).4 A QUALITY-GUARANTEED SCHEMEThe traditional single resource renting scheme cannot guaranteethe quality of all requests but wastes a great amountof resources due to the uncertainty of system workload.To overcome the weakness, we propose a double rentingscheme as follows, which not only can guarantee the qualityof service completely but also can reduce the resource wastegreatly.4.1 The Proposed SchemeIn this section, we first propose the Double-Quality-Guaranteed (DQG) resource renting scheme which combineslong-term renting with short-term renting. The maincomputing capacity is provided by the long-term rentedservers due to their low price. The short-term rented serversprovide the extra capacity in peak period. The detail of thescheme is shown in Algorithm 1.The proposed DQG scheme adopts the traditional FCFSqueueing discipline. For each service request entering thesystem, the system records its waiting time. The requests areassigned and executed on the long-term rented servers inthe order of arrival times. Once the waiting time of a requestreaches D, a temporary server is rented from infrastructureAlgorithm 1 Double-Quality-Guaranteed (DQG) Scheme1: A multiserver system with m servers is running and waitingfor the events as follows2: A queue Q is initialized as empty3: Event – A service request arrives4: Search if any server is available5: if true then6: Assign the service request to one available server7: else8: Put it at the end of queue Q and record its waiting time9: end if10: End Event11: Event – A server becomes idle12: Search if the queue Q is empty13: if true then14: Wait for a new service request15: else16: Take the first service request from queue Q and assign itto the idle server17: end if18: End Event19: Event – The deadline of a request is achieved20: Rent a temporary server to execute the request and releasethe temporary server when the request is completed21: End Eventproviders to process the request. We consider the novel servicemodel as an M/M/m+D queuing model [48, 49, 50]. TheM/M/m+D model is a special M/M/m queuing model withimpatient customers. In an M/M/m+D model, the requestsare impatient and they have a maximal tolerable waitingtime. If the waiting time exceeds the tolerable waiting time,they lose patience and leave the system. In our scheme, theimpatient requests do not leave the system but are assignedto temporary rented servers.Since the requests with waiting time D are all assignedto temporary servers, it is apparent that all service requestscan guarantee their deadline and are charged based on theworkload according to the SLA. Hence, the revenue of theservice provider increases. However, the cost increases aswell due to the temporarily rented servers. Moreover, theamount of cost spent in renting temporary servers is determinedby the computing capacity of the long-term rentedmultiserver system. Since the revenue has been maximizedusing our scheme, minimizing the cost is the key issue forprofit maximization. Next, the tradeoff between the longtermrental cost and the short-term rental cost is considered,and an optimal problem is formulated in the following toget the optimal long-term configuration such that the profitis maximized.4.2 The Profit Optimization ProblemAssume that a cloud service platform consists of m longtermrented servers. It is known that part of requests needtemporary servers to serve, so that their quality can beguaranteed. Denoted by pext(D) the steady-state probabilitythat a request is assigned to a temporary server, or putdifferently, pext(D) is the long-run fraction of requests whosewaiting times exceed the deadline D. pext(D) is differentfrom FW(D). In calculating FW(D), all service requests,whether exceed the deadline, will be waiting in the queue.However, in calculating pext(D), the requests whose waiting0018-9340 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. Seehttp://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI10.1109/TC.2015.2401021, IEEE Transactions on ComputersTRANSACTIONS ON COMPUTERS, VOL. *, NO. *, * 2015 6times are equal to the deadline will be assigned to thetemporary servers, which will reduce the waiting time ofthe following requests. In general, pext(D) is much less thanFW(D). Refer to [50], we can known that pext(D) is:pext(D) =(1 ρ)(1 FW(D))1 ρ(1 FW(D)). (5)0 5 10 15 20 25 30 35 40 45 5000.030.060.090.120.150.180.210.24Deadlinepextaver_r=1Fig. 4: The probability of waiting time exceeding D.That is to say, there are about λpext(D) service requestsin one unit of time which need short-term rented servers.Fig. 4 gives the probability versus different deadline whereλ = 5.99, r = 1, m = 6 and s = 1. Hence, the cost onshort-term rented servers in one unit of time is calculatedas:Cshort = λpext(D)rs(γ + δP), (6)where rs is the average execution time of each request.Among the requests entering the service system, aboutpext(D) percentage requests are not executed by the m longtermrented servers. Hence, the system utilization of them servers is ρ(1 pext(D)). Since the power for speeds is ξsα, the average amount of energy consumed by along-term rented server in one unit of time is Plong =ρ(1 pext(D))ξsα + P_. Hence, the cost of the long-termrented servers in one unit of time is calculated as:Clong = m(β + δPlong). (7)The following theorem gives the expected charge to aservice request.Theorem 4.1. The expected charge to a service request is ar.Proof 4.1. Because the waiting time W of each request isless than or equal to D, the expected charge to a servicerequest with execution requirement r is ar according tothe SLA. Since r is a random variable, ar is also randomvariable. It is known that r is an exponential randomvariable with mean r, so its probability distributionfunction is fr(z) = 1r ez/r. The expected charge to aservice re∫quest is 10fr(z)R(r, z)dz =∫ 101rez/razdz=ar∫ 10ez/rzdz = a∫ 10zdez/r= a[zez/r___10∫ 10ez/rdz]= a[zez/r___10+ rez/r___10]= ar.(8)The theorem is proven.The profit of a service provider in one unit of time isobtained asProfit = Revenue Clong Cshort, (9)where Revenue = λar,Clong = m(β + δ(ρ(1 pext(D))ξsα + P_)),andCshort = λpext(D)rs(γ + δ(ξsα + P_)).We aim to choose the optimal number of fixed servers mand the optimal execution speed s to maximize the profit:Profit(m, s) = λar λpext(D)rs(γ + δ(ξsα + P_)) m(β + δ(ρ(1 pext(D))ξsα + P_)).(10)Fig. 5 gives the graph of function Profit(m, s) where λ =5.99, r = 1, D = 5, a = 15, P_ = 3, α = 2.0, ξ = 9.4192,β = 1.5, γ = 3, and δ = 0.3.01020301 0 3 2−150−100−50050The Server Speed The Server SizeProfitFig. 5: The function Profit(m, s).From the figure, we can see that the profit of a serviceprovider is varying with different server size and differentexecution speed. Therefore, we have the problem of selectingthe optimal server size and/or server speed so that theprofit is maximized. In the following section, the solutionsto this problem are proposed.5 OPTIMAL SOLUTIONIn this section, we first develop an analytical methodto solve our optimization problem. Using the analyticalmethod, the ideal optimal solutions are obtained. Becausethe server size and the server speed are limited and discrete,we give an algorithmic method to get the actual solutionsbased on the ideal optimal ones.5.1 An Analytical Method for Ideal SolutionsWe firstly solve our optimization problem analytically, assumingthat m and s are continuous variables. To thisend, a closed-form expression of pext(D) is needed. In thispaper, we use the same closed-form expression as [2], whichisΣm1k=0(mρ)kk!_ emρ. This expression is very accuratewhen m is not too small and ρ is not too large [2]. SinceStirling’s approximation of m! isp2πm(me )m, one closedformexpression of πm isπm _ p 1ρ2πm(1ρ)( e_eρ )m+1,0018-9340 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. Seehttp://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI10.1109/TC.2015.2401021, IEEE Transactions on ComputersTRANSACTIONS ON COMPUTERS, VOL. *, NO. *, * 2015 7andpext(D) _ (1 ρ)emμ(1ρ)D1+p2πm(1ρ)( e_eρ )mρemμ(1ρ)D.For convenience, we rewrite pext(D) _ (1ρ)K1K2ρK1, whereK1 = emμ(1ρ)D, and K2 = 1 +p2πm(1 ρ)_, where_ = (eρ/eρ)m.In the following, we solve our optimization problemsbased on above closed-form expression of pext(D).5.1.1 Optimal SizeGiven λ, r, a, P_, α, β, γ, δ, ξ, D, and s, our objective is tofind m such that Profit is maximized. To maximize Profit, mmust be found such that∂Profit∂m= ∂Clong∂m ∂Cshort∂m= 0,where∂Clong∂m= β + δP_ δλrξsα1 ∂pext(D)∂m,and∂Cshort∂m= λ(γ + δP_)rs∂pext(D)∂m+ λrδξsα1 ∂pext(D)∂m.Sinceln_ = mln(eρ/eρ) = m(ρ ln ρ 1),and∂ρ∂m= λrm2s= ρm,we have1_∂_∂m= (ρ ln ρ 1) + m(1 1ρ)∂ρ∂m= ln ρ,and∂_∂m= _ln ρ.Then, we get∂K1∂m= μDK1,and∂K2∂m=p2πm_(12m(1 + ρ) ln ρ(1 ρ)).Furthermore, we have∂pext(D)∂m=1(K2ρK1)2[ ρmK1(K2K1)+ (ρ1)μDK1K2(1+ρ)K12m(K21)+ (1ρ)K1(ln ρ)(K21)].We cannot get a closed-form solution to m, but we canget the numerical solution to m. Since ∂Profit/∂m is not anincreasing or decreasing function of m, we need to find thedecreasing region of m, and then use the standard bisectionmethod. If there are more than one maximal values, theyare compared and the maximum is selected. When usingthe bisection method to find the extreme point, the iterationaccuracy is set as a unified value 1010.In Fig. 6, we demonstrate the net profit in one unit oftime as a function of m and λ where s = 1, r = 1, and theother parameters are same as with those in Fig. 5. We noticethat there is an optimal choice of msuch that the net profit ismaximized. Using the analytical method, the optimal valueof m such that ∂Profit/∂m = 0 is 4.8582, 5.8587, 6.8590,1 2 3 4 5 6 7 8 9 10111213141516171819200102030405060708090The Server SizeProfitlamda=4.99lamda=5.99lamda=6.99lamda=7.99Fig. 6: Net profit versus m and λ.0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 201020304050The Server SpeedOptimal Sizelamda=4.99lamda=5.99lamda=6.99lamda=7.99(a) Optimal size versus s and _.0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 20102030405060708090The Server SpeedMaximal Profitlamda=4.99lamda=5.99lamda=6.99lamda=7.99(b) Maximal profit versus s and _.Fig. 7: Optimal size and maximal profit vs. s and λ.7.8592 for λ = 4.99, 5.99, 6.99, 7.99, respectively. When thenumber of servers m is less than the optimal value, theservice provider needs to rent more temporary servers toexecute the requests whose waiting times are equal to thedeadline; hence, the extra cost increases, even surpassingthe gained revenue. As m increases, the waiting times aresignificantly reduced, but the cost on fixed servers increasesgreatly, which also surpasses the gained revenue too. Hence,there is an optimal choice of m which maximizes the profit.In Fig. 7, we demonstrate the optimal size and maximalprofit in one unit of time as a function of s and λ. It means,for each combination of s and λ, we find the optimal numberof servers and the maximal profit. The parameters are same0018-9340 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. Seehttp://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI10.1109/TC.2015.2401021, IEEE Transactions on ComputersTRANSACTIONS ON COMPUTERS, VOL. *, NO. *, * 2015 8as those in Fig. 6. From the figures we can see that a higherspeed leads to a less number of servers needed for each λ,and different λ values have different optimal combinationsof speed and size. In addition, the greater the λ is, the morethe maximal profit can be obtained.5.1.2 Optimal SpeedGiven λ, r, a, P_, α, β, γ, δ, ξ, D, and m, our objective isto find s such that Profit is maximized. To maximize Profit, smust be found such that∂Profit∂s= ∂Clong∂s ∂Cshort∂s= 0,where∂Clong∂s= δξλrsα2[(α 1)(1 pext(D)) s∂pext(D)∂s],and∂Cshort∂s=rλ(γ + δP_)s2(s∂pext(D)∂s pext(D))+ λrδξsα2[s∂pext(D)∂s+ (α 1)pext(D)],Since∂ρ∂s= λrms2 = ρs,and1_∂_∂s= m(1 1ρ)∂ρ∂s,we have∂_∂s=ms(1 ρ)_.Now, we get∂K1∂s= DK1mr,and∂K2∂s=p2πm(ρ + m(1 ρ)2)_s.Furthermore, we have∂pext(D)∂s=1(K2 ρK1)2[ρsK1(K2 K1)+ (ρ 1)K1p2πm(ρ + m(1 ρ)2)_s+ (ρ 1)DK1K2mr].Similarly, we cannot get the closed-form expression ofs, so we can use the same method to find the numericalsolution of s. In Fig. 8, we demonstrate the net profit inone unit of time as a function of s and λ, where m = 6.The rest parameters are the same as that in Figs. 6 and 7.We notice that there is an optimal choice of s such thatthe net profit is maximized. Using the analytical method,the optimal value of s such that respectively. When theservers run at a slower speed than the optimal speed, thewaiting times of service requests will be long and exceedthe deadline. So, the revenue is small and the profit is notoptimal. When s increases, the energy consumption as wellas the electricity cost increases. Hence, the increased revenueis much less than the increased cost. As a result, the profit isreduced. Therefore, there is an optimal choice of s such thatthe net profit is maximized.In Fig. 9, we demonstrate the optimal speed and maximalprofit in one unit of time as a function of m and λ. The0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 1.5102030405060708090The Server SpeedProfitlamda=4.99lamda=5.99lamda=6.99lamda=7.99Fig. 8: Net profit versus s and λ.1 3 5 7 9 11 13 15 17 19 21 23 2500.20.40.60.811.21.41.6The Server SizeOptimal Speedlamda=4.99lamda=5.99lamda=6.99lamda=7.99(a) Optimal speed versus m and _.1 3 5 7 9 11 13 15 17 19 21 23 250102030405060708090The Server SizeMaximal Profitlamda=4.99lamda=5.99lamda=6.99lamda=7.99(b) Maximal profit versus m and _.Fig. 9: Optimal speed and maximal profit versus m and λ.parameters are same as that in Figs. 6–8. From the figureswe can see that if the number of fixed servers is great, theservers must run at a lower speed, which can lead to anoptimal profit. In addition, the optimal speed of servers isnot faster than 1.2, that is because the increased electricitycost surpasses the increased cost that rents extra servers. Thefigure also shows us that different λ values have differentoptimal combinations of speed and size.5.1.3 Optimal Size and SpeedGiven λ, r, a, P_, α, β, γ, δ, ξ, D, our third problem is to findm and s such that Profit is maximized. Hence, we need tofind m and s such that ∂Profit/∂m = 0 and ∂Profit/∂s = 0,where ∂Profit/∂m and ∂Profit/∂s have been derived in the0018-9340 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. Seehttp://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI10.1109/TC.2015.2401021, IEEE Transactions on ComputersTRANSACTIONS ON COMPUTERS, VOL. *, NO. *, * 2015 90.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 1.525303540455055The Server SpeedProfitm=3m=4m=5m=6Fig. 10: Net profit versus m and s.0.5 0.75 1 1.25 1.5 1.75 2020406080100120140160Average rMaximal Profitlamda=4.99lamda=5.99lamda=6.99lamda=7.99Fig. 11: Maximal profit versus λ and r.last two sections. The two equations are solved by usingthe same method as [2]. In Fig. 10, we demonstrate the netprofit in one unit of time as a function of m and s. Here λis 5.99, and r = 1. The optimal value is m = 6.2418 ands = 0.9386, which result in the maximal profit 58.0150. InFig. 11, we demonstrate the maximal profit in one unit oftime in different combinations of λ and r. The figure showsthat the service providers can obtain more profit when theservice requests are with greater λ and r.5.2 An Algorithmic Method for Actual SolutionsIn above subsection, the optimal solutions find using theanalytical method are ideal solutions. Since the number ofrented servers must be integer and the server speed levelsare discrete and limited in real system, we need to find theoptimal solutions for the discrete scenarios. Assume thatS = fsij1 _ i _ ng is a discrete set of n speed levels withincreasing order. Next, different situations are discussed andthe corresponding methods are given as follows.5.2.1 Optimal SizeAssume that all servers run at a given execution speed s.Given λ, r, a, P_, α, β, γ, δ, ξ, and D, the first problem is tofind the number of long-term rented servers m such that theprofit is maximized. The method is shown in Algorithm 2.5.2.2 Optimal SpeedAssume that the service provider rents m servers. Given λ,r, a, P_, α, β, γ, δ, ξ, andD, the second problem is to find theAlgorithm 2 Finding the optimal sizeInput: s, _, r, a, P_, _, _, , _, _, and DOutput: the optimal number Opt size of fixed servers1: Profit max ← 02: find the server sizemusing the analytical method in Section5.1.13: m_l← ⌊m⌋, m_u← ⌈m⌉4: Profitl← Profit(m_l ; s), Profitu← Profit(m_u; s)5: if Profitl > Profitu then6: Profit max ← Profitl7: Opt size ← m_l8: else9: Profit max ← Profitu10: Opt size ← m_u11: end ifoptimal execution speed of all servers such that the profit ismaximized. The method is shown in Algorithm 3.Algorithm 3 Finding the optimal speedInput: m, _, r, a, P_, _, _, , _, _, and DOutput: the optimal server speed Opt speed1: Profit max ← 02: find the server speed s using the analytical method inSection 5.1.23: s_l← si, s_u← si+1 if si < s ≤ si+14: Profitl← Profit(m; s_l ), Profitu← Profit(m; s_u)5: if Profitl > Profitu then6: Profit max ← Profitl7: Opt speed ← s_l8: else9: Profit max ← Profitu10: Opt speed ← s_u11: end if5.2.3 Optimal Size and SpeedIn this subsection, we solve the third problem, which is tofind the optimal combination of m and s such that the profitis maximized. Given λ, r, a, P_, α, β, γ, δ, ξ, and D, themethod is shown in Algorithm 4.Algorithm 4 Finding the optimal size and speedInput: _, r, a, P_, _, _, , _, _, and DOutput: the optimal number Opt size of fixed servers and theoptimal execution speed Opt speed of servers1: Profit max ← 02: find the server size m and speed s using the analyticalmethod in Section 5.1.33: m_l← ⌊m⌋, m_u← ⌈m⌉4: find the optimal speed s_l and s_u using Algorithm 3 withserver size m_l and m_u, respectively5: Profitl← Profit(m_l ; s_l ), Profitu← Profit(m_u; s_u)6: if Profitl≤ Profitu then7: Profit max ← Profitu8: Opt size ← m_u , Opt speed ← s_u9: else10: Profit max ← Profitl11: Opt size ← m_l , Opt speed ← s_l12: end if5.3 Comparison of Two Kinds of SolutionsIn Tables 1, 2, and 3, the ideal optimal solutions and theactual optimal solutions are compared for three different0018-9340 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. Seehttp://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI10.1109/TC.2015.2401021, IEEE Transactions on ComputersTRANSACTIONS ON COMPUTERS, VOL. *, NO. *, * 2015 10cases. Table 1 compares the ideal optimal size and the actualoptimal size under the given server speed. Table 2 comparesthe ideal optimal speed and the actual optimal speed underthe given server size. In Table 3, two kinds of solutions arecompared for different combinations of λ and r. Here, mcan be any positive integer, and the available speed levelsare S = f0.2, 0.4, _ _ _ , 2.0g. According to the comparisonswe can see that the ideal maximal profit is greater thanthe actual maximal profit. In the tables, we also list therelative difference (RD) between the ideal optimal profit andthe actual optimal profit, which is calculated asRD =Idep ActpActp,where Idep and Actp are the maximal profit in ideal andactual scenarios. From the results we know that the relativedifference is always small except some cases in Table 2. Thatis because a small difference of speed would lead to a bigdifference of profit when the server size is large.6 PERFORMANCE COMPARISONUsing our resource renting scheme, temporary servers arerented for all requests whose waiting time are equal to thedeadline, which can guarantee that all requests are servedwith high service quality. Hence, our scheme is superiorto the traditional resource renting scheme in terms of theservice quality. Next, we conduct a series of calculationsto compare the profit of our renting scheme and the rentingscheme in [2]. In order to distinguish the proposedscheme and the compared scheme, the proposed schemeis renamed as Double-Quality-Guaranteed (DQG) rentingscheme and the compared scheme is renamed as Single-Quality-Unguaranteed (SQU) renting scheme in this paper.6.1 The Compared SchemeFirstly, the average charge of the using the SQU rentingscheme is analyzed.Theorem 6.1. The expected charge to a service request usingthe SQU renting scheme isar(1 Pqe(1ρ)mμD).Proof 6.1. Recall that the probability distribution function ofthe waiting time W of a service request isfW(t) = (1 Pq)u(t) + mμπme(1ρ)mμt.Since W is a random variable, so R(r,W) is also a randomvariable. The expected charge to a service requestwith execution requirement r isR(r) = R(r,W)=∫ 10fW(t)R(r, t)dt=∫ D0[(1 Pq)u(t) + mμπme(1ρ)mμt]ardt= (1 Pq)ar + mμπmar1 e(1ρ)mμD(1 ρ)mμ= ar(1 Pqe(1ρ)mμD).Therefore, the expected charge to a service request is theexpected value of R(r):R(r)=∫ 10fr(z)R(z)dz=∫ 101rez/raz(1 Pqe(1ρ)mμD)dz=ar(1 Pqe(1ρ)mμD)∫ 10ez/rzdz= ar(1 Pqe(1ρ)mμD).The theorem is proven.By the above theorem, the profit in one unit of time usingthe SQU renting scheme is calculated as:λar(1 Pqe(1ρ)mμD) m(β + δ(ρξsα + P_)). (11)Using the SQU renting scheme, a service provider mustrent more servers or scale up the server speed to maintaina high quality-guaranteed ratio. Assumed that the requiredquality-guaranteed ratio of a service provider is ψ and thedeadline of service requests is D. By solving equationFW(D) = 1 πm1 ρemμ(1ρ)D _ ψwith given mor s, we can get the corresponding s orm suchthat the required quality-guaranteed ratio is achieved.6.2 Profit Comparison under Different Quality-Guaranteed RatioLet λ be 5.99 and the other parameters be the same asthose in Section 5. In the first example, for a given numberof servers, we compare the profit using the SQU rentingscheme with quality-guaranteed ratio 100%, 99%, 92%, 85%and the optimal profit using our DQG renting scheme. Becausethe quality-guaranteed ratio 100% cannot be achievedusing the SQU renting scheme, hence, we set 99.999999% _100%. The results are shown in Fig. 12. From the figure, wecan see that the profit obtained using the proposed scheme isalways greater than that using the SQU renting scheme, andthe five curves reach the peak at different sizes. In addition,the profit obtained by a service provider increases whenthe qualtiy-guaranteed ratio increases from 85% to 99%, butdecreases when the ratio is greater than 99%. That is becausemore service requests are charged with the increasing ratiofrom 85% to 99%; but once the ratio is greater than 99%, thecost to expand the server size is greater than the revenueobtained from the extra qualtiy-guaranteed requests, hence,the total profit is reduced.In the second example, we compare the profit of theabove five scenarios under the given server speed. Theresults are given in Fig. 13. The figure shows the trend ofprofit when the server speed is increasing from 0.1 to 2.9.From the figure, we can see that the curves increase firstlyand reach the peak at certain speed, and then decrease alongwith the increasing speed on the whole. The figure verifiesthat our proposed scheme can obtain more profit than theSQU renting scheme. Noticed that the changing trends ofthe curves of the SQU renting scheme with 100%, 99%,92%, and 85% quality-guaranteed ratio are interesting. Theyshow an increasing trend at the beginning and then decreaseduring a small range of speed repeatedly. The reason is0018-9340 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. Seehttp://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI10.1109/TC.2015.2401021, IEEE Transactions on ComputersTRANSACTIONS ON COMPUTERS, VOL. *, NO. *, * 2015 11TABLE 1: Comparison of the two methods for finding the optimal sizeGiven Speed 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0IdealSolutionOptimal Size 29.1996 14.6300 9.7599 7.3222 5.8587 4.8827 4.1854 3.6624 3.2555 2.9300Maximal Profit 11.5546 45.5262 54.6278 57.5070 57.8645 56.9842 55.3996 53.3498 51.0143 48.4578ActualSolutionOptimal Size 29 15 10 7 6 5 4 4 3 3Maximal Profit 11.5268 45.4824 54.6014 57.3751 57.8503 56.9727 55.3259 53.0521 50.8526 48.4513Relative Difference 0.2411% 0.0964% 0.0483% 0.2299% 0.0246% 0.0202% 0.1332% 0.5612% 0.3180% 0.01325%TABLE 2: Comparison of the two methods for finding the optimal speedGiven Size 5 7 9 11 13 15 17 19 21 23IdealSolutionOptimal Speed 1.1051 0.8528 0.6840 0.5705 0.4895 0.4288 0.3817 0.3440 0.3132 0.2875Maximal Profit 57.3742 57.7613 56.0783 53.3337 49.9896 46.2754 42.3167 38.1881 33.9366 29.5933ActualSolutionOptimal Speed 1.0 0.8 0.8 0.6 0.6 0.4 0.4 0.4 0.4 0.4Maximal Profit 57.0479 57.3751 54.7031 53.1753 48.4939 45.4824 42.2165 37.4785 32.6795 27.8795Relative Difference 0.5721% 0.6732% 2.5140% 0.2979% 3.0843% 1.7435% 0.2373% 1.8934% 3.8470% 6.1474%TABLE 3: Comparison of the two methods for finding the optimal size and the optimal speedr 0.50 0.75 1.00 1.25 1.50 1.75 2.00_ = 4:99IdealSolutionOptimal Size 2.5763 3.8680 5.1608 6.4542 7.7480 9.0420 10.3362Optimal Speed 0.9432 0.9422 0.9413 0.9406 0.9399 0.9394 0.9388Maximal Profit 24.0605 36.0947 48.1539 60.1926 72.2317 84.3121 96.3528ActualSolutionOptimal Size 3 4 5 6 7 9 10Optimal Speed 1.0 1.0 1.0 1.0 1.0 1.0 1.0Maximal Profit 23.8770 35.7921 48.0850 60.1452 72.0928 83.9968 96.2230Relative Difference 0.7695% 0.8454% 0.14355% 0.0789% 0.1927% 0.3754% 0.1349%_ = 5:99IdealSolutionOptimal Size 3.1166 4.6787 6.2418 7.8056 9.3600 10.9346 12.4995Optimal Speed 0.9401 0.9393 0.9386 0.9380 0.9375 0.9370 0.9366Maximal Profit 28.9587 43.4364 57.9339 72.4121 86.9180 101.3958 115.9086ActualSolutionOptimal Size 3 4 6 7 9 10 12Optimal Speed 1.0 1.0 1.0 1.0 1.0 1.0 1.0Maximal Profit 28.9158 43.1208 57.8503 72.2208 86.7961 101.2557 115.7505Relative Difference 0.1484% 0.7317% 0.1445% 0.2649% 0.1405% 0.1384% 0.1365%1 2 3 4 5 6 7 8 9 10111213141516171819202122232425010203040506070The number of serversProfitDQGSQU 100%SQU 99%SQU 92%SQU 85%Fig. 12: Profit versus m and different quality-guaranteedratios.analyzed as follows. When the server speed is changingwithin a small speed range, in order to satisfy the requireddeadline-guaranteed ratio, the number of servers rented bya service provider keeps unchanged. At the beginning, theadded revenue is more than the added cost, so the profit isincreasing. However, when the speed becomes greater, theenergy consumption increases, leading to the total increasedcost surpassing the increased revenue, hence, the profitdecreases.In the third example, we explore the changing trend ofthe profit with different D, and the results are shown as0.1 0.3 0.5 0.7 0.9 1.1 1.3 1.5 1.7 1.9 2.1 2.3 2.5 2.7 2.90102030405060The SpeedProfitDQGSQU 100%SQU 99%SQU 92%SQU 85%Fig. 13: Profit versus s and different quality-guaranteedratios.Fig. 14. Fig. 14(a) gives the numerical results when the serverspeed is fixed at 0.7, and Fig. 14(b) shows the numericalresults when the number of servers is fixed at 5. We analyzethe results as follows.From Fig. 14(a), we can see that the profit obtainedusing the SQU renting scheme increases slightly with theincrement of D. That is because the service charge keepsconstant but the extra cost is reduced whenD is greater. As aconsequence, the profit increases. The second phenomenonfrom the figure is that the curves of SQU 92% and SQU 85%have sharp drop at some points and then ascend gradually0018-9340 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. Seehttp://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI10.1109/TC.2015.2401021, IEEE Transactions on ComputersTRANSACTIONS ON COMPUTERS, VOL. *, NO. *, * 2015 125 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 2530354045505560Deadline DProfitDQGSQU 100%SQU 99%SQU 92%SQU 85%(a) Fixed server speed s = 0:7.5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 253035404550556065Deadline DProfitDQGSQU 100%SQU 99%SQU 92%SQU 85%(b) Fixed server size m = 5.Fig. 14: Profit versus D and different quality-guaranteedratios.and smoothly. The reasons are explained as follows. Whenthe server speed is fixed, enough servers are needed tosatisfy the given quality-guaranteed ratio. By calculating,we know that the number of required servers is the samefor all D values in a certain interval. For example, [5,7] and[8,25] are two intervals of D for the curve of SQU 92%,and the required servers are 10 and 9, respectively. For allD within the same interval, their costs are the same witheach other. Whereas, their actual quality-guaranteed ratiosare different which get greater with the increasing D. Hence,during the same interval, the revenue gets greater as wellas the profit. However, if the deadline increases and entersa different interval, the quality-guaranteed ratio sharplydrops due to the reduced servers, and the lost revenuesurpasses the reduced cost, hence, the profit sharply dropsas well. Moreover, we can also see that the profit of SQU100% is much less than the other scenarios. That is becausewhen the quality-guaranteed ratio is great enough, addinga small revenue leads to a much high cost.From Fig. 14(b), we can see that the curves of SQU 92%and SQU 85% descend and ascend repeatedly. The reasonsare same as that of Fig. 14(a). The deadlines within the sameinterval share the same minimal speed, hence, the cost keepsconstant. At the same time, the revenue increases due tothe increasing quality-guaranteed ratio. As a consequence,the profit increases. At each break point, the minimal speedsatisfying the required quality-guaranteed ratio gets smaller,which leads to a sharp drop of the actual quality-guaranteedratio. Hence, the revenue as well as the profit drops.6.3 Comparison of Optimal ProfitIn order to further verify the superiority of our proposedscheme in terms of profit, we conduct the following comparisonbetween the optimal profit achieved by our DQGrenting scheme and that of the SQU renting scheme in [2]. Inthis group of comparisons, λ is set as 6.99,D is 5, r is varyingfrom 0.75 to 2.00 in step of 0.25, and the other parametersare the same as Section 5. In Fig. 15, the optimal profitand the corresponding configuration of two renting schemesare presented. From Fig. 15(a) we can see that the optimalprofit obtained using our scheme is always greater than thatusing the SQU renting scheme. According to the calculation,our scheme can obtain 4.17 percent more profit on theaverage than the SQU renting scheme. This shows that ourscheme outperforms the SQU renting scheme in terms ofboth of quality of service and profit. Figs. 15(b) and 15(c)compare the server size and speed of the two schemes. Thefigures show that using our renting scheme the capacityprovided by the long-term rented servers is much less thanthe capacity using the SQU renting scheme. That is becausea lot of requests are assigned to the temporary servers usingour scheme, and less servers and slower server speed areconfigured to reduce the waste of resources in idle period.In conclusion, our scheme can not only guarantee the servicequality of all requests, but also achieve more profit than thecompared one.7 CONCLUSIONSIn order to guarantee the quality of service requests andmaximize the profit of service providers, this paper hasproposed a novel Double-Quality-Guaranteed (DQG) rentingscheme for service providers. This scheme combinesshort-term renting with long-term renting, which can reducethe resource waste greatly and adapt to the dynamicaldemand of computing capacity. An M/M/m+D queueingmodel is build for our multiserver system with varyingsystem size. And then, an optimal configuration problemof profit maximization is formulated in which many factorsare taken into considerations, such as the market demand,the workload of requests, the server-level agreement, therental cost of servers, the cost of energy consumption, andso forth. The optimal solutions are solved for two differentsituations, which are the ideal optimal solutions and theactual optimal solutions. In addition, a series of calculationsare conducted to compare the profit obtained by theDQG renting scheme with the Single-Quality-Unguaranteed(SQU) renting scheme. The results show that our schemeoutperforms the SQU scheme in terms of both of servicequality and profit.In this paper, we only consider the profit maximizationproblem in a homogeneous cloud environment, because theanalysis of a heterogenous environment is much more complicatedthan that of a homogenous environment. However,we will extend our study to a heterogenous environment inthe future.0018-9340 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. Seehttp://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI10.1109/TC.2015.2401021, IEEE Transactions on ComputersTRANSACTIONS ON COMPUTERS, VOL. *, NO. *, * 2015 130.75 1 1.25 1.5 1.75 2406080100120140Average rOptimal ProfitDQGSQU(a) Comparison of Profit.0.75 1 1.25 1.5 1.75 205101520Average rOptimal SizeDQGSQU(b) Comparison of Server Size.0.75 1 1.25 1.5 1.75 20.90.920.940.960.981Average rOptimal SpeedDQGSQU(c) Comparison of Server Speed.Fig. 15: Comparison between our scheme with that in [2].