Future Generation Computer Systems ( )

Contents lists available at ScienceDirect

Future Generation Computer Systems

journal homepage: www.elsevier.com/locate/fgcs

Capacity-driven utility model for service level agreement negotiationof cloud servicesNadia Ranaldo, Eugenio Zimeo Department of Engineering, University of Sannio, Italy

h i g h l i g h t s

We propose a capacity-aware utility model to support negotiation of cloud services. The utility function takes into consideration the available resources dynamically. The approach improves the provider utility and reduces SLA violations.

a r t i c l e i n f o

Article history:Received 15 May 2014Received in revised form18 January 2015Accepted 9 March 2015Available online xxxx

Keywords:Cloud computingQoS managementSLANegotiationCapacity planning

a b s t r a c t

Dynamic customers requirements and providers resources availability in the Cloud market make itinadequate static approaches to guarantee Quality of Service (QoS) levels and to define pricing. In thiscontext, negotiation guided by dynamic information is a viable way to achieve high satisfaction levelsfor both contract parties. We propose to exploit capacity planning to support bilateral negotiationprocesses with the aim of optimizing the utility for service providers, by avoiding contracts that couldincur in Service Level Agreements (SLAs) violations, keeping, at the same time, competitive prices. Theproposed technique exploits a non-additive utility function defined in the region of acceptable SLAproposals, taking into account desired QoS and expected resources availability, costs and penalties. Theexperimental analysis shows the benefit of the proposed dynamic approach with respect to static onesin a scenario characterized by a set of customers and differentiated classes of applications provided by acloud environment.

2015 Elsevier B.V. All rights reserved.

1. Introduction

Service Level Agreements (SLAs) [1] represent key elements toachieve full success in Cloud computing, since they represent thedesired guarantees between service providers and customers. SLAsallow to formally describe the offered functions, the QoS levelsthe provider promises to meet, the responsibilities [2] of both thecontract parties, and the penalties applied in case QoS levels arenot satisfied.

Platform as Service (PaaS) providers (e.g. Google App Engineand Force.com), often offer a pool of differentiated services withpre-fixed prices, related to the complexity of the deployed appli-cations, measured through metrics such as the number of applica-tions and database objects. For these services, SLAs are currently

Correspondence to: Department of Engineering, University of Sannio, viaTraiano, Benevento, 82100, Italy. Tel.: +39 0824 305538; fax: +39 0824 30552.

E-mail addresses: ranaldo@unisannio.it (N. Ranaldo), zimeo@unisannio.it(E. Zimeo).

used to define only the granted service availability (uptime) leveland a credit-based penalty system in case of violation. They do notoffer, yet, the possibility to define custom agreements that couldbetter satisfy both customers and providers.

Coarse-grained and static QoS guarantees are no longer satis-factory in a market characterized by continuously changing con-ditions. They, in fact, require providers to quickly react in orderto maintain high levels of competitiveness and customer satisfac-tion (birth of new high competitive providers, customers demandof cloud services for new business fields, fluctuations of electricalpower price, optimal data center resources exploitation).

In this dynamic scenario, negotiation of fine-grained SLAs couldbe a viable approach for service providers to be competitive andto reach more profitable agreements for both customers andproviders [3].

The level of flexibility of the negotiation process depends on theunderlying protocol. It could be (1) unilateral, if a party (typicallythe provider) proposes a SLA and the other one can only decide toaccept or reject it, or (2) bilateral, if both the parties have an activerole in proposing and defining SLAs. The latter allows to resolve

http://dx.doi.org/10.1016/j.future.2015.03.0070167-739X/ 2015 Elsevier B.V. All rights reserved.

2 N. Ranaldo, E. Zimeo / Future Generation Computer Systems ( )

conflicts deriving by different and continuously changing goals,policies andpreferences of customers andproviders throughdialogbetween them.

In many bilateral negotiation strategies, each negotiation actoradopts a decision model based on a utility function, whichrepresents the (perceived) satisfaction level associated to a SLAproposal. In particular, given n negotiable SLA parameters, theutility function assigns a utility value to each point in thecorresponding n-dimensional space of suchparameters. The regionin such space in which the utility value is considered acceptableduring the negotiation process is called acceptable region. Eachpoint in this region represents a SLAproposal andhas a utility valuebetween a minimum and a maximum.

Since customers and providers adopt different utility functionsthat are not known to the counter-parts, an agreement is possibleonly if the intersection between the two acceptable regions, callednegotiation space, is not empty. In this case, an agreement is a pointin the negotiation space where the utility assumes a satisfactoryvalue for both customer and provider.

An agreement is reached through a process, typically basedon time. For example, time-based decision functions [4] allowto make time-dependent concessions with respect to an initialutility value (e.g. the maximum one) with the aim to reach theagreementwithin a prefixednegotiation time. In particular,when aSLA proposal is received from a negotiation party, the parametersvalues are verified against the acceptable region and, if they areadmissible, the related utility value is computed. On the basis ofsuch evaluations and of elapsed time, the strategymakes decisionsabout the acceptance or rejection of the proposal, the counter-proposal generation or negotiation termination.

In the literature, typically, the decision models are basedon multi- and independent-attribute utility functions that areadditive with respect to negotiation parameters, that is, the utilitycan be evaluated considering one parameter at a time, and thetotal utility can be computed by adding (linear combination) theutility contributions derived from the value of each negotiableparameter [5]. With this approach, a SLA proposal is acceptableif each negotiable parameter value is within the correspondinginterval of acceptable values.

In a more realistic Cloud market, some negotiable parameters,such as price and QoS levels, cannot be considered additiveindependent: the service price depends on resources cost, that,in turn, depends on the agreed QoS terms. Moreover, utilityshould take into account strategic business policies and dynamicinformation about the negotiation context, such as markettrend, actual customers requirements and providers resourcesavailability and performance. In fact, before a SLA is signed, theprovider has to check whether the requested set of resourceswill be available when desired, to avoid future SLA violations.Moreover, an offer with the same QoS level and price could beaccepted (refused) on the basis of different conditions: sustainable(not sustainable) service usage conditions (e.g. the forecasted dailyload peak) and a high (low) competitive market phase, also in caseof potential economic loss.

In this paper, we focus on bilateral SLA negotiation of PaaSservices for hosting multi-tier Web applications in a scenariowhere the number of users is variable and the workload is notstationary but, typically, exhibits peaks and dipswith daily, weeklyor also seasonal cycles [6,7].

In order to meet QoS terms, the provider allocates appropriateresources to each tier of the application architecture. Currently,we adopt replication only at the application server tier, whilea Web server is used as a load balancer and a single databaseserver is shared. Thanks to virtualization technologies, replicationis dynamically managed by a resource management system whichhandles a set of independent and homogeneous virtual machines(the overall Cloud provider capacity).

The virtual machines, allocated on a set of hardware resourcesof the provider data center, are exploited to host various instancesof the application server. The number of virtual machines,allocated to the application server tier of each signed SLA, changesdynamically during the day by means of a predictive resourceallocation mechanism. This mechanism aims to define the bestresource allocation plan able to maximize the profit and to avoidQoS violations under a daily fluctuating workload.

The proposed utility model, which dynamically defines theacceptable regions on the basis of available capacity, is used atprovider-side to guide negotiation strategies. To this aim, theadopted utility function is non-additive to represent the overallprovider economic profit deriving by a new contract, net of costof assigned capacity, penalty payment in case of QoS guaranteesviolations and eventual variation in profits of already signed SLAs.

Utility is a function of two negotiable parameters, which are thecontract price (una-tantum payment) and the maximum responsetime that can be perceived by the end-user without incurring ina penalty, and other non-negotiable parameters (constraints andpre-conditions). These constraints and preconditions are definedby both customer (contract duration and starting day, applicationcomponent size, forecasted daily workload plan) and provider(service availability and penalty).

Price is a function of capacity cost and market conditions.Capacity cost depends on the product virtual machines x daily timeslots assigned to a SLA, whereas market conditions (monopoly vscompetition) are captured by using two factors that express (a) theprobability to choose that provider and (b) the possible profit.

From the considerations above, the proposed utility function isbased on effective customers requirements, specified in the initialnegotiation phase (as pre-conditions), and on a capacity planningtechnique, which suggests the best profitable resources allocationplan for every new SLA by avoiding (or reducing) violations.

To validate the proposal and to show the benefit in predictingutility of new potential contracts, an in-depth experimentalanalysis has been conducted.

1.1. Main contribution

To the best of our knowledge, this is the first proposal thatadopts capacity planning in the first phase of a contract life-cycleto guide bilateral negotiation strategies through the definition ofthe providers acceptable region and utility value. By adoptingthis approach, the provider reduces the risk of incurring in SLAviolations since the technique allows to find actual free slots in theglobal resource allocation plan,which is defined by considering theresources needed to satisfy all the signed SLAs. Our proposal, unlikethe traditional ones based on additive and static utility functions,allows the provider to propose, during negotiation processes,offerswith competitive prices and feasible performance.Moreover,it maintains the potential violation of QoS terms under fixedtolerable levels and avoids the stipulation of new contracts in casethey conduct to unprofitable revenues or customer unsatisfaction.

Our proposalwas firstly presented in [8], inwhich a preliminaryexperimental validation, based on a simple linear applicationperformance model was adopted to investigate the proposedutility function and the capacity-driven evaluation technique. Amore realistic experimental scenario and the comparison of theproposed approach with the traditional one based on additiveutility functions have been presented in [9].

This paper extends both the previous ones, giving deeper detailsabout the utility function formalization and the heuristic adoptedfor its evaluation, and presenting an in-depth experimentalanalysis to validate the approach. The analysis shows theachievement of high satisfaction levels for both providers andcustomers: providers can gain advantages both in the short period

N. Ranaldo, E. Zimeo / Future Generation Computer Systems ( ) 3

(economic profit) and in long period (customer loyalty), whereascustomers can stipulate contracts with more competitive prices andwith greater guarantees on QoS terms fulfillment.

The rest of the paper is organized as follows. Section 2 discussesrelated work. Section 3 introduces the SLA model for the PaaS ser-vice and the capacitymodel. Sections 4 and 5 present, respectively,the proposed utility model and a dynamic evaluation techniquebased on capacity planning. Section 6 shows the benefit of the pro-posed technique with respect to static approaches through exper-imental analysis. Finally, Section 7 presents conclusion and futurework.

2. Related work

Many bilateral negotiation strategies proposed in the literatureare based on a multi-attribute additive utility function, assumingthat the negotiable parameters are additive independent of eachother. The most popular form is given by a linear combinationof linear utility functions defined for each negotiation parameterand normalized within the corresponding acceptable interval ofvalues [4]. Moreover, the utility functions and related acceptableregions are defined statically and require human intervention[1012], limiting the applicability of those approaches in highlydynamic environments, such as Cloud.

The paper [13] proposes the adoption of non-additive multi-objective utility functions for satisfying both business and perfor-mance goals in unilateral negotiation of Cloud services for Grid. Theutility function takes into account various objectives (economicrevenue maximization and reputation, priority to tasks or servicesexecuted in off-peak hours). When a provider receives a proposal,the utility function is maximized, taking into account economicfactors and resources availability information, to propose an offerto the customer.Wepropose a similar approach (based on themax-imization of a non-additive utility function) butwe define dynamicacceptable regions based on short-term capacity planning formoreflexible bilateral negotiation processes.

Spillner and Schill in [14] propose the semi-automatic adjust-ment of SLA templates, published by providers in an advertise-ment service registry and adopted as starting point of negotiationprocesses. Their approach is based on a performance predictionmodel, that exploits both run-time and historical monitoring data,to define the sustainable QoS level before reaching the resourcelimit and eventually incurring in SLA terms violations. Differentlyfrom this proposal, which requires manually adjustment of SLAtemplates by providers, we propose a capacity planning tech-nique to drive a negotiation process that follows a business policyautomatically.

Some papers face with bilateral negotiation mechanisms basedon non-additive utility functions, which are more realistic thanadditive ones. As an example, the paper [15] proposes twonegotiation models able to handle non-monotonic and discretenon-linear utility functions, based respectively on a multipleoffer and an approximating mechanism. The paper [16] proposesthe definition of a bilateral negotiation protocol based on analternative projection strategy for generating offers with the aimto reach, in a finite time, a satisfying agreement among agentscharacterized by a generic nonlinear utility function. However,these papers adopt pre-defined nonlinear utility functions and donot face with their realistic modeling that should take into accountcapacity availability and cost, customers requirements andmarketconditions.

Capacity planning of IT infrastructures, both for optimizedshort-term resource management and long-term investmentplans, can be employed by service providers to manage SLAs andpromised QoS levels in the most profitable way [17].

The problem of a self-adaptive capacity planning for optimiz-ing economic profits related to SLA of Internet Services has beeninvestigated in the literature. Some approaches leverage the queu-ing theory to solve an optimization resource allocation problemunder constraints on the service rate. In particular, the paper [18]takes into account the profit with respect to the penalty and [19]the reward in case a surge workload is supported. The paper [20]proposes a capacity planning method based on a queuing networkapproach and an analytical performance model of Cloud service topredict the optimal configuration of a Cloud application, consider-ing both the provider profit and the customer satisfaction levels. Asin our proposal, resource virtualization for performance isolationand dynamic resource allocation are exploited. However, these pa-pers apply capacity planning techniques to optimally manage datacenter capacity, by considering a set of signed SLAs, but not to de-fine the offer space and the potential utility for the provider of newSLAs under negotiation.

The paper [21] deals with the dynamic capacity allocationof a data center to a hosted application on the basis of theworkload demand, in order to reduce SLA violations and powerconsumption. To this aim, historical workload traces are analyzedto identify a base daily pattern represented by an aggregation(discretization) of the original workload time-series. The patternis defined through a dynamic programming problem aiming tominimize the discretization error and the number of intervals inthe discretization. Overloading is handled with capacity increaseat short-term scale.

In our paper,we consider customerworkload as a pre-conditionfor SLA negotiation and capacity planning is adopted in thisinitial phase of contract life-cycle. The technique proposed in thepaper cited above could be used in our solution during resourceutilization phase (after a contract has been established) and todefine a new customer profile (workload definition) for successiveSLA negotiations.

3. A PaaS for web hosting

Resource allocation strategies of a PaaS can be influencedby application requirements. We classify applications accordingto two dimensions: the functional and technological ones. Theformer affects the number of users simultaneously accessingthe application, the way they interact with it, and, as aconsequence, the aggregated request rate. This dimension isdefined by the following sub-dimensions: the application class,the interactive level (interactive-intensive vs. batch processing),and geographic extension (national vs. continental or world-wide spread notoriety). The technological dimension characterizesthe way the data center resources are mainly exploited bythe application. It affects the performance pattern exhibited byexecuting the application on a pool of resources. The relatedsub-dimensions are: architecture (monolithical vs. multi-tierarchitecture), and memory management (volatile vs. persistence-intensive data management).

We focus on a PaaS for hosting Web applications, calledVirtual Web Platform (VWP) service. A VWP service offers a virtualplatform used to host interactive-intensive Web applications withnational geographic extension, composed ofmultiple components,deployed on the provider resources according to a multi-tierarchitecture and with a volatile-intensive data management.

3.1. SLA model

The SLA model is structured in four sections: (1) servicedescription, (2) QoS target, (3)measurement and (4) penalty model.

4 N. Ranaldo, E. Zimeo / Future Generation Computer Systems ( )

The service description defines the contract validity period,denoted as number of days D, the starting day B, the total serviceprice P , expresses in euros (e).

The QoS target section defines the QoS guarantee terms and theservice usage conditions under which the provider is responsiblefor such guarantees terms (pre-conditions). Pre-conditions aredefined through the workload plan W expected for the Webapplication with a daily pattern, modeling the typical requestrate fluctuation of Web applications [6,7]. A day is partitioned inK disjoint time intervals, each characterized by a representativeworkload value wk, k : 1, . . . , K . wk is expressed as number ofrequests per second (r/s) and is based on the number of requestsreceived and completed during the k-th time slot.

We consider two QoS guarantee terms: (1) the response timemust not overcome the maximum response time T expressedin seconds (s); (2) the service availability, expressing the timepercentage during which the service is capable of being used, mustbe equal or greater than the valueMinAvail defined by the provider.The service is considered available when the difference betweenthe measured response time and T is less than the maximumallowed value, indicated withtmax.

The measurement section defines the measurement processesfor response time, workload and service availability in order tomonitor service usage conditions and QoS guarantee terms.

The response time is influenced by various delay components,some of them are not under the provider responsibility (such asdata transfer delay outside the data center performed on publicnetworks used without contract regulation). For this reason, wedefine the response time as the time interval beginning from theHTTP request receipt on the provider infrastructure to the HTTPresponse completion and transfer beginning. Moreover, since theresponse time depends on the processing time of the specific in-voked Web component, in order to obtain comparable measures,we introduce a customized Web component, called BenchmarkingWeb Component (BWC), defined by the customer to characterizetheWeb application in terms of typical operations load. Finally, wedefine as measurement sample the average of a set of single mea-surements retrieved during a small interval time, calledmonitoringtime unit. This approach avoids to detect, as QoS violation, isolatedperformance degradations during transitory situations, typical ofadaptation actions on resource allocation.

The penalty model defines the penalty that the provider mustpay when the QoS guarantee terms are not satisfied. We considera penalty directly proportional to price P, QoS violation degreeand duration. In particular, it is expressed as the summation ofelementary penalties eventually derived from each monitoringtime unit. Indicated with Pendkj, d : 1, . . . ,D, k : 1, . . . , K , j :1, . . . , J the penalty in the j-th monitoring time unit of the k-thtime slot of the d-th contract day, the total penalty, Pen, is given by:

Pen =D

d=1

Kk=1

Jj=1

Pendkj. (1)

Pendkj depends on the difference between the measuredresponse time and T , denoted as tdkj, and on price related to amonitoring time unit, denoted as p, as follows:

Pendkj =

0 if tdkj 0p

tdkjtmax if 0 < tdkj < tmax

p if tdkj tmax

, (2)with > 0, p = P

DKJ.

A summary of parameters characterizing an SLA signed andmanaged by a provider, and the manner in which they are defined,is reported in Table 1.

Fig. 1. Capacity model.

3.2. Capacity model

To meet QoS guarantee terms, the Cloud provider of VWPservices must allocate appropriate resources to each tier of thehosting platform of multi-tier Web applications. Currently, weadopt a replication schema to the application server tier, while theWeb server is used as a load balancer and a unique database serveris shared.

The mapping of an application server replica onto hardware re-sources of the provider data center is managed by a resource vir-tualization layer so that each virtual machine hosts an applicationserver (see Fig. 1). Under this assumption, we model the overallCloud capacity as the set ofM independent and homogeneous vir-tual machines that can be simultaneously launched on the datacenter hardware resources. We assume that such virtual machineshave the same provisioned performance, measured through an ap-plication benchmarking technique.

The number of virtual machines allocated to the applicationserver tier of each signed SLA changes, during the day and for eachcontract day, adopting a dynamic resource provisioning technique.Such technique is, currently, guided by the solution to a utilityoptimization problem which defines the best capacity allocationplan able to maximize the profit and to avoid QoS violationsunder different workloads provisioned in daily time slots. Theoptimization problem, that is the same adopted for the definitionof acceptable regions and utility values of SLA proposals duringnegotiation phase, is formulated in Section 4 and a heuristic to findits solution is discussed in Section 5.

The capacity performance model adopted by the utilityoptimization problem is based on a benchmarking technique thatexploits the measurement process described in Section 3.1. Itdefines the performance function t(n, w), a function that returns theresponse time than can be reached assigning n virtual machines tothe application server tier under workloadw. On the other hand, amore accurate provisionalmodel, based onmonitoring data duringservice operation and/or Web workload modeling techniques,could be adopted in the future, with the aim to improve estimationand reduce service costs (e.g. by avoiding energy consumption ofassigned but not-exploited resources).

Given R = {SLAi}, i : 1, . . . ,N the set of already signed SLAs,the capacity assigned (called also resource allocation plan) to the

N. Ranaldo, E. Zimeo / Future Generation Computer Systems ( ) 5

Table 1SLA parameters.

Fixed by the customerD Contract validity durationB Starting day of the contract validitywk, k : 1, . . . , K Workload rate that must be supported during the k-th daily time slot

Negotiated T Maximum response timeP Contract price (all-inclusive)

Fixed by the provider MinAvail Minimum service availability (percentage)tmax Maximum difference between the measured response time and T to consider a service available.

VWP service related to each SLAi is denoted by NRi = {nRidk}, i :1, . . . ,N, d : 1 :, . . . ,D, k : 1, . . . , K . It defines the number ofvirtual machines assigned in each time slot k of each contract dayd. In the following, we consider the same time slot partitioning (K)and contract validity period (D and B) for all SLAs.

The fraction of capacity allocated to a VWP service has an eco-nomic cost for the provider, that wemodel through a cost functionlinearly proportional to the resource allocation plan. In particular,the capacity cost, Ci, related to a signed SLA, SLAi, is given by:

Ci = cD

d=1

Kk=1

nidk, (3)

where c represents the elementary cost for virtual machine usageper time slot.

4. Utility model

The proposed utility model evaluates the profit that theprovider achieves by accepting a new contract with the relatedSLA, indicated as SLAN+1, taking into account QoS guarantee terms,service usage conditions, capacity availability and utility derivingfrom each already signed contract in R.

The utility, denoted by U(TN+1, PN+1), deriving by the newcontract is defined as the difference between the overall profitaccommodating the new contract and the one gained by the al-ready signed contractswithout accepting the newone. Adopting anadditive model with respect to the profit deriving from each singleSLA, utility is given by:

U(TN+1, PN+1) =N+1i=1

UQi (SLAi,NQi )

Ni=1

URi (SLAi,NRi ), (4)

where Q denotes the union of R (already signed SLAs) with SLAN+1,and NQi = {nQidk} the resource allocation plans of SLAs consideringthe new set Q of signed SLAs. URi and U

Qi are the contract utility de-

riving by each SLA considering respectively R and Q as the set ofsigned SLAs. Generically, contract utility Ui is given by:

Ui(SLAi,Ni) = Pi Ci Peni, (5)where Pi is the contract price, Ci is the cost of assigned capacity andPeni is the provisioned penalty, all depending on SLAi and Ni.

In order to vary the price in response to changingmarket supplyand demand, we adopt a dynamic price function, whose minimumvalue is the cost of the resources identified by the allocation plan(capacity cost Ci). In particular, Pi is defined as

Pi = Ci( des+ (1 des)g), > 1, 0 < des 1, 0 g 1, (6)where the profit factor , evaluated on the basis of historicalmarket data about accepted/average price, expresses the possibleprofit deriving by a new contract, given its cost. The interestlevel of the provider in signing a new contract, that captures themarket model (monopoly vs competition), is modeled throughthe parameter des, calculated as the probability for a provider

to be chosen by a customer among the available ones. Price isinversely proportional to des, in fact, the higher is des, the higher iscompetitiveness among providers, and as a consequence, lower isthe price to increase customers interest. Both and des are time-dependent factors, since information about market conditionsvaries in time.

Finally, g is a factor, useful during the negotiation process, tovary the price between the minimum value, obtained with g =1 (the capacity cost, Ci), and the maximum allowed one, giventhe current market conditions, obtained with g = 0 (the valueCi des).

Peni is evaluated on the basis of the provisioned applicationperformance captured by the performance function t(n, w),already introduced in Section 3.2, adopting the assigned capacityand the workload plan declared by the customer.

The acceptable region, indicated with AR, represents the regionin the bi-dimensional space of negotiable parameters (TN+1, PN+1)for the SLA proposals of the new contract, whose utility U isacceptable. For themulti- and independent-attribute utility modelproposed in [4], the acceptable region is composed of all proposalswhose parameters values are within their respective acceptableintervals defined by means of the related static minimum andmaximum values. On the contrary, for the proposed utility model,negotiation parameters are not independent, in particular theinterval of acceptable prices depends on the cost of the resourceallocation plan, that, on its turn, depends on the desiredmaximumresponse time.

We formalize AR as follows: indicated with [Tmin, Tmax] = IntTthe interval of acceptable response times, called acceptable per-formance interval, a proposal (TN+1, PN+1) belongs to AR if TN+1 iscontained within the acceptable performance interval and if PN+1belongs to the interval of acceptable prices related to TN+1, calledacceptable price interval, indicatedwith [Pmin(TN+1), Pmax(TN+1)] =IntP(TN+1):

AR = {(TN+1, PN+1) : TN+1 IntT and PN+1 IntP(TN+1)}. (7)

5. A heuristic for utility evaluation based on capacity planning

The proposed heuristic adopts SLA parameters of a negotiationrequest and dynamic information about capacity availability todecide whether the request can be handled and, in positive case,to evaluate the acceptable region and utility value of proposals inthat region. In particular, the hosting of a new service is guided bythe principle of optimizing utility (4): given a proposal with certainvalues for price and response time, a capacity planning problemis performed in order to find the optimal resource allocation planthat allows to obtain the best utility value, taking into accountthe available resources in various time slots of the contract validityperiod and the utility gained by the already signed SLAs. Moreover,it is necessary to define the conditions under which such utility(and the related proposal) is considered acceptable. The problemoffinding the best resource allocation plan NQN+1, related to SLAN+1is formulated as follows:

max(U(TN+1, PN+1)), (8)

6 N. Ranaldo, E. Zimeo / Future Generation Computer Systems ( )

constrained by

N+1i=1

nQidk M, d,k. (9)

nQidk noptidk (Ti,Wi), i,d,k. (10)nQidk 0, nQidk nmax, nQidk N,i,d,k (11)whereNopti (Ti,Wi) = {noptidk (Ti, wik)}, i : 1, . . . ,N+1, is the optimalresource allocation plan of SLAi given a certain maximum responsetime Ti and workload plan Wi. It is the allocation plan that, atminimum cost, allows to obtain a response time less that Ti in eachtime slot k under the related workloadwik:

noptidk (Ti,Wi) = min(min(n : t(n, wik) Ti), nmax). (12)where nmax is themaximumnumber of assignable virtualmachinesto a VWP service in a time slot.

The optimization problem (8) finds the resource allocationplans for all SLAs that maximize utility (4), taking into accountthe effective available capacity (constraint (9)). Constraint (10)maintains at minimum the cost for the new contract, allowing toavoid waste of resources and to offer competitive prices. It statesthat, for each contract, the number of assigned virtual machines, ineach time slot, must be less or equal than the optimal one.

We consider two resource allocation policies, the conservativeand the progressive one. With the former, the allocation plan forthe new service is spread out on effective available resources anddoes not affect the allocation plan of the already signed services.With the progressive allocation policy the hosting of a new servicetakes into account changes in the allocation plans for the alreadystipulated contracts, so potentially causing a variation in their cost,penalty and utility. In the latter case, in addition to constraints(9)(11), other constraints can be formulated to limit re-allocationactions that could cause uncontrolled reduction of utility derivingby each contract and related customer satisfaction level. Theseconstraints are, for example, limitations on the maximum numberof virtual machines that can be subtracted to the already signedSLAs and on the maximum performance degradation.

The utility optimization problem (8) can result in negativeor positive values of utility, in case the proposal leads to a lossof profit for the provider or to an effective gain, respectively.The overall business policy is responsible to dynamically guidethe decision whether a proposal is satisfactory or not, with acertain competitiveness level. In particular, we adopt the followingparametric conditions under which a proposal (TN+1, PN+1) isdefined acceptable:

Response time acceptability condition: a response time TN+1 is ac-ceptable if the utility of the proposal (TN+1, Pmax(TN+1)) is equalor greater than percentage PUmax of the utility Uoptmax that canbe gained with NoptN+1(TN+1,WN+1) and the related maximumprice:

U(TN+1, Pmax(TN+1)) PUmax Uoptmax. (13)Anhigh value of PUmax reduces the risk tied to penalty paymentwhen the optimal allocation plan cannot be assigned at all. Thiscondition influences the acceptable performance interval and,in particular, Tmin.

Price acceptability condition: price PN+1 is acceptable, in rela-tion to TN+1, if the corresponding utility is included betweenthe minimum and the maximum one. In particular, the maxi-mum utility, UQmax(TN+1), is the one corresponding to the max-imum allowed price (g = 0 in (6)), and the minimum utility,UQmin(TN+1), is a percentage, PUmin, of UQmax(TN+1):

PUmin UQmax(TN+1) U(TN+1, PN+1) UQmax(TN+1). (14)

Service availability condition: the percentage of nQN+1dk whose re-lated response time overcomes TN+1 more than tmax must beless thanMinAvail. Considering all the contract validity durationD, it can be expressed as follows:

nQn+1dk : (t(nQn+1dk, wN+1k) TN+1) > tmax< MinAvail DK . (15)

5.1. A heuristic

In order to evaluate the utility in a computationally feasiblemanner, we propose a heuristic aiming to find an approximation ofthe utility function and of related acceptable region solving (8) for alimited number of cases and exploiting an interpolation technique.

The algorithm consists of the following main steps:

Step 1: Evaluation of Tmax, IntP(Tmax), and the utility consideringTmax and the boundaries of IntP(Tmax); Step 2: Evaluation of Tmin, IntP(Tmin), and the utility consideringTmin and the boundaries of IntP(Tmin); Step 3: Evaluation of a set of Z response timeswithin IntT , calledTzs, each IntP(Tz) and utility considering Tzs and the boundariesof related IntP(Tz).

Step 1Tmax is the maximum response time provisioned in the K time

slots adopting the minimum number nmin of assignable virtualmachines:

Tmax = max(t(nmin, wN+1k))K . (16)Since we consider interactive-intensive applications, we limit

the value of Tmax to an upper bound, called TBOUND, beyond whichwe consider the response time not acceptable for the customer.IntP(Tmax) is evaluated through an iterative approach. The first stepis to solve (8) adopting the conservative allocation policy. In thiscase, (4) becomes:

U(TN+1, PN+1) = PN+1 CN+1 PenN+1. (17)In order to maximize (17), we find the allocation plan NQN+1

adopting a best effort capacity planning approach taking intoaccount constraints (9)(11):

NQN+1 = {nN+1dk}

=max

M

N+1i=1

nQidk

, noptN+1dk(Tmax,WN+1)

. (18)

Then, if the service availability condition for NQN+1 is satisfied,CN+1 is evaluated adopting (3). Further, the response timeacceptability condition (13) is checked evaluating Pmax(Tmax) andthe relative penalty, adopting respectively (6) with g = 0 and (1):Pmax(Tmax) = CN+1 des. (19)

If such condition is satisfied, the price acceptability condition(14) is adopted to evaluate Pmin(Tmax). Considering the conservativeresource allocation policy, it can be evaluated directly startingfrom Pmax(Tmax), since different prices do not influence the bestassignable allocation plan for problem (8).

If the service availability and the response time acceptabilityconditions are not satisfied, the progressive allocation policy istaken into account. The basic idea is the following: for each timeslot in which the number of allocated virtual machines, nQN+1dk, isless than both nmax and the optimal number n

optN+1dk, we find a re-

allocation plan, involving the already signed SLAs, that leads to thegreatest utility increase. The process is stopped when the accept-ability conditions are satisfied. If the re-allocation actions do not al-low to satisfy the acceptability conditions, this means that the new

N. Ranaldo, E. Zimeo / Future Generation Computer Systems ( ) 7

(a) Office. (b) Business. (c) Private.

Fig. 2. Normalized workload plans for various applications classes.

SLA does not lead to an acceptable utility for any value of price andresponse time. In this case the acceptable region cannot be definedand the current negotiation request is refused.Step 2

Tmin is evaluated adopting an iterative approach aiming tofind the minimum response time that satisfies the acceptabilityconditions. Initially, Tmin is defined as the minimum response timeobtained in each time slot:

Tmin = min(t(nN+1dk, wN+1k))D,K . (20)Assuming to adopt the conservative allocation policy and amonotonically increasing trend of the application performancefunction, (20) can be expressed as follows:

Tmin = min(t(nN+1k, wN+1k))K ,

with nN+1k = min

M Ni=1

nidk

, nmax

D

, k. (21)

If the acceptability conditions are not satisfied, the progressiveallocation policy is exploited. If, also in this case, the acceptabilityconditions are not satisfied, greater values of Tmin are attempted. Inparticular, an attempt value is obtained summing to the previousone a little amount > 0, until the acceptability conditions aresatisfied or Tmax is reached. If a Tmin (Tmax ), > 0, is found,the heuristic proceeds to the next step. On the contrary, it stopsand the negotiation request is refused.Step 3

Because of non-linearity of the model, the utility function isevaluated for the set TZ of response times internal to IntT . A simpletechnique to define TZ is based on the partition of the interval IntTinto equal-length parts (but other solutions could be adopted):

Tz = Tmin + z Tmax Tminz + 1 . (22)For each Tz, IntP(Tz) and utilities considering the boundaries of

such interval are evaluated adopting the acceptability conditionsin a similar manner to Step 1.

Given a generic proposal (TN+1, PN+1), it is considered withinthe acceptable region under the following conditions:

Tmin TN+1 Tmax, Pmin(TN+1) PN+1 Pmax(TN+1).Pmin(TN+1) and Pmax(TN+1) are evaluated adopting an interpo-

lation technique which takes into account the distance betweenresponse time TN+1 and the nearest response times in the set {Tz}.

6. Experimental results and discussion

In this section, we analyze the non-linear behavior of theproposed non-additive utility model with respect to price and

response time, its dependency on the workload plan and availablecapacity, and the heuristic accuracy.Moreover, we discuss the ben-efit that the dynamic evaluation approach introduces with respectto the static one in terms of satisfaction level for both providers andcustomers. To this end, we compare the dynamic approach withtwo static ones: (1) the additive multi-parameter (AMP) utilitymodel, statically defined, typically adopted by negotiation strate-gies driven by time-based decision functions [4]; (2) the proposedutility model, statically evaluated before processing any negotia-tion request.

6.1. Experimental setup

The experimental results are related to the conservativeresource allocation policy and the following parameters:

the contract validity period D (the same for all the contracts)has a duration of 180 days;

static competitive (balanced demand/supply) market condi-tions, characterized by a provider interest level des = 0.5;

= 4 for price evaluation, c = 4.17 103 e, for costevaluation;

regarding the service availability condition, MinAvail = 100%and tmax = T : the service has to be always available and aproposal is not acceptable for the provider if there is a time slotinwhich the number of assignable resources leads to a responsetime that overcomes 2T . In this case, the service acceptabilitycondition can be expressed as follows:

nQidk : t(nQidk) 2T , i,d,k. (23) TBOUND = 2 s; PUmax = 30% and PUmin = 10%; = 5 ms and = 5 ms, for Tmin definition; nmax = 2000, nmin = 1; the number of virtual machines initially available in each dailytime slot isM = 20 000;

Z = 28.We define the daily workload pattern for three application

classes (see Section 3.1): (1) office: support of office work and ter-tiary industry, (2) business: support to production cycle of manu-facturing industry, and (3) private: services for e-commerce, onlinebanking, games, news and entertainment portals, connecting peo-ple services, social networks, etc. Fig. 2 shows the normalizedworkload plans W = {w1, . . . , w12} for the three applicationclasses, considering the partition of a day in 12 time slots of twohours-duration. The patterns are based on the analysis of dailytraces of real applications reported in [21] and their discretizationwith respect to time slots. The absolute workload plans adopted inour experimentation are obtained scaling the normalized ones byfixing the peak workload value, called V .

The application performance function of each application classis based on a queuing performance model, frequently adopted toabstract a multi-tier application hosted in a data center [22,23].

8 N. Ranaldo, E. Zimeo / Future Generation Computer Systems ( )

(a) Acceptable region. (b) Maximum utility.

Fig. 3. Acceptable region and maximum utility varying workload plan.

(a) Acceptable region. (b) Maximum utility.

Fig. 4. Acceptable region and maximum utility varying available capacity.

We use an M/M/n/PS queue to capture the mean response timereachable adoptingnhomogeneous virtualmachines at applicationserver-tier receiving incoming requests from a unique queuemanaged by the HTTP load balancer. In particular, indicated withts the mean service time of the BWC component, and with w themean request rate, both service time and request rate describedby a Poisson distribution, the application performance function isgiven by:

t(n, w) = 11ts wn

. (24)

The minimum number of virtual machines required to reach agiven certain response time is obtained by (24) approximating theresult to the nearest integer number. For queue stability, incomingworkloadw has to satisfy the following condition:

w nts. (25)

If such condition is not satisfied, we assume the service isunavailable. In the following for each negotiation request ts =0.049 s.

6.2. Utility model analysis

Fig. 3 shows results of utility model evaluation consideringvarious workload plans within the related acceptable performanceinterval. In particular, Fig. 3(a) shows the two-dimensional

acceptable region while Fig. 3(b) shows the maximum utility forthe three workload patterns and for two peak values, 15 000 and35 000 r/s. In Fig. 3(a) we note that for proposals characterizedby workloads with V = 15 000 r/s it is possible to reach lowerresponse times with respect to workloads with V = 35 000 r/s.Moreover, the maximum acceptable prices decrease by increasingresponse times with a trend that is mainly influenced by theperformance model adopted in the experimentation. In particular,for response times close to the best reachable response time (ts),the optimal allocation plans are high expensive (high resourcesdemand especially in time slots with workload near the peak) and,consequently, prices reach high values. On the contrary, for highresponse times (close to TBOUND) the maximum acceptable pricesslowly decrease towards a value that is influenced by the allocationplan made of the minimum number of resources required in eachtime slot to satisfy the queue stability condition.

As it is possible to note in Fig. 3(b), for workloads with hugepeak values, the maximum utility that potentially can be reachedis higher than the one reachable with lower peak values. This istied to the necessity to assign more expensive allocation plans,that leads to higher costs and consequently higher prices and po-tential gains. Moreover, the maximum utility trend is increasingin the first part, reaches a maximum and then decreases. This isdue to the impact of costs and penalty (performance degradation)that for lower response times are huger and consequently limitsthe profits. Fig. 4 shows the influence of available capacity on util-ity model. In particular Fig. 4(a) shows the acceptable region and

N. Ranaldo, E. Zimeo / Future Generation Computer Systems ( ) 9

Fig. 5. Percentiles of relative error in logarithmic scale varying number ofevaluations Z .

Fig. 4(b) shows the maximum utility considering the office work-load pattern with V = 35 000 r/s and a number of available vir-tual machines in each time slot varying from 1716 to 2000, that arerespectively theminimum number of virtual machines required toaccept the negotiation request and themaximumassignable one ineach time slot. Increasing the number of available virtualmachinesin each time slot takes to lower response timeswithing the accept-able performance intervals and higher maximum utility value, be-cause ofmore expensive allocation plans and, consequently, higherprices and lower penalty (lower performance degradation).

The proposed utility model allows to evaluate the provisionedutility for a certain price P and response time T within the accept-able region, adopting an interpolation technique on a finite number(Z + 2) of exact utility evaluations. The accuracy error introducedby this approach is mainly influenced by the non-linearity of theutility function and by the number Z of exact evaluations. In orderto evaluate the accuracy error, we compare the provisioned utilityU(T , P) with the exact one. The exact utility Uexact(T , P) is definedas the utility provisioned adopting the actual best allocation planassignable to a new contract. It is obtained solving optimizationproblem (8) taking into account the conservative allocation policy,SLA parameters, current capacity availability and already signedSLAs. Given Uexact , we define the relative error, E, as:

E = |Uexact U|Uexact

100. (26)Fig. 5 shows the percentiles, from the 1-th to the 100-th, of therelative error varying Z from 0 to 50. The samples are obtainedconsidering 200 values of response times uniformly distributedwithin the acceptable performance interval, and for each responsetime, 200 price values uniformly distributed within the respectiveacceptable price interval. The reported results refer to the officeworkload pattern with V = 35 000 r/s. We observe that with Z =30 the 80% of samples reports a relative error lower than 0.6%whilethe 90% of samples about 2%. More and more increasing Z , relativeerror remains significative for a more and more decreasing num-ber of samples.We can state that such results are satisfying for sup-portingnegotiation strategies, because theutility provision of a SLAfor a new contract can tolerate a limited degree of inaccuracy forthe benefit of a feasible computational complexity.

6.3. Dynamic versus static approaches

In this section, we evaluate the effectiveness of the proposeddynamic approach in increasing customer satisfaction level andprovider profit and reputation with respect to static approaches.To this aim, we define two parameters, the price-based indicator,

UC (P), as the customer satisfaction level with respect to theprice, and the performance-based indicator, UC (T ), as the customersatisfaction level with respect to agreed maximum response timeT and the actual response time during the contract validity period.Denoted with UCP (P) the provider profit perceived by the customerwith the cost awareness of the optimal allocation plan (Costopt),given by:

UCP (P) = P Costopt , (27)the customer satisfaction level is maximum when P is equal toCostopt , while it is minimum when the maximum price, Pmax isapplied.

UC (P) is defined as the normalized form of (27) as follows:

UC (P) = UCP (P) UCP (P)min

UCP (P)max UCP (P)min. (28)

(28) allows to obtain values within the interval [0, 1] for priceswithin the interval [Costopt , Pmax]. For prices less than Costopt ,UC (P) is greater than 1, for prices greater than Pmax it becomes neg-ative. As a consequence, UC (P) is adopted as an indicator that aproposal is in the customer acceptable region and represents a po-tential negotiation point if its value is within the interval [0, 1].

Since in our experimentation scenario Pmax = 2 Costopt , (28)becomes:

UC (P) = 2 PCostopt . (29)

Denoted with Tdk, d = 1, . . . ,D, k = 1, . . . , K , the actualresponse time in the k-th time slot of d-th day with the assignedallocation plan, we introduce the following parameter UCP (T ), use-ful to represent the performance degradation perceived by thecustomer:

UCP (T ) =D

d=1

Kk=1

(T Tdk),

Tdk =Tdk T , if (Tdk T ) > 00, if (Tdk T ) 0

, d,k.

(30)

WhenTdk = 0,UCP (T ) has the best value, on the contrary, it isat minimum level, but still acceptable, when a maximum degrada-tion level is reached. Defining such level in a proportional way to Tby means of the factor deg, UC (T ), expressed as the normalizationform of UCP (T ), assumes values within the interval [0, 1]within thetwo above-mentioned boundary cases. In particular, it is given by:

UC (T ) = UCP (T ) UCP (T )min

UCP (T )max UCP (T )min

= 1

Dd=1

Kk=1

Tdk

deg T . (31)

For performance equal or better than the agreed one, (31) is1, for a degradation level greater than the maximum allowed, itbecomes negative and indicates a certain level of provider reputa-tion reduction.

6.3.1. Non-additive versus additive utility modelThe additive utility function, called UAMP(T , P), adopted for

comparison with our approach, is given by a linear combination oflinear utility functions defined for each negotiation parameter and

10 N. Ranaldo, E. Zimeo / Future Generation Computer Systems ( )

Fig. 6. Acceptable region (AR) evaluated for the static utility model (consideringthe office workload pattern and peak value V 25 000 r/s), and for the InnerAMP and Outer AMP models, two additive multi-parameter utility modelsevaluated considering as acceptable region, respectively, the rectangle inscribedand circumscribed AR.

Table 2Parameters of the Inner AMP and Outer AMP models.

Inner AMP Outer AMP

Tmin [s] 0.27 0.07Tmax [s] 2.0 2.0Pmin [e] 6 780 5 672Pmax [e] 10 313 24 504

normalized within the corresponding static acceptable interval ofvalues [4]. In particular, it is given by:

UAMP(T , P) = wTVT (T )+ wPVP(P),VT (T ) = T TmaxTmax Tmin , VP(P) =

P PmaxPmax Pmin ,

wT + wP = 1, wT > 0, wP > 0,(32)

whereVP(P) andVT (T ) are, respectively, the linear utility functionswith respect to price P andmaximum response time T , normalizedin the related interval of acceptable values [Pmin, . . . , Pmax] and[Tmin, . . . , Tmax]. For such model, the acceptable region is therectangle in the bi-dimensional space (T , P), whose projection onaxes is given by the interval of acceptable values of T and P . Inthis region, the utility varies linearly within the value 0 (for thepoint (Tmin, Pmin)) and the value 1, gained with respect to the bestprofitable value of each negotiable parameter, corresponding tothe point (Tmax, Pmax).

The comparison was conducted considering the static evalua-tion of the proposed model, considering a well-defined negotia-tion request. In particular, such model, called static utility model,is evaluated using parameters defined in Section 6.1, consideringa capacity availability of M = 20 000 and a negotiation requestcharacterized by the office workload pattern and V = 25 000 r/s.

In Fig. 6 we plot the acceptable region AR of the static util-ity model and of two AMP models characterized by differentacceptable regions correlated with the acceptable region (AR) ofthe static utility model. For the first, called Inner AMP model, theacceptable region corresponds to the rectangle that roughly in-scribes AR. For the second, called Outer AMP model, the acceptableregion corresponds to the rectangle that roughly circumscribes AR.Table 2 summarizes parameters that characterize the acceptableregion and utility function (32) of the two AMP models, for whichwe consider the weights wP andwT both equal to 0.5.

While with the additive approach, utility has an increasinglinear trend varying response time andprice,with the non-additiveapproach utility has a non-linear trend, as it is possible to note inFig. 7, in which we compare utility evaluated with the static model

Fig. 7. Utility evaluation for the proposed non-additive model and Inner AMP andOuterA MP models fixing price and varying response time.

and with the Inner AMP and Outer AMP models, considering twofixed values of prices (24 504 e and 5672 e) and varying responsetime.

A comparison of utility estimations for both the approaches,considering the vertices of acceptable regions of Inner AMP andOuter AMP models, is reported in Table 3. In particular, the tablereports: utility U provisioned by the static utility model normal-ized within the acceptable region AR and utility UAMP provisionedby the AMP model related to the vertices under study; if a point(T , P), corresponding to a SLA proposal, is feasible or not for ne-gotiation adopting the static utility model. Such condition, calledPotential Negotiation Point Feasibility (FEANP), is verified if the SLAproposal is included in acceptable region AR. When a SLA pro-posal is external to AR, it is not included in the negotiation processfor the agreement achievement with the customer. Moreover, thetable reports Uexact , that is the effective utility evaluated adoptingthe best allocation plan resulting from the optimization problem(8), the provisioned penalty and the value of the parameter UC (P).

Analyzing the results, we can state that the Inner AMP modelproduces utility estimates more similar to the proposed approachthan the Outer AMP model. On the other hand, the Inner AMPmodel corresponds to a low-risk approach, rejecting all pointscharacterized by high performance, with the disadvantage ofprecluding the chance of high profits that can be gained withhigh demanding customers, that are ready to pay much for highperformance services. On the other hand, the Outer AMP modelcorresponds, on one hand, to a high-risk approach, since highperformance is offered for low prices, and on the other one, to aout-of-market behavior, because of too high prices required for lowperformance. Themain estimation error for static AMP approachesare pointed out for points (0.07 s, 5672 e) and (2 s, 24 504 e),that, instead, are not feasible for the proposed approach, because ofrespectively too low price (negative utility and Uexact ) and too highprice (utility greater than 1 and negative UC (P)). Finally, we canconclude that, with respect to the AMP utility model, the proposedapproach has the advantage of reducing the acceptable region tothe proposals with feasible performance and competitive prices,and so it can be effectively adopted by an integrative negotiationstrategy to quickly reach an agreement as good as possible for bothcustomer and provider.

6.3.2. Proposed dynamic versus static approachesWe conducted a comparative analysis between the static and

dynamic evaluation of the proposed utility model to show its ca-pability in leading towards negotiation agreements profitable forboth provider and customer. In particular, for this experimenta-tionwe consider the static utilitymodel, as defined in Section 6.3.1,

N. Ranaldo, E. Zimeo / Future Generation Computer Systems ( ) 11

Table 3Non-additive versus AMP utility model.

T [s] P [e] U FEANP UAMP Uexact [e] Pen [e] UC (P)!FEANP

Outer AMP

0.07 5 672 1.19 !FEANP 0 8 965 2 376 1.542.0 5 672 0.04 FEANP 0.5 516 0 0.900.07 24 504 0.23 FEANP 0.5 1 968 10 263 0.002.0 24 504 2.48 !FEANP 1 19 348 0 2.75

Inner AMP

0.27 6 780 0.05 FEANP 0 616 0 0.902.0 6 780 0.18 FEANP 0.5 1 623 0 0.690.27 10 312 0.51 FEANP 0.5 4 150 0 0.332.0 10 312 0.64 FEANP 1 5 156 0 0.00

that remains the same during a temporal sequence of negotiationprocesses, each launched to respond to a customer request for aVWP service negotiation. The dynamic utility model, on the con-trary, corresponds to the proposed model evaluated every time anegotiation request is received, taking into account the effectivecustomer requirements and current capacity availability. We con-sider a sequence of negotiation requests (called NRs), each char-acterized by the same contract validity period, starting day andapplication performance features, but differentiated on the basisof distinct workload plans (different application classes and peakvalues). Since our experimentation does not focus on the evalua-tion of a specific negotiation strategy, we simulate the negotiationprocess of a NR in the following manner: the negotiation strategyleads to a final result, called potential negotiation point (PNP), corre-sponding to a point (T,P) randomly chosen in the acceptable region,and the utility model guides the decision if the NR and the relatedPNP could lead to the actual signing or not of a new contract.

In particular, given a certain NR and the related PNP, the utilitymodel guides the following decisions: Negotiation Request Acceptance (NRACC): a NR is accepted andleads to a negotiation process when it is possible to define theacceptable region. Typically, a NR is refused when the provideravailable capacity makes not convenient to negotiate for anewcontractwith certain application performance features andworkload plan features;

Potential Negotiation Point Feasibility (FEANP): as alreadydefinedin the previous section, a PNP is feasible if it is included in thecurrent acceptable region. If the NR is accepted and related PNPis feasible, such PNP can be proposed by the provider to thecustomer for contract stipulation, but it becomes an effectiveagreement if both the following conditions are met:

Positive Price-based Indicator (POSUCP): UC (P) 0, consideringthe actual capacity availability;

Satisfied Service Availability (SSA): in each time slot the numberof available resources is enough to satisfy the service availabilitycondition. If such condition is not satisfied, this means that thePNP is accepted for contract stipulation and that the providerrealizes to not be able to satisfy QoS terms only later, when theresourcemanagement system tries to allocate resources to hostthe new service.

If one of these two conditions is not satisfied, we supposethat the contract is not stipulated at all, and that the providerreputation is affected negatively, since the customer is notsatisfied with the provider behavior: in case of violation of thePOSUCP condition, the provider is proposing too high prices,while in case of violation of the SSA condition, the providerinitially proposes a PNP and then refuses to agree on it.

Finally, if a contract is stipulated, the customer is satisfiedwith respect to QoS guarantee terms if the following conditionis verified:

Positive Time-based Indicator (POSUCT): UC (T ) 0, consideringthe actual capacity availability and deg = 0.1KD, that corre-sponds to an average tolerable performance degradation of 10%with respect to the agreed level.

We conducted a comparative analysis mainly evaluating thetotal provider utility (supposed to be zero at the beginning),estimated by the static and the dynamic utility model at the endof a negotiation sequence, and comparing it with Uexact . Moreover,we evaluated the negative impact on the provider reputation in thecase the utility model is not able to prevent situations leading tocustomer unsatisfaction. In particular, we evaluate the percentageof NRs, called customer unsatisfaction percentage (CUP), for whichthe customer is not satisfied, corresponding to the cases in whichthe conditions NRACC and FEANP are satisfied and at least oneof the condition POSUCP, SSA or POSUCT are not satisfied. Ahigh value of the customer unsatisfaction percentage means thatthere are many cases in which the customer is unsatisfied and soindicates a highnegative impact of the adoptedutilitymodel on theprovider reputation. If such percentage is very low, thismeans thatmost of the negotiation requests and related negotiation points,considered acceptable and feasible, lead to customer satisfaction.This, as a consequence, indicates a very low negative impact ofthe utility model on the provider reputation and an increase ofcustomer loyalty (the provider earns a higher probability of beingselected for future contract stipulations).

Experimental results refer to a sequence of 50 NRs, whose asso-ciated PNPs are spread on the acceptable region, AR, of the staticutility model. Different experiments were conducted consider-ing different techniques to choose, randomly, the parameters Tand P of each PNP. In this paper we report the results obtainedconsidering for parameter T a sequence following a uniform dis-tribution within the acceptable performance interval IntT. Suchsequence allows to simulate the stipulation of contracts with var-ious QoS levels that, in order to be satisfied, require different ca-pacity allocations. To each response time of the sequence of PNPs,is associated a price P adopting a normal distribution within theacceptable price interval related to such response time. The nor-mal distribution for prices was chosen since it more realisticallymodels potential negotiation agreements between a provider anda customer around medium prices. Table 4 reports comparison re-sults considering four sequences of NRs, characterized by the fol-lowing workload plans:

1. The workload plan of each NR is the same adopted for the staticutility model (office workload pattern, V = 25 000 r/s);

2. The workload patterns of various workload plans derive froma discrete uniform distribution of application classes (office,business and private) and a uniformdistribution of V within theinterval [10 000 r/s, 40 000 r/s] (with avg V = 25 000 r/s);

3. like (2) but considering the interval [10 000 r/s, 30 000 r/s],(with avg V = 20 000 r/s);

4. like (2) but considering the interval [20 000 r/s, 40 000 r/s],(with avg V = 30 000 r/s).In each scenario, the initial number of available virtual ma-

chines in each time slot isM = 20 000.For the first sequence of NRs, until the number of available re-

sources is greater or equal to the optimal one, necessary to grant

12 N. Ranaldo, E. Zimeo / Future Generation Computer Systems ( )

Table 4Comparison of dynamic and static utility model for various sequences of NRs.

Workload Pattern V [r/s] Sta. Dyn. U [e] Uexact [e] Pen [e] Avg (E) Std (E) SLAsigned CUP %

Office 25 000 Sta. 47 720 47 911 0 0.4 1.2 15 70Dyn. 47 720 47 911 0 0.4 1.2 15 0

Uniform Distr. avg 25 000 Sta. 51 527 35 484 2723 72.7 76 16 70Dyn. 38 841 38 821 0 0.05 0.1 16 0

Uniform Distr. avg 20 000 Sta. 58 123 55 626 0 30.4 24 18 64Dyn. 55 517 55 626 0 0.14 0.3 18 0

Uniform Distr. avg 30 000 Sta. 51 916 29 548 1844 406 659 16 70Dyn. 50 349 50 544 0 0.47 1.6 18 0

the maximum response time of the PNP, the static and dynamicutility models give the same results. In particular, the first 15 NRsare accepted, their related PNPs are feasible and take to contractstipulations, the three conditions POSUCP, SSA and POSUCT are sat-isfied, no penalty payment are provisioned, the total provisionedutility is the same (47 720 e) and the average value and standarddeviation of relative error E of the utility estimation for each con-tract is very low. After the 15-thNR, the resource occupation is 0.99in the time slot 7 (involved by the peak workload), and remaining36 virtual machines are not enough to grant service availability forany other NR. Since the static utility model does not point out thissituation (it does not take into account effective capacity availabil-ity), it considers each NR further to 15-th acceptable and relatedPNP feasible. As a consequence, with the static approach such NRslead to the violation of SSA condition, causing a high provider rep-utation loss (CUP = 70%). On the contrary, the dynamic utilitymodel, taking into account the effective capability availability, re-fuses from the 15-th to the last NR, since NRACC condition is notsatisfied, and avoids any negative influence on provider reputation(CUP = 0%).

For the second sequence of NRs, regards the static approach, wecan note a huge relative error for single utility estimations (the av-erage value is 72.7%), that causes a huge discrepancy between thefinal provisioned utility (51 527 e) and the exact one (35 484 e)at the end of the negotiation sequence. Such results are causedby the wrong estimations performed by the static approach whenworkload plans are different from the one adopted for the staticutility model evaluation. In particular, for lighter workload plans(lower V and less demanding patterns), the static utility is lessthan the exact one since the price acceptable intervals are eval-uated for huger (more expensive) allocation plans than effectivelynecessary to grant QoS guarantee terms, and vice versa, for hugerworkload plans, it is greater that the exact one. In the first case,the static approach can incur in the violation of POSUCP condition,since price of PNP is too high with respect to the cost of the ef-fectively required allocation plan and the utility becomes greaterthan the maximum allowed one. In the second case, performancedegradation (and penalty), not correctly provisioned, can happenand further influence utility estimation error. On the contrary, thedynamic approach, taking into account the effectively requiredworkload plan of the NR and current capacity availability, ensuresaccurate utility provisions as for the first scenario. Summarizing,with the static approach 16 NRs lead to contract stipulations, whilethe remaining ones cause: 1 violation of the POSUCT condition(with UC (T ) = 1.82), 6 violations of the POSUCP condition, 27violations of the SSA condition, that take to a high provider reputa-tion degradation (CUP = 70%). In the dynamic approach, instead,beyond to 16 contract stipulations, the remaining ones are imme-diately discarded for violation of FEANP condition (9 NRs) and ofNRACC condition (25 NRs), so avoiding any negative influence onprovider reputation.

The third and the fourth sequence point of NRs point out thesame advantages of the dynamic approach with respect to the

static one, already noticed for the second sequence, in a moreevident manner. For such sequences, in fact, the average value ofV is chosen so as to be, respectively, less and greater that 25 000r/s, the value adopted for the evaluation of the static utility model.In particular, in the third scenario, there are a lot of NRs for whichthe workload plan is characterized by a peak value lower than 25000 r/s. As a consequence, for many NRs (in particular 13 NRs)the static approach accepts the related PNPs as feasible and causesthe violation of POSUCT condition. On the contrary, the dynamicapproach correctly discard them by means of the violation ofFEANP condition. Among the first 31 NRs, 18 of them are acceptedand considered feasible by both the approaches, but the averageestimation error is 30.4% for the static approach and 0.14% forthe dynamic one. The last 19 NRs are rejected by the dynamicapproach (because of violation of NRACC condition) since availablecapacity is very scarce (91 virtual machines in the 7-th timeslot, characterized by the peak workload, after the 18th agreedcontract). On the contrary, they are accepted by the static approachand lead to the violation of SSA condition, since such approach isnot able to detect, in advance, the lack of resources required tosatisfy QoS requirements of new contracts. Summarizing, in thestatic approach there is a great customer unsatisfaction percentage(64%), while in the dynamic approach it still remains to 0%.

In the last scenario, the utility estimation errors are very high forthe static approach (both in excess and in defect), because of highlyvariable workload plans of NRs (see Table 4). In particular, Fig. 8compares the provisions performed by the static and dynamicapproaches for the first 31 NRs. In case a NR is rejected, it islabeledwith !NRACC,while in case it is accepted and SSA conditionis satisfied, the bar of related provisioned utility is plotted andeventual violations of other conditions are labeled. The NRs thatlead to provider reputation loss are graphically shown using thedotted contour for the utility bar.

The figure points out as the dynamic approach is able to rejectNRs with huge workloads that lead to negative (for 1, 3, 7, 19 and23 NR) and very low (for 6, 10 and 13 NR) utilities, because of lowprices of the related PNPs or lack of available capacity. Moreover,the first NR, with a workload plan with business application pat-tern and V = 38 260 r/s, is also characterized by a low perfor-mance level (UC (T ) = 0.15) and a penalty provision of 1844 e.Adopting the static approach the above-mentioned NRs lead tocontract stipulation, leading to an high difference between the finalutility estimation (51 916e) and the exact one (29 548e).With thedynamic approach, more contracts are stipulated than the staticone (18 versus 16) and a much higher final utility (50 544 e ver-sus 29 548 e) is reached without any performance degradation. Fi-nally, as for the previous sequences, when the capacity availabilitybecomes insufficient to grant SSA condition, only the dynamic ap-proach is able to a-priori discard further NRs, so avoiding any rep-utation loss. This happens for NRs from the 16-th to the last NR,except the 24-th.

N. Ranaldo, E. Zimeo / Future Generation Computer Systems ( ) 13

Fig. 8. Static and dynamic utility evaluation for the sequence of NRs whose workload plans are characterized by a uniform distribution of application classes and of peakvalues with average of 30 000 r/s. Utility bars with dotted contour represent NRs that lead to provider reputation loss.

7. Conclusion

In this paper, we have proposed a technique based on capacityplanning to support Cloud providers in bilateral negotiation ofhigh-level QoS parameters and prices related to PaaS services.The technique aims at achieving high satisfaction levels for bothproviders and customers through a heuristic that dynamicallyevaluates a non-additive utility function and the acceptableregion, by taking into account application performance, capacityavailability and a price function based on cost and market.

By adopting a queuing-based performancemodel andworkloadpatterns based on real daily workload traces, the experimentalanalysis has demonstrated that the proposed solution leads theprovider to accurately predict the utility that can be gained by newcontracts in order to avoid their stipulation in case they conduct tounprofitable revenues or customer unsatisfaction.

The current proposal aims at optimizing the provider side (evenif it takes into consideration customer satisfaction). In fact, it needsthe declaration of customer workload as a pre-condition for SLAnegotiation. The declaration of bad workload has not negativeimpacts on providers satisfaction (utility), but may influencecustomer utility (unfulfillment of QoS terms at customer side, if thedeclared workload is under-estimated with reference to the actualone, or unuseful costs, if the declared workload is over-estimatedwith reference to the actual one).

Therefore, we are planning to investigate a progressive alloca-tion policy and amore sophisticated approach for the dynamic def-inition of resource allocation plans assigned to new contracts, ableto modify the initial allocation plan, resulting from the proposedutility optimization problem, to a better exploitation of data cen-ter resources. In particular, we will exploit machine learning algo-rithms to learn from monitoring data with respect to both actualincoming workload of hosted applications and resource perfor-mance [24] in order to model customer workload profile.

We are also investigating an integrative negotiation strategybased on time-based decision functions able to quickly reach anagreement with high satisfaction levels for both providers andcustomers.

Acknowledgment

This paper is partially supported by Italian Ministry ofEducation, University and Research within the framework of PRINIDEAS Integrated Design and Evolution of Adaptive Systems grant number J38C13001510001.

References

[1] L. Wu, R. Buyya, Service level agreement SLA in utility computing systems,in: V. Cardellini, E. Casalicchio, K. Castelo Branco, J. Estrella, F. Monaco (Eds.),Grid and Cloud Computing: Concepts, Methodologies, Tools and Applications,IGI Global, 2012, pp. 286310.

[2] A. Andrieux, K. Czajkowski, A. Dan, K. Keahey, H. Ludwig, T. Nakata, J. Pruyne,J. Rofrano, S. Tuecke, M. Xu, Web Services Agreement Specification (WS-Agreement), Tech. Rep., Global Grid Forum, GRAAP WG, 2005.

[3] K. Czajkowski, I. Foster, C. Kesselman, Agreement-based resource manage-ment, Proc. IEEE 93 (3) (2005) 631643. http://dx.doi.org/10.1109/JPROC.2004.842773.

[4] H. Raiffa, The Art and Science of Negotiation, Harvard University Press, 1982.[5] P. Wakker, Additive Representations of Preferences: A New Foundation of

Decision Analysis, Kluwer Academic Publishers, Dordrecht, Boston, London,1989.

[6] M. Arlitt, C. Williamson, Internet Web servers: workload characterizationand performance implications, IEEE/ACM Trans. Netw. 5 (5) (1997) 631645.http://dx.doi.org/10.1109/90.649565.

[7] A. Williams, M. Arlitt, C. Williamson, K. Barker, Web workload characteriza-tion: Ten years later, in: X. Tang, J. Xu, S. Chanson (Eds.), Web Content Deliv-ery, in:Web Information Systems Engineering and Internet Technologies BookSeries, vol. 2, Springer US, 2005, pp. 321.

[8] N. Ranaldo, E. Zimeo, Exploiting capacity planning of cloud providersto limit SLA violations, in: The 3rd International Conference on CloudComputing and Services Science, CLOSER 2013, 2013, pp. 184195.http://dx.doi.org/10.5220/0004377001840195.

[9] N. Ranaldo, E. Zimeo, Capacity-aware utility function for SLA negotiationof cloud services, in: IEEE/ACM 6th International Conference on Utility andCloud Computing, UCC 2013, Dresden, Germany, December 912, 2013, 2013,pp. 292296. http://dx.doi.org/10.1109/UCC.2013.58.

[10] H. Li, S. Su, H. Lam, On automated e-business negotiations: Goal, policy,strategy, and plans of decision and action, J. Org. Comput. Electron. Commer.16 (1) (2006) 129. http://dx.doi.org/10.1080/10919390609540288.

[11] M. Chhetri, J. Lin, S. Goh, J. Yan, J.Y. Zhang, R. Kowalczyk, A coordinatedarchitecture for the agent-based service level agreement negotiation of Webservice composition, in: The 17thAustralian Software Engineering Conference,ASWEC 2006, 2006, pp. 9099. http://dx.doi.org/10.1109/ASWEC.2006.1.

[12] F. Zulkernine, P. Martin, An adaptive and intelligent SLA negotiationsystem for Web services, IEEE Trans. Serv. Comput. 4 (1) (2011) 3143.http://dx.doi.org/10.1109/TSC.2010.44.

[13] M. Macas, J. Guitart, Using resource-level information into nonadditivenegotiationmodels for cloudmarket environments, in: 12th IEEE/IFIPNetworkOperations and Management Symposium, NOMS 2010, 2010, pp. 325332.

[14] J. Spillner, A. Schill, Dynamic SLA template adjustments based on serviceproperty monitoring, in: IEEE International Conference on Cloud Computing,CLOUD 2009, 2009, pp. 183189. http://dx.doi.org/10.1109/CLOUD.2009.56.

[15] F. Ren, M. Zhang, Bilateral single-issue negotiation model consideringnonlinear utility and time constraint, Decis. Support Syst. 60 (0) (2014) 2938.Automated negotiation technologies and their applications. http://dx.doi.org/10.1016/j.dss.2013.05.018.

[16] R. Zheng, N. Chakraborty, T. Dai, K. Sycara, M. Lewis, Automated bilateralmultiple-issue negotiation with no information about opponent, in: 201346th Hawaii International Conference on System Sciences (HICSS), 2013,pp. 520527. http://dx.doi.org/10.1109/HICSS.2013.626.

[17] J. Allspaw, The Art of Capacity Planning: Scaling Web Resources, OReillyMedia, Inc., 2008.

14 N. Ranaldo, E. Zimeo / Future Generation Computer Systems ( )

[18] J. Almeida, V. Almeida, D. Ardagna, C. Francalanci, M. Trubian, Resourcemanagement in the autonomic service-oriented architecture, in: The 2006IEEE International Conference on Autonomic Computing, ICAC 2006, 2006,pp. 8492. http://dx.doi.org/10.1109/ICAC.2006.1662385.

[19] B. Abrahao, V. Almeida, J. Almeida, A. Zhang, D. Beyer, F. Safai, Self-adaptive SLA-driven capacity management for Internet services, in: The 10thIEEE/IFIP Network Operations and Management Symposium, NOMS 2006,2006, pp. 557568. http://dx.doi.org/10.1109/NOMS.2006.1687584.

[20] Y. Kouki, T. Ledoux, SLA-driven capacity planning for cloud applications,in: 2012 IEEE 4th International Conference on Cloud Computing Technologyand Science 0, 2012, pp. 135140. http://doi.ieeecomputersociety.org/10.1109/CloudCom.2012.6427519.

[21] A. Gandhi, Y. Chen, D. Gmach,M. Arlitt,M.Marwah, Hybrid resource provision-ing for minimizing data center SLA violations and power consumption, Sus-tain. Comput. Inf. Syst. 2 (2012) 14. http://dx.doi.org/10.1016/j.suscom.2012.01.005.

[22] J. Bi, Z. Zhu, R. Tian, Q. Wang, Dynamic provisioning modeling for virtualizedmulti-tier applications in cloud data center, in: The 3rd IEEE InternationalConference on Cloud Computing, CLOUD 2010, 2010, pp. 370377. http://dx.doi.org/10.1109/CLOUD.2010.53.

[23] D. Menasce, M. Bennani, Autonomic virtualized environments, in: The 2006IARIA International Conference on Autonomic and Autonomous Systems, ICAS2006, 2006, http://dx.doi.org/10.1109/ICAS.2006.13.

[24] D. Huang, B. He, C. Miao, A survey of resource management in multitierWeb applications, IEEE Commun. Surv. Tutor. 16 (3) (2014) 15741590.http://dx.doi.org/10.1109/SURV.2014.010814.00060.

Nadia Ranaldo received the Ph.D. degree in ComputerScience from University of Sannio, Benevento, Italy, in2005. She is a Research Assistant in the Department ofEngineering, University of Sannio. Her main researchinterests include frameworks for distributed systems,parallel computing, wireless and sensor networks, re-sourcemanagement and capacity planning, Grid andCloudcomputing.

Eugenio Zimeo received the M.S. degree in ElectronicEngineering from the University of Salerno, Italy, and thePh.D. degree in Computer Science from the Universityof Naples, Italy, in 1999. Currently, he is an AssociateProfessor at the University of Sannio, Benevento, Italy. Hisprimary research interests include software architecturesand frameworks for distributed systems, high perfor-mancemiddleware, service oriented, Grid and Cloud com-puting, and wireless sensor networks. He has publishedabout 90 scientific papers in journals and conferences ofthe field and heads many large research projects.

Capacity-driven utility model for service level agreement negotiation of cloud servicesIntroductionMain contribution

Related workA PaaS for web hostingSLA modelCapacity model

Utility modelA heuristic for utility evaluation based on capacity planningA heuristic

Experimental results and discussionExperimental setupUtility model analysisDynamic versus static approachesNon-additive versus additive utility modelProposed dynamic versus static approaches

ConclusionAcknowledgmentReferences