cloud service negotiation in internet of things environment a mixed approach

Cloud Service Negotiation in Internet of ThingsEnvironment: A Mixed Approach

Xianrong Zheng, Student Member, IEEE, Patrick Martin, Kathryn Brohman, and Li Da Xu, Senior Member, IEEE

Abstract—Internet of Things (IoT) allows connected objects tocommunicate via the Internet. IoT can benefit from the unlimitedcapabilities and resources of cloud computing. Also, when coupledwith IoT, cloud computing can in turn deal with real world things ina more distributed and dynamic manner. As the cloud marketbecomes more open and competitive, Quality of Service (QoS) willbemore important. However, cloud providers and cloud consumershave different, and sometimes opposite, preferences. If such aconflict occurs, a Service Level Agreement (SLA) cannot be reachedwithout negotiation. A tradeoff negotiation approach can outper-form a concession approach in terms of utility, but may incur morefailures if information is incomplete. To balance utility and successrate, we propose a mixed approach for cloud service negotiation,which is based on the “game of chicken.” In particular, if one isuncertain about the strategy of its counterpart, it is best to mixconcession and tradeoff strategies in negotiation. To evaluate theeffectiveness of this approach, we conduct extensive simulations.Results show that a mixed negotiation approach can achieve ahigher utility than a concession approach, while incurring fewerfailures than a tradeoff approach.

Index Terms—Cloud computing, Internet of Things (IoT), mixednegotiation approach, Quality of Service (QoS).

I. INTRODUCTION

I NTERNET OF THINGS (IoT) is expected to be a world-wide network of interconnected objects [7]. IoT allows

objects like computers, sensors, mobile phones, etc. to commu-nicate via the Internet. It is characterized by limited capacitiesand constrained devices, and its development depends on newtechnologies including cloud computing. IoT can benefit fromthe unlimited capabilities and resources of cloud computing.Also, when coupled with IoT, cloud computing can in turn dealwith real world things in amore distributed and dynamicmanner.

In this sense, IoT and cloud computing can complement eachother.

Cloud services are Internet-based IT services. Infrastructure asa Service (IaaS), Platform as a Service (PaaS), and Software as aService (SaaS) are three representative examples [1], [23].Compared with other models, cloud services are easier to accessand use, cost-efficient, and environmentally sustainable. As theyeliminate large upfront expenses in hardware and expensivelabor costs for maintenance, cloud services are beneficial tosmall- and medium-sized enterprises. Moreover, large-sizedenterprises with computationally intensive tasks can obtainresults quickly, since their applications can scale up promptly.

As the cloud market becomes more open and competitive,Quality of Service (QoS) will be more important. However,cloud providers and cloud consumers have different and some-times opposite preferences. For example, a cloud consumerusually prefers a high reliability, whereas a cloud provider mayonly guarantee a less than maximum reliability in order to reducecosts and maximize profits. If such a conflict occurs, a ServiceLevel Agreement (SLA) cannot be reached without negotiation.Automated negotiation occurs, when software agents negotiateon behalf of their human counterparts. It has been studied inelectronic commerce and artificial intelligence for many yearsand is considered as the most flexible approach to procureproducts and services [10].

In bilateral negotiation, negotiation strategies are critical. Tocreate a proposal, a negotiation agent can adopt two strategies—concession and tradeoff. Our previous studies show that atradeoff negotiation approach can outperform a concession onein terms of utility [31], [32]. However, if information is incom-plete, it may cause miscalculations, and so underperform theconcession one in terms of success rate. To balance utility andsuccess rate, we, in this paper, present amixed approach for cloudservice negotiation, which is based on the “game of chicken.” Inother words, if a party’s counterpart uses a concession strategy, itis best to adopt a tradeoff one; if a party’s counterpart uses atradeoff strategy, it is best to adopt a concession one; and if aparty is uncertain about the strategy of its counterpart, it is best tomix concession and tradeoff. In fact, those are the three Nashequilibria of a negotiation game with two pure strategies.

The paper’s main contributions are as follows.1) A multi-attribute bilateral negotiation mechanism involv-

ing five quality attributes. To accommodate these attri-butes, we adopt nonlinear utility functions and redesign theconcession and tradeoff strategies.

2) A mixed negotiation approach based on the “game ofchicken,” which can balance utility and success rate. Inparticular, if a party has no knowledge of which strategy

Manuscript received September 19, 2013; revised December 31, 2013;accepted January 22, 2014. Date of publication February 20, 2014; date ofcurrent version May 02, 2014. This work was supported in part by the NaturalSciences and Engineering Research Council of Canada (NSERC), in part by theSocial Sciences andHumanitiesResearchCouncil ofCanada (SSHRC), in part bythe National Natural Science Foundation of China (NSFC) under Grant71132008, and in part by the U.S. National Science Foundation (NSF) underGrant SES-1318470 and Grant 1044845. Paper no. TII-13-0645.

X. Zheng and P. Martin are with the School of Computing, Queen’sUniversity, Kingston, ON K7L 2N8, Canada (e-mail: [email protected]; [email protected]).

K. Brohman is with the School of Business, Queen’s University, Kingston,ON K7L 3N6, Canada (e-mail: [email protected]).

L. D. Xu is with the Institute of Computing Technology, Chinese Academy ofSciences, Beijing 100190, China; with Shanghai Jiao Tong University, Shanghai200240, China; with the University of Science and Technology of China, Anhui230026, China; and also with Old Dominion University, Norfolk, VA 23529USA (e-mail: [email protected]).

Color versions of one ormore of the figures in this paper are available online athttp://ieeexplore.ieee.org.

Digital Object Identifier 10.1109/TII.2014.2305641

1506 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 10, NO. 2, MAY 2014

1551-3203 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

that its counterpart will play, it is best to mix concessionand tradeoff in negotiation.

3) Extensive simulations to evaluate the effectiveness of themixed negotiation approach. We first test the impact ofdifferent parameters on negotiation results and then con-ductMonte Carlo simulations. Results show that themixednegotiation approach can achieve a higher utility than aconcession approach, while incurring fewer failures than atradeoff approach, which demonstrates its effectiveness.

The rest of the paper is structured as follows. Section IIreviews related work. Section III presents a motivating examplewith conflicts requiring negotiation. Section IV introduces theutility functions we use to model agents’ preferences. Section Vdescribes multi-attribute bilateral negotiations where concessionand tradeoff strategies are detailed. Section VI proposes a mixedapproach for cloud service negotiation, which is based on the“game of chicken.’ Section VII reports and analyzes simulationresults. Section VIII summarizes the paper and mentions poten-tial limitations.

II. RELATED WORK

IoT allows connected objects to communicate via the Internet,whereas cloud computing promises unlimited resources deliv-ered over the Internet [7], [8]. Zhou et al. [33] review the state ofthe art of integrating IoT and cloud computing and propose acloud-based IoT platform to facilitate things applicationdevelopment.

In conducting service research, many ideas and methods havebeen proposed [21], [25], [27], [28]. QoS is important indiscovering, selecting, and composing Web services [12],[19], grid services [18], and cloud services [3], [14], [20]. Liet al. [7] report that commercial cloud services are not yet stableand ask for more attention to the performance, reliability,scalability, and security issues of cloud services. Wang et al.[24] argue that QoS and SLAs are increasingly emphasized inenterprise cloud services, and automated SLA and adaptiveresource management are needed.

Automated negotiation occurs when software agents negotiateon behalf of their human counterparts. It has been studied inartificial intelligence and electronic commerce for many years[4], [10]. Jennings et al. [5] argue that negotiation is the mostfundamental mechanism to manage runtime dependenciesamong agents, and thus underpins cooperation and coordination.Lomuscio et al. [10] argue that automated negotiation underpinsthe next generation of electronic commerce systems, and developa classification scheme for negotiation in electronic commerce. Itoffers a systematic basis on which different negotiation mechan-isms can be compared and contrasted.

Automated negotiation has been proposed as an idealapproach to procure cloud resources and services. Sim [16]proposes a market-driven, agent-based negotiation mechanismto procure resources and services in the cloud marketplace. Inparticular, the mechanism supports parallel negotiation activi-ties, namely, multiple negotiations at the same time. It is reportedthat with it, agents can achieve more utility and a high successrate. Stantchev and Schröpfer [17] propose an approach for SLAmapping between business processes and IT infrastructures. It

aims to formalize, negotiate, and enforce QoS requirements forcloud services. However, no negotiation approaches are speci-fied. Yaqub et al. [29] present a generic negotiation platform forSLA@SOI (Service-Oriented Infrastructure). However, it focus-es on negotiation protocols and not the negotiation strategies thatwe deal with in this paper.

III. MOTIVATING EXAMPLE

Internet startups are able to reside on a cloud to build theirservices even without their own infrastructure. A storage cloudallows users to store their data in data centers without worryingabout backup, such that they can focus on their core businesses.Amazon Simple Storage Service (Amazon S3), Microsoft Win-dows Azure Blob Storage (Azure Blob), and Aliyun OpenStorage Service (Aliyun OSS) are three well-known storageclouds [30].

Here, we present a motivating example, where a StorageConsumer (SC) negotiates over QoS with a Storage Provider(SP). It contains conflicts that cannot be resolved withoutnegotiation. Suppose that, five attributes, i.e., Availability(AVAL), Reliability (REL), Responsiveness (RESP), Security(SECY), and Elasticity (ELAS), are used to describe a storagecloud, as shown in Table I. The numbers are built upon ourexperiences with real-world storage clouds [32]. Refer to [32] forthe definitions and the metrics of the five attributes.

Assume that the minimum and the maximum availability are0% and 100%, respectively. The SP has the ability to provide anavailability of 99%, but it only wants to offer an availability of84%. In other words, its preferred value and reserved value ofavailability are 84% and 99%, respectively. As a higher avail-ability means higher resources consumption, it makes sense forthe SP to offer a reasonable level of availability. In contrast, theSC wants to have an availability of 95%, but if not possible, anavailability of 80% is acceptable too. In other words, its preferredvalue and reserved value of availability are 95% and 80%,respectively. As a higher availability indicates a better quality,it is reasonable for the SC to expect a higher level of availability.Here, a conflict over preferred values of availability (95% vs.84%) occurs between the two parties.

It is also shown in Table I that for the SC, availability is ahigher-is-better attribute, for which a symbol is assigned besideits preferred values. By contrast, for the SP, availability is alower-is-better one, for which a symbol is assigned beside itspreferred values. However, the two parties differ in their pre-ferences over availability. The SP puts a weight of 0.20 onavailability, whereas the SC places a weight of 0.10 on it. Forconciseness, we list corresponding numbers for other attributesin Table I, without going into details.

Indeed, the conflicts shown in Table I cannot be resolvedwithout negotiation. Human negotiation becomes inefficient anderror-prone, due to the inclusion of five attributes, complexpreferences, and time constraints. This necessitates automatednegotiation. Here, we make two assumptions.

Assumption 1 (Preference Gap): Assume that there exists apreference gap between the SP and the SC. In other words, the SPand the SC have a different, if not opposite, weight over anattribute, .

ZHENG et al.: CLOUD SERVICE NEGOTIATION IN IOT ENVIRONMENT: A MIXED APPROACH 1507

Assumption 2 (Incomplete Information): Assume that the SPand the SC are self-interested, and so always keep importantparameters secret. In other words, both parties do not disclosetheir reserved values, weights, and exact preferences, but canindicate their preference direction.

It should be observed that Assumption 1 is reasonable, sincewe only assume some gaps about agents’ preferences.Without it,there is no chance tomake a tradeoff. Assumption 2 is reasonabletoo, since we only assume incomplete information about agents’preferences, which is more realistic and practical than assumingcomplete information.

IV. UTILITY FUNCTIONS

In economics, a utility function can be used to representpreferences [2]. It is adopted here to measure the level ofsatisfaction that a user receives from a cloud service provider.In this paper, we use a general exponential function to model anagent’s preferences over a single attribute. For a lower-is-betterattribute, its utility is measured by

where represents utility, denotes the value of the attribute( ), and are positive constants, and is a scalingfactor computed such that and . Here, wechoose . It follows that and , and thus

Similarly, for a higher-is-better attribute, its utility is measuredby

where represents utility, denotes the value of the attribute( ), and are positive constants, and is a scalingfactor computed such that and . Here, wechoose . It follows that and , and thus

We now use a weighted sum function to model an agent’sutility over multiple attributes. For a proposal containing

attributes, which are linearly additive, its totalutility is measured by

where represents utility, denotes a proposal, is weight,and is the utility function of attribute . If isa lower-is-better attribute, then ; if is a higher-is-betterone, then .

In the motivating example, availability is a lower-is-betterattribute for the SP. The SP’s preferred value of availability is0.84, so its utility is determined as

Similarly, the SP’s preferred utilities of reliability, respon-siveness, security, and elasticity are determined as 0.165, 0.590,0.128, and 0.068, respectively. We assume that the SP’s weightsfor availability, reliability, responsiveness, security, and elastic-ity are 0.20, 0.30, 0.30, 0.10, and 0.10, respectively. So, the totalutility of the SP’s preferred proposal, , is determined as

Similarly, the utility of the SP’s reserved proposal isdetermined as 0.087. In the motivating example, availability isa higher-is-better attribute for the SC. Without going into detail,the utilities of the SC’s preferred proposal and reserved proposalare determined as 0.873 and 0.681, respectively.

It should be noted that since and are monotonic,we can easily map between a value of an attribute and its utility.Without this property,we cannotmake an effective concession ortradeoff. Moreover, since is additive, we can easily extendthe motivating example to handle as many attributes as needed.

Here, we make another two assumptions.Assumption 3 (Nonlinear Preference): Assume that the

utilities of attributes AVAL, REL, RESP, SECY, and ELASchange nonlinearly with their values, and (2) and (4) can be usedby the SP and the SC tomeasure the utilities of the five attributes,respectively.

Assumption 4 (Linearly-Additive Attribute): Assume thatattributes AVAL, REL, RESP, SECY, and ELAS areadditive, and (5) can be used by the SP and the SC tomeasure the utility of a proposal containing the five attributes.

We observe that Assumption 3 is reasonable, since agents’preferences in most cases are nonlinear. Assumption 4 is alsoreasonable, since agents usually, if not always, are willing to

TABLE IQOS CONFLICTS BETWEEN SP AND SC


substitute the utility of one attribute for that of another, and theirreserved values stop them going beyond bottom lines. Thisjustifies using a weighted sum function to measure the utilityof a proposal containing multiple attributes.

It should be cautioned that even (5) could cause the compen-sation problem, where a poor QoS can be counterbalanced by agood QoS, reserved values in our example can keep the negativeimpact of compensation within bounds. In fact, without a certaindegree of compensation, it is impossible to adopt a tradeoffnegotiation approach, since it utilizes the positive impact ofcompensation to exploit preference gaps. So, the compensationproblem cannot be considered as a serious drawback.

V. MULTI-ATTRIBUTE BILATERAL NEGOTIATION

Here, we introduce multi-attribute bilateral negotiations, witha focus on their negotiation protocol and negotiation strategies.In bilateral negotiations, two agents have a common interest incooperation, but have conflicting interests regarding the particu-lar way of doing so [30]. In multi-attribute negotiations, multipleissues are negotiated among agents, where a win–win solution ispossible. However, amulti-attribute negotiation ismore complexand challenging than a single-attribute one, because of complexpreferences over multiple issues and the multiple-dimensionalsolution space. For multi-attribute bilateral negotiations, whichwe deal with in the paper, their negotiation protocol and negoti-ation strategies merit special attention [4], [5], [10].

A. Negotiation Protocol

A negotiation protocol specifies the “rules of encounter”among agents [10]. In this paper, we adopt an alternating-offersprotocol for cloud service negotiation [15]. In multi-attribute bilateral negotiations, two agents alternately exchangetheir proposals and counter proposals, until one of them accepts aproposal, a failure to reach an agreement happens, or the deadlineis reached. If the first case occurs, the negotiation ends success-fully with an agreement established; otherwise, it fails andterminates with no deal made. Three points should be mentionedhere. First, a failure could happen if one cannot suggest a validproposal [30], because of incomplete information. Second,instead of time, a predefined maximum negotiation round isused to model the deadline. In fact, automated negotiation cancomplete in seconds, which makes time unsuitable to modelthe deadline in our case. Third, in each negotiation round,multiple attributes are negotiated simultaneously, which wouldbe tedious, if not impossible, for a human negotiator to do so.

B. Negotiation Strategies

Once the negotiation protocol is chosen, negotiation strategiesbecome critical. Two negotiation strategies, concession andtradeoff [13], can be used to make a proposal. When the deadlineapproaches or something undesirable happens, a party has toconcede in order to make a deal. With a concession strategy, theparty gradually reduces its utility until all conflicts are resolved.Indeed, the party who adopts the concession strategy can movetoward the preferences of its counterpart, even under incomplete

information. If no miscalculations happen, its proposal has ahigher chance of being accepted. As the concession strategydecreases the party’s utility, it is considered when no alternativesexist.

However, if two parties have different preferences, conflictscould be resolved without concession.With a tradeoff strategy, aparty yields on its less important attributes, but demandsmore onits more important attributes. As a result, a proposal moreattractive to its counterpart is created, but no utility is reduced.In particular, if it succeeds, the tradeoff strategy can generatemore utility than the concession one [31]. However, if informa-tion is incomplete, the party who adopts the tradeoff strategycould move away from the preferences of its counterpart, or inthe worst case, move in the opposite direction. So, itsproposal becomes less attractive, and it is very likely that afailure happens.

VI. MIXED NEGOTIATION APPROACH

We outline above two negotiation strategies—concession andtradeoff. With a concession strategy, an agent may receive lessutility, but has a higher chance to reach an agreement. With atradeoff strategy, the agent may get more utility, but incurs morefailures, if information is incomplete. To balance utility andsuccess rate, we propose a mixed negotiation approach for cloudservice negotiation, which is based on the “game of chicken.”

A. Two-Player Negotiation Game

In a negotiation game, a selfish agent’s utility remains thesamewith a tradeoff strategy, whereas its utility is decreasedwitha concession one. As the agent attempts to maximize its utility, itseems that it should stick to the tradeoff strategy instead of theconcession one. If the agent and its counterpart both adopt thetradeoff strategy, unfortunately, it is very likely that a failurehappens, whereupon both receive the worst utility. It thusbecomes a dilemma. This indicates that how to play concessionand tradeoff strategies is of utmost importance. However, to thebest of our knowledge, no previouswork dealswith this problem.In fact, we first identify the problem and model it with the “gameof chicken,” which goes as follows [2]. Two boys, say Alan andBob, want to prove their manhood. They drive toward each otherat breakneck speed. The one who swerves loses face andbecomes a “chicken,” whereas the other who stays, of course,proves his manhood and becomes a hero to his friends. If bothswerve, nothing is proved. If neither swerves, they crash intoeach other with potentially disastrous results.

A possible payoff matrix of the game of chicken is shown inTable II, where a number only has a relative significance,namely, the greater the number, the higher the payoff. A Nashequilibrium is “a situation in which each player in a gamechooses the strategy that yields the highest payoff, given thestrategies chosen by the other players” [2]. The “game ofchicken” has two pure strategy Nash equilibria. One is for Alanto swerve and forBob to stay,whereas the other is forAlan to stayand for Bob to swerve. In fact, if Alan swerves, Bob is better offstaying (payoff 1) than swerving (payoff 0). Conversely, if Alanstays, Bob is better off swerving (payoff −1) than staying


(payoff −10). So, those are the two pure strategy Nashequilibria. Below, we give a formal description for Nashequilibrium [6].

Definition 1 (Nash Equilibrium): A Nash equilibrium is astrategy profile , such that each player,

, has no incentive to deviate from its currentstrategy, , given the strategy profile, , of the other players,where .

A general payoff matrix of a two-player negotiation gamewithconcession and tradeoff strategies is shown in Table III, where

and > >> . It should be noted that, here, the game is asym-

metric, in that the two players are distinguishable from eachother, and is more applicable, in that it generalizes the “game ofchicken.”We establish Theorem 1 to determine its pure strategyNash equilibria. Refer to [30] for the proof.

Theorem 1 (Negotiation With Complete Information): Letbe a two-player negotiation game, whose payoff matrix is shownin Table III. There exist two pure strategy Nash equilibria for .One is for Player 1 to make a concession and for Player 2 tomakea tradeoff, whereas the other is for Player 1 tomake a tradeoff andfor Player 2 to make a concession.

In the above negotiation game, the best move of a player is todo the opposite of what its counterpart decides. So, it is importantfor the player not to have its behavior anticipated by its counter-part. In other words, its behavior should be unpredictable.A good way to achieve this is to let chance decide. In contrastto the case with pure strategies, where a player attempts tomaximize its payoff, a player, here, employs a mixed strategyto maximize its expected payoff [30]. Below, we give a formaldescription for the mixed strategy, the payoff function, and theexpected payoff concepts.

Definition 2 (Mixed Strategy):For a player , itsmixed strategyis a probability distribution over a set of

pure strategies, , where playswith probability .

Definition 3 (Payoff Function): For a player, , its payofffunction, , is a real-valued function , such that

is the payoff to when strategy is chosen,where is a set ofpure strategies.

Definition 4 (Expected Payoff): For a player , its mixedstrategy and its payoff function ’sexpected payoff of , is , where

is a pure strategy of .We establish Theorem 2 to determine the mixed strategy Nash

equilibrium of a two-player negotiation game with concessionand tradeoff strategies. Refer to [30] for the proof, and [6] forincomplete information game.

Theorem 2 (Negotiation With Incomplete Information): Letbe a two-player negotiation game, whose payoff matrix is shownin Table III. There exists a mixed strategy Nash equilibrium

for , where Players 1 and 2 play and

, respectively, and , and

.

B. Game-Theoretic Description

A mixed strategy is “a choice among two or more purestrategies according to prespecified probabilities,” where a purestrategy is a specific choice of possible strategies [2]. A mixednegotiation approachworks as follows. In preparing a proposal, aparty plays a concession strategy with a certain probability anda tradeoff strategy with another probability. In the case that aconcession strategy is played, the utility of its reference proposal,onwhich a counter proposal is based, is reduced; in the case that atradeoff strategy is played, the utility of its reference proposalremains the same. Similarly, the values of its attributes areadjusted, accordingly, in favor of its counterpart. So, the partycan encourage its counterpart to accept the proposal with a higherprobability, but at a reasonable price. In fact, the idea of mixedstrategies can be traced back to Nash’s 1950 seminal paper onEquilibrium Points in n-Person Games, where a mixed strategyis defined as probability distributions over a finite set of purestrategies [11].

A graphical representation of a mixed negotiation approachis depicted in Fig. 1. Without loss of generality, a two-dimensional space is assumed here. Also, utility functionsare assumed to be nonlinear and additive (i.e., Assumptions 3and 4). Let , , and be the indifference curves of a party’spreferred proposal, its counter proposal, and the preferredproposal of its counterpart, respectively. In economics, anindifference curve connects a set of consumption baskets thatyield the same level of utility to a consumer [2]. Again, let point

correspond to the party’s initial proposal, and point itscounter proposal. Especially, when the party randomizes itschoices of strategies, it moves along from point to point, with probability , where a tradeoff strategy is played, and so

no utility is reduced from . It then moves from point to point, with probability , where a concession strategy is

played, and thus a certain amount of utility is reduced from ,but is closer to , since < . Similarly, the partymoves toward the preferences of its counterpart, but only areasonable amount of utility is reduced from its preferred pro-posal by doing so.

TABLE IIGAME OF CHICKEN

TABLE IIITWO-PLAYER NEGOTIATION GAME


C. Algorithmic Description

Algorithm 1 implements a mixed negotiation approach. Itworks as follows. First, in line 1, agent sends ―its initialproposal―to agent , and waits for a response. If doesnot accept and ’s counter proposal is not acceptable to ,then adopts a mixed approach in thewhile loop of lines 2–15 tocreate a new proposal; otherwise, true is returned in line 16.Here, a party’s acceptance criterion is that the utility receivedfrom a proposal is no less than that of its reserved proposal,and the values received from the proposal do not go beyondits reserved values. In Section VII, we relax the criterion alittle bit.

Next, in line 4, uses function random to generate a randomnumber between 0 and 1 for variable . In lines 5–10, if < ,which implies that a concession strategy is triggered, usesfunction concession to create a new proposal, where

< . In line 6, is increased by one, eachtime the condition is triggered. It should be mentioned thatconcession is a function that implements a concession strategyof a multi-attribute negotiation. Refer to [30] for its algorithmicdescription. If , which implies that a tradeoff strategy istriggered, uses function tradeoff to create a newproposal, where

. In line 9, is increased by 1, each time thecondition is triggered.

Algorithm 1: Mixed Approach ( )

Input: array with raw values of attributes;

array with weights of attributes;

array with flags of attributes; A flag indicateswhether an attribute is higher-is-better;

parameters and < < which indicatethe rate of concession and the rate of tradeoff at a time,respectively;

parameter < < which indicates the probabilityof playing tradeoff, or for short

Output: true if succeed and false otherwise

1 agent sends to agent and waits for a response

2 while agent does not accept and ’s counter proposal is not

3 acceptable to agent

4

5 if < then

6

7

8 else

9

10

11

12 if is out of bounds then

13 return FALSE

14 else

15 agent sends to agent and waits for aresponse

16 return TRUE

It should also be mentioned that tradeoff is a function thatimplements a tradeoff strategy of a multi-attribute negotiation.Refer to [30] for its algorithmic description. In line 11, countsthe total number of negotiation rounds.

Finally, in lines 12–15, if is out of bounds, false is returned;otherwise, agent sends , whose values are adjusted, to agentas a new proposal, and waits for a response again. The processrepeats until either success or failure occurs. In this process, ’sutility of the current proposal can remain the same ( moves alongits current indifference curve) or be reduced ( moves down to itsnext indifference curve). It can be proved that Algorithm 1converges and terminates in a finite number of rounds. Referto [30] for the proof.

Agent ’smixed behavior is illustrated in Fig. 2. Suppose thatnow sits at the top-left point, where it can choose or . If ischosen, it moves, horizontally, to point (utility remains thesame), where it can choose or again. If is chosen, it moves,vertically, to point (utility is reduced), where it can chooseor again. The process repeats until itmoves to point , where itreaches ―the indifference curve of the preferred proposal ofits counterpart. As a result, moves toward the preferences of itscounterpart with a higher success rate, but only a moderatedecrease in the amount of utility. While Algorithm 1 is notalways guaranteed to find a solution, even if one exists, anysolution it finds is always guaranteed to be correct, in terms of theacceptance criterion specified earlier. In this sense, Algorithm 1is a Las Vegas algorithm―a randomized algorithm thatalways gives a correct result, i.e., it always produces correctresults or it informs of a failure [26].

It should be noted that the mixed approach we adopt innegotiation exhibits a certain degree of intelligence. Just asTuring [22] pointed out in Computing Machinery and Intelli-gence, “Intelligent behavior presumably consists in a departurefrom the completely disciplined behavior involved in computa-tion, but a rather slight one, which does not give arise to randombehavior, or to pointless repetitive loops.”

Fig. 1. Mixed negotiation approach.


VII. EVALUATION AND ANALYSIS

We conduct extensive simulations to evaluate the mixedapproach for cloud service negotiation. First, we describe theexperimental setup. Next, we describe the parameter setup.Finally, we report and analyze simulation results.

A. Experimental Setup

All simulations are conducted on a Lenovo ThinkCentredesktop with a 2.80-GHz Intel Pentium Dual-Core CPU and a2.96-GB RAM, running Microsoft Windows 7 ProfessionalOperating System. The simulations are implemented with Javaunder NetBeans IDE 7.2.1 with JDK 7u13.

An alternating-offers protocol is adopted as the negotiationprotocol, and a mixed negotiation strategy is compared withconcession and tradeoff strategies. The negotiation processworks as follows. First, without loss of generality, a SP sendsits initial proposal to a SC. Next, if the proposal is accepted by theSC, negotiation ends successfully; otherwise, the SC uses eithermixed, tradeoff, or concession negotiation approach to create acounter proposal. After that, the SC sends back the counterproposal to the SP, and the negotiation process repeats. Theprocess ends once a proposal or a counter proposal is accepted,and it fails if no proposal is acceptable to both parties.

Java multithreading, which allows multiple tasks in a programto be executed concurrently, is the ideal technique to simulate thenegotiation process. A thread is the flow of execution, frombeginning to end, of a task.Wemodel the behaviors of the SP andthe SC as two threads. In particular, we use thread synchroniza-tion techniques to coordinate their behaviors, and a shared objectto exchange their proposals and counter proposals.

In our software prototype, there is a QoS matrix to benegotiated, where the SP and the SC can specify their QoSrequirements, i.e., their preferred values, reserved values, andweights over quality dimensions AVAL, REL, RESP, SECY,and ELAS. In a real negotiation, those values would be keptprivate. In our simulations, we attempt to resolveQoS conflicts inthe motivating example. In other words, we use Table I as theQoS matrix to be negotiated. Also, there is a parameter variance

< < that can be used to generate a random numberwithin a certain interval of a value, such that the impact of aspecific data set on negotiation results can be reduced, if notcompletely removed.

As to negotiation strategies, the SP and the SC can choose aconcession, a tradeoff, or a mixed approach. So, in total, thereexist nine combinations, i.e., CC, CT, CM, TC, TT, TM,MC, MT, MM, respectively, where stands for a concessionapproach, a tradeoff one, and a mixed one. Also, there areparameters the rate of concession < < , the rate oftradeoff < < , and the probability of playing tradeoff

< < (or for short).As to negotiation results, success or failure can happen. In the

case that success occurs, QoS conflicts are resolved, and the newvalues agreed to by the SP and the SC are output. In the case thatfailure happens, relevant information about the failure is output.Also, there is parameter tolerance < < , within which asolution whose values go beyond a party’s reserved values, butits utility is no less than its reserved utility is still acceptable. So, arigid cutoff value is avoided, and the chance of success increased.In fact, it is the acceptance criterion that we adopt in oursimulations.

To fully understand the impact of different parameters onnegotiation results, we conduct a series of simulations wherethe SP and the SC both adopt a concession, a tradeoff, and amixed approach, respectively. Refer to [30] for those parts. Wekeep, without further details,

, in our simulations. Unless specified other-wise, we keep the gap in preferred values as 0.11, 0.20, 0.60,0.05, and 0.06, and the gap in reserved values as 0.19, 0.10, 0.20,0.15, and 0.14 for AVAL, REL, RESP, SECY, and ELAS,respectively, in our simulations. Also, we set the maximumnegotiation round , as most negotiations can finish in nomore than 20 rounds.

It should be mentioned that Theorem 2 gives some idea abouthow to set up , but it assumes that agents’ preferences are known.In our simulations, agents’ preferences are kept private so wecannot apply Theorem 2 here. However, it does give us some hintson how to choose . As a general rule, if competition is high, asmall value is preferred; otherwise, a large value is preferred.

B. Monte Carlo Simulations

First, we study which strategy performs better when a party’scounterpart plays a concession strategy. Without loss of gener-ality, we let the SP adopt a concession approach, whereas the SCcan choose a concession, a tradeoff, or a mixed approach. Theresults are shown in Table IV. Take CC of run 1 as an example.The negotiation ends successfully at round 8 when the SCaccepts a proposal from the SP, where the agreed values forAVAL, REL, RESP, SECY, and ELAS are 0.873, 0.820, 0.464,0.858, and 0.926, respectively. The SP’s received utility is 0.178,which is greater than its reserved utility 0.087, but less than itspreferred utility 0.244. In contrast, the SC’s received utility is0.777,which is also greater than its reserved utility 0.688, but lessthan its preferred utility 0.879. It can be verified that the acceptedproposal is a valid solution, according to the acceptance criterionspecified earlier.

In our simulations, the average negotiation round is 8.90 forCC, and its success rate is 100% (0 failures out of 10 runs). Theaverage utility is 0.158 and 0.722 for SP and SC, respectively,and the total utility is 0.880. In our simulations, the average

Fig. 2. Agent ’s mixed behavior.


negotiation round is 8.60 for CT, and its success rate is 70%(3 failures out of 10 runs). The average utility is 0.109 and 0.562for SP and SC, respectively, and the total utility is 0.671. In oursimulations, the average negotiation round is 8.80 for CM, and itssuccess rate is 90% (1 failure out of 10 runs). The average utilityis 0.133 and 0.686 for SP and SC, respectively, and the totalutility is 0.819. Here, CC outperforms CT and CM in terms ofboth utility and success rate. It seems that the result does notagree with the theory, which states that the best strategy tocounter a concession strategy is tradeoff. In fact, the theory onlyworks with complete information, but according to Assumption2, information is incomplete in our simulations. Even if a partyknows the strategy of its counterpart, it does not know its privatevalues. So, CCperforms better thanCT andCM―CCdoes notincur a failure, but CT and CM do.

Second, we study which strategy performs better when aparty’s counterpart plays a tradeoff strategy. Without loss ofgenerality, we let the SP adopt a tradeoff approach, whereas theSC can choose a concession, a tradeoff, or amixed approach. Theresults are shown in Table V. In our simulations, the averagenegotiation round is 9.20 for TC, and its success rate is 100%(0 failures out of 10 runs). The average utility is 0.147 and 0.706for SP and SC, respectively, and the total utility is 0.852. In oursimulations, the average negotiation round is 10.60 for TT, andits success rate is 20% (8 failures out of 10 runs). The averageutility is 0.020 and 0.174 for SP and SC, respectively, and thetotal utility is 0.195. In our simulations, the average negotiationround is 9.20 for TM, and its success rate is 80% (2 failures out of10 runs). The average utility is 0.101 and 0.617 for SP and SC,respectively, and the total utility is 0.718. Here, TC outperformsTT and TM in terms of both utility and success rate. It should benoted that the result agrees with the theory, which states that thebest strategy to counter a tradeoff strategy is concession. In fact,the theory works with both complete and incomplete informa-tion, as in our simulations. So, TC performs better than TT andTM―TC does not incur a failure, but TT and TM do.

Third, we study which strategy performs better when a party’scounterpart plays a mixed strategy. Without loss of generality,we let the SP adopt amixed approach,whereas the SC can choosea concession, a tradeoff, or a mixed approach. The results areshown in Table VI. In our simulations, the average negotiationround is 9.20 for MC, and its success rate is 100% (0 failures outof 10 runs). The average utility is 0.147 and 0.706 for SP and SC,respectively, and the total utility is 0.852. In our simulations, theaverage negotiation round is 10.60 for MT, and its success rate is20% (8 failures out of 10 runs). The average utility is 0.020 and0.174 for SP and SC, respectively, and the total utility is 0.195. Inour simulations, the average negotiation round is 9.60 for MM,and its success rate is 100% (0 failures out of 10 runs). Theaverage utility is 0.133 and 0.752 for SP and SC, respectively,

TABLE VNEGOTIATION RESULTS (TC, TT, AND TM)

indicates negotiation round, success, and failure.

TABLE VINEGOTIATION RESULTS (MC, MT, AND MM)


TABLE IVNEGOTIATION RESULTS (CC, CT, AND CM)



and the total utility is 0.885. Here,MM outperformsMC in termsof utility, with one exception (the SP’s utility of MM is less thanthat of MC), and it outperforms MT in terms of both utility andsuccess rate. It should be noted that the result agrees with thetheory, which states that the best strategy to counter a mixedstrategy is a mixed one. In fact, the theoryworks with incompleteinformation, as in our simulations. So, MM performs better thanMC and MT―MM and MC do not incur a failure, but MTdoes.

In short, the game of chicken has two pure strategy Nashequilibria, and one mixed strategy Nash equilibrium, where theformer works with complete information, and the latter incom-plete information. In real negotiations, however, information isnot complete—an agent may not know the payoffs and strategiesof its counterpart—in most cases, if not all. So, a mixed approachbased on the “game of chicken” becomes a promising approachfor cloud service negotiation. In fact, as demonstrated here, amixed approach, which can balance utility and success rate,achieves a higher utility than a concession approach, whileincurring fewer failures than a tradeoff one.

VIII. CONCLUSION

IoT and cloud computing complement each other. IoT canbenefit from the unlimited capabilities and resources of cloudcomputing. Also, when coupled with IoT, cloud computing canin turn deal with real world things in a more distributed anddynamic manner.

To succeed in a competitive market, cloud providers need tooffer superior services that meet customers’ expectations. How-ever, cloud providers and cloud consumers have different andsometimes oppositeQoSpreferences. If such a conflict occurs, anagreement cannot be reached, without negotiation.

A tradeoff approach can outperform a concession one in termsof utility, but may incur more failures if information is incom-plete. To balance utility and success rate, we propose a mixedapproach for cloud service negotiation, which is based on the“game of chicken.” In particular, if a party is uncertain about thestrategy of its counterpart, it is best to mix concession andtradeoff strategies. In fact, it is amixed strategyNash equilibriumof a negotiation game with two pure strategies, which providesthe theoretical basis for our approach.

To demonstrate the effectiveness of the mixed approach, weconduct extensive simulations. Results show that when a partyhas no knowledge of the strategy of its counterpart, a mixedapproach outperforms a concession one in terms of utility, and itoutperforms a tradeoff one in terms of success rate. It should benoted that the mixed approach works under incomplete informa-tion, and so is applicable for real negotiations, where informationis generally not complete.

This work, however, has one potential limitation. The nonlin-ear utility functions we adopt in our simulations are general ones,whichmay not be able to represent a specific agent’s preferences.However, it is not realistic to have an accurate utility function atthis time, since we do not know how and at what cost to engineerQoS requirements. Even with the limitation, the main results andconclusions of the paper are not affected.

In conclusion, when one is uncertain about the strategy of itscounterpart, a mixed negotiation approach, which exhibits acertain degree of intelligence, can achieve a higher utility than aconcession one, while incurring fewer failures than a tradeoffone. It thus becomes a promising approach for cloud servicenegotiation.

REFERENCES

[1] M. Armbrust et al., “A view of cloud computing,” Commun. ACM, vol. 40,no. 4, pp. 50–58, 2010.

[2] D. Besanko and R. R. Braeutigam,Microeconomics, 3rd ed. Hoboken, NJ,USA: Wiley, 2008.

[3] Q. Duan, Y. Yan, and A. V. Vasilakos, “A survey on serivce-orientednetwork virtualizaiton toward convergence of networking and cloud com-puting,” IEEE Trans. Netw. Service Manag., vol. 9, no. 4, pp. 373–392,Dec. 2012.

[4] P. Faratin, C. Sierra, and N. Jennings, “Negotiation decision functions forautonomous agents,” Robot. Auton. Syst., vol. 24, no. 3-4, pp. 159–182,1997.

[5] N. R. Jennings et al., “Automated negotiation: Prospects, methods andchallenges,” Group Decis. Negotiation, vol. 10, no. 2, pp. 199–215, 2001.

[6] K. Leyton-Brown and Y. Shoham, Essentials of Game Theory: A Concise,Multidisciplinary Introduction. San Rafael, CA, USA: Morgan &Claypool, 2008.

[7] Q. Li et al., “Applications integration in a hybrid cloud computing envi-ronment: Modelling and platform,” Enterpr. Inf. Syst., vol. 7, no. 3,pp. 237–271, 2013.

[8] S. Li et al., “Integration of hybrid wireless networks in cloud servicesoriented enterprise information systems,” Enterpr. Inf. Syst., vol. 6, no. 2,pp. 165–187, 2012.

[9] S. Li, L. Xu, andX.Wang, “Compressed sensing signal and data acquisitionin wireless sensor networks and Internet of things,” IEEE Trans. Ind.Informat., vol. 9, no. 4, pp. 2177–2186, Nov. 2013.

[10] A. R. Lomuscio, M. Wooldridge, and N. R. Jennings, “A classificationscheme for negotiation in electronic commerce,”GroupDecis. Negotiation,vol. 12, no. 1, pp. 31–56, 2003.

[11] J. F. Nash, “Equilibrium points in n-person games,” in Proc. Natl. Acad.Sci., vol. 36, 1950, pp. 48–49.

[12] D. Paulraj, S. Swamynathan, and M. Madhaiyan, “Process model-basedatomic service discovery and composition of composite semantic webservices using web ontology language for services (OWL-S),” Enterpr.Inf. Syst., vol. 6, no. 4, pp. 445–471, 2012.

[13] H. Raiffa, The Art and Science of Negotiation. Cambridge, MA, USA:Harvard Univ. Press, 1982, pp. 148–165.

[14] L. Ren et al., “A methodology towards virtualisation-based high perfor-mance simulation platform supporting multidisciplinary design of complexproducts,” Enterpr. Inf. Syst., vol. 6, no. 3, pp. 267–290, 2012.

[15] A. Rubinstein, “Perfect equilibrium in a bargaining model,” Econometrica,vol. 50, no. 1, pp. 97–110, 1982.

[16] K. M. Sim, “Agent-based cloud computing,” IEEE Trans. Serv. Comput.,vol. 5, no. 4, pp. 564–577, Nov. 2012.

[17] V. Stantchev and C. Schröpfer, “Negotiating and enforcing QoS and SLAsin grid and cloud computing,” in Proc. 4th Int. Conf. Grid PervasiveComput., LNCS 5529. Geneva, Switzerland, 2009, pp. 25–35.

[18] F. Tao et al., “Research on manufacturing grid resource service optimal-selection and composition framework,” Enterpr. Inf. Syst., vol. 6, no. 2,pp. 237–264, 2012.

[19] F. Tao et al., “Modelling of combinable relationship-based compositionservice network and the theoretical proof of its scale-free characteristics,”Enterpr. Inf. Syst., vol. 6, no. 4,, pp. 373–404, 2012.

[20] F. Tao et al., “FC-PACO-RM: A parallel method for service compositionoptimal-selection in cloud manufacturing system,” IEEE Trans. Ind.Informat., vol. 9, no. 4, pp. 2023–2033, Nov. 2013.

[21] J. M. Tien, “Services: A system’s perspective,” IEEE Syst. J., vol. 2, no. 1,pp. 146–157, Mar. 2008.

[22] A. M. Turing, “Computing machinery and intelligence,” Mind, vol. 59,no. 236, pp. 2023–2033, 1950.

[23] L. M. Vaquero et al., “A break in the clouds: Towards a cloud definition,”ACM SIGCOMM Comput. Commun. Rev., vol. 39, no. 1, pp. 50–55,2009.

[24] H.Wang,W. He, and F. K.Wang, “Enterprise cloud service architectures,”Inf. Technol. Manag., vol. 13, no. 4, pp. 445–454, 2012.


[25] J. W. Wang et al., “On a unified definition of the service systems: What isits identity,” IEEE Syst. J., to be published, doi: 10.1109/JSYST.2013.2260623.

[26] X. Wang, Design and Analysis of Computer Algorithms. Beijing,China: Publishing House of Electronics Industry, pp. 216–222, 2001(in Chinese).

[27] L. Xu, “Guest editorial: Advances of systems research in service industry,”IEEE Syst. J., to be published, doi: 10.1109/JSYST.2013.2279921.

[28] L. Xu, “Introduction: Systems science in industrial sectors,” Syst. Res.Behav. Sci., vol. 30, no. 3, pp. 211–213, 2013.

[29] E. Yaqub et al., “A generic platform for conducting SLA negotiations,”in Service Level Agreements for Cloud Computing, P. Wieder, et al., Eds.New York, NY, USA: Springer, 2011, pp. 187–206.

[30] X. Zheng, “QoS Representation, Negotiation and Assurance in CloudServices,” Ph.D dissertation, School of Computing Queen’s Univ.,Kingston, ON, Canada, Feb. 2014.

[31] X. Zheng, P. Martin, and K. Brohman, “Cloud service negotiation: Con-cession vs. tradeoff approaches,” in Proc. 12th IEEE/ACM Int. Symp.Cluster, Cloud Grid Comput. (CCGrid 2012). Ottawa, ON, Canada, 2012,pp. 515–522.

[32] X. Zheng, P. Martin, and K. Brohman, “Cloud service negotiation: Aresearch roadmap,” in Proc. 10th IEEE Int. Conf. Services Comput., SantaClara, CA, USA, 2013, pp. 627–634.

[33] J. Zhou et al., “CloudThings: A common architecture for integrating theInternet of things with cloud computing,” in Proc. 17th IEEE Int. Conf.Comput. SupportedCooperativeWorkDesign,Whistler,BC,Canada, 2013,pp. 651–657.

Xianrong Zheng (S’13), photograph and biography not available at the time ofpublication.

Patrick Martin, photograph and biography not available at the time ofpublication.

Kathryn Brohman, photograph and biography not available at the time ofpublication.

Li Da Xu (M’86–SM’11), photograph and biography not available at the time ofpublication.
