Method of Resource Estimation Based on QoS in...

Research ArticleMethod of Resource Estimation Based on QoS inEdge Computing

Guangshun Li Jianrong Song Junhua Wu and Jiping Wang

School of Information Science and Engineering Qufu Normal University Rizhao 276800 China

Correspondence should be addressed to Guangshun Li 30752585qqcom

Received 13 October 2017 Accepted 21 December 2017 Published 22 January 2018

Academic Editor Shangguang Wang

Copyright copy 2018 Guangshun Li et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

With the development of Internet ofThings the number of network devices is increasing and the cloud data center load increasessome delay-sensitive services cannot be responded to timely which results in a decreased quality of service (QoS) In this paper wepropose a method of resource estimation based on QoS in edge computing to solve this problem Firstly the resources are classifiedand matched according to the weighted Euclidean distance similarityThe penalty factor and Grey incidence matrix are introducedto correct the similarity matching functionThen we use regression-Markov chain prediction method to analyze the change of theload state of the candidate resources and select the suitable resource Finally we analyze the precision and recall of the matchingmethod through simulation experiment validate the effectiveness of thematchingmethod and prove that regression-Markov chainprediction method can improve the prediction accuracy

1 Introduction

With the development of Internet of Things (IoT) [1] moreandmore devices especiallymobile devices constantly accessthe Internet CISCO predicts that 50 billion devices will con-nect to the Internet by 2020 [2] These devices will generatelarge amounts of data at the end of the network which leadsto the increment of the burden of cloud data centerMoreoverthe remote distance between the mobile devices and clouddata center makes the transmission delay increase whichmakes some delay-sensitive services can not get response andprocessing rapidly IoT services such as connected vehicleand video streaming require high bandwidth and low latencycontent delivery to guarantee QoS Edge computing [3] playsan important role by using network resources near the localnetwork to provide a low latency service and improve QoS

Edge computing is located between the end devices andcloud data center It transfers data and computing power fromthe cloud to the edge of the network which processes datalocally [4] Due to the fact that edge servers are close to endusers the latency between mobile devices and data center isreducedMoreover the resources are close to the user that canimprove QoS to some extent

In the IoT environment the number of network devicesis increasing constantly The resources with the same servicefunction are also increasing gradually The complexity ofuser requirements makes resource management a challengeAmong the many available resources selecting the suitableresources to meet the needs of users has become the maingoal

Although the latency is reduced the resources in edgecomputing are limited compared to cloud computing Theavailability and short-term prediction of resources load [56] have some influence on task scheduling applicationperformance and the QoS of users The method of resourceprediction can provide the appropriate resource for usersby analyzing the load of the resource itself Therefore theestimation of Resource QoS is significant to user satisfactionand task allocation in edge computing

In this paper we first propose an edge computingframework Then we introduce the penalty factor and Greyincidencematrix to improve the accuracy of similaritymatch-ing and select resources which satisfy the needs of usersWe use the regression-Markov chain prediction method forcandidate resources to analyze the change of the load state of

HindawiWireless Communications and Mobile ComputingVolume 2018 Article ID 7308913 9 pageshttpsdoiorg10115520187308913

2 Wireless Communications and Mobile Computing

the resource itself Finally we prove the effectiveness of theestimation method through simulation experiment

The remainder of this paper is organized as follows Therelated works are introduced in Section 2 We describe thearchitecture of the edge computing in Section 3 In Section 4we describe the method of dynamic resource estimationThe experiment results are presented in Section 5 Section 6concludes the paper

2 Related Work

Edge computing is the extension of cloud computing tonetwork edge There is no standard architecture The workof resource estimation has not been well addressed Mao etal [7] developed a joint radio and computational resourcemanagement algorithm for multiuser MEC systems with theobjective of minimizing the average weighted sum powerconsumption of the mobile devices and the MEC serverShekhar and Gokhale [8] proposed Dynamic Data DrivenCloud and Edge Systems (D3CES) as a framework for adap-tive resource management across cloud and edge resourcesto provide QoS guarantees for performance-sensitive appli-cations Wang et al [9] developed an Edge NOde ResourceManagement (ENORM) framework to manage edge nodesThey proposed amechanism for provisioning and autoscalingedge node resources Their method reduced data transmis-sion and communication frequency between the edge nodeand cloud

Aazam and Huh [10] proposed a dynamic resourceestimation and pricing model for IoTTheir work was mainlyfocused on considering different types of customers anddevices and the service-oriented relinquish probabilitiesBased on previous studies Aazam et al [11] proposed aQoE-based resource estimation model to enhance QoSTheydevised MeFoRE method on the basis of service relinquishprobability and historical records And they considered pre-vious QoE and Net Promoter Score records to enhance QoSZuo et al [12] proposed a resource estimationmodel based onentropy optimization and dynamicweightedmethod to accu-rately grasp the dynamic load and availability information ofresources They used the objective function and constraintcondition of maximum entropy to select the resources whichmeet the QoS of userThismethod achieved optimal schedul-ing and guaranteed the QoS of user Dynamic weighted loadevaluation was carried out on the filtered resources to realizeload balancing and improve system utilization

Zhao et al [13] proposed a multidimensional resourcemodel for dynamic resource matching in the IoT with themultidimensional description of the IoT resources Theyconsidered the ontology structure and ontology descriptionto calculate the similarity by matching the hardware featuresand software feature between resources Zhou et al [14]proposed a multiple QoS resource selection and estimationalgorithm They used Multiple Attribute Decision Makingand Analytic Hierarchy Process (AHP) algorithm Theirmethod mainly considered the preference information ofusers Dong et al [15] analyzed the QoS of a single resourcenode They used a local optimization algorithm to select the

appropriate cloud resources that meet the needs of usersDing et al [16] proposed QoS-aware resource matchingand recommendation method in cloud computing Theydescribed a resource matching algorithm that considers bothfunctional requirements and QoS attributes and designed aresource recommendation method for cloud computing

3 Edge Computing

Edge computing located between themobile end devices andcloud data center provides computing storage and networkservices Edge computing has the ability of decentralizedcomputation and storage compared with cloud computingThe resources near the user can support mobile devices forreal-time communication The main goal of edge computingis to preprocess data reduce the delay and serve the appli-cations of low delay and real-time response Currently thecommon edge computing architecture is a three-tier networkarchitecture as shown in Figure 1

As the underlying node the end devices not only con-sume data but also produce data In particular mobiledevices will require more resource from edge servers ratherthan cloud data center The end devices include all IoTdevices such as smart mobile phone intelligent vehicle andvirtual sensor nodes These devices are able to sense datacommunicate with the upper layer through sensor networks3G WiFi and so on and transmit the collected raw dataCompared with the servers in data centers most of the IoTdevices or sensors at the edge are somewhat limited in bothcomputing power and battery capacity [17] Edge devicesinclude some traditional network devices (routers switchesetc) and some specially deployed devices (local servers)Edge devices have the ability of computing processing stor-age and forwarding data to the cloud servers Edge serverscan be connected directly to nearby mobile devices via one-hopwireless linkMicro data center (MDC) resources includecomputing resources network resources storage resourcesand software resources On the one hand resources comefrom the cached local resources on the other hand they comefrom the data obtained by the monitoring equipment As thetop layer cloud computing layer includes data center (DC)and some cluster servers The cloud servers can receive datasent by edge devices for processing and storage

Figure 2 shows the request process of the service in theedge computing The user first submits the request to thesystem administrator through the end devices The systemadministrator stores the submitted information and transmitsthe statistical QoS requirements to the edge servers throughthe edge gateway Monitoring equipment generates statisticslog by tracking resource utilization and availability of sensorapplications and servicesThemonitoring data is transmittedto the edge servers The edge servers analyze the QoS ofrequested service for end users and process locally Theservice is divided into several tasks which are processedlocally as far as possible Each task selects the best matchingresource from the resource pool based on the estimationresultsThen the selected resource is allocated and scheduledto meet the QoS requirements

Wireless Communications and Mobile Computing 3

DC

Resource

MDC

Storage

Computing Network

3G4G WiFi

End device

Edge server

Cloud server

Software

Figure 1 Edge computing architecture

Systemadministrator

Statistical QoSrequirements

Edgegateway

Cloudgateway

Monitor

Cloud data center

Edge server

Task processingTask 1

Resource pool

Task 2

Serviceresponse

Request service

End user

QoSrequirements

Serviceresponse Service

response

Service response

Serviceresponse

Service response Local computing

Cloudcomputing

QoS

requirements

Monitoring data

Task n

Figure 2 The request process of service in edge computing

4 Dynamic Estimation Method

In this section we describe the dynamic estimation methodand estimate the resource with the same service functionThespecific process is shown in Figure 3

As shown in the figure the dynamic estimation methodincludes two phases similarity matching algorithm andregression-Markov chain prediction method First we estab-lish a Resource QoS matrix to calculate the matching degreebetween QoS requirements of users and resources throughsimilarity matching Then resources which satisfy the needsof the users are selected by threshold Finally we analyze

the change of resource load and select the optimal resourcethrough regression-Markov chain prediction method

41 Similarity Matching Method

411 Resource Description The resources in the edge nodescan be provided to meet the QoS requirements of users Inthis paper resource set which have the same service functionis defined as follows119877119904 = 1199031 1199032 119903119899 (1)


Start

Available resource set Rs

Resource matrix R

Resource subset RA Resource subset RB

Resource matchingmatrix PA

Resource matchingmatrix PB

RA resource put into

Regression-Markov chainprediction

End

Yes

No

Yes

No

Distance similaritＳ gt

Matching degreesiＧ gt k

resource set Rs

Figure 3 Flow chart of resource estimation method

Each resource has different QoS attributes According tothese attributes the resource is evaluated to determine thematching between the needs of users and resources Resourceset 119877119902119894 denotes a collection of the QoS attributes of resource119877119902119894 = 1199031199021 1199031199022 119903119902119898 (2)

where index 119898 represents that each resource has 119898 differentQoS attributes including response time availability reliabil-ity and price The price attribute in this paper is defined asfollows 119901 = 119880 lowast 119887120583120593 lowast 119863dev (3)

where 119880 is the basic price of service 120583 is the number ofservice requests completed in unit time 120593 is the number ofservice requests received in unit time 119887 is a price adjustmentfactor determined by the service provider 119863dev representsthe device type which can generally be divided into staticdevices small mobile devices and large mobile devices Therelative reserved resources of each device are 1 125 and 15times respectively [10]

The QoS attributes requested by the users are the same asthe resources The QoS attributes set is defined as follows119906119902 = 1199061199021 1199061199022 119906119902119898 (4)

In this paper we construct a QoS attribute matrix ofresources as the decision matrix

119877 = (119903119894119895)119899times119898 =(11990311 11990312 sdot sdot sdot 119903111989811990321 11990322 sdot sdot sdot 1199032119898 1199031198991 1199031198992 sdot sdot sdot 119903119899119898) (5)

where 119903119894119895 is the 119895th QoS attribute value of 119894th resourceSince the measurement units of the QoS attributes are

different it is meaningless to process the matrix directlyTherefore according to the relationship between attributesand user satisfaction we use the following formula to stan-dardize processing

119911119894119895 = 119903119894119895 minus 119903min

119895119903max119895 minus 119903min

119895

119902 is positive attribute119903max119895 minus 119903119894119895119903max119895 minus 119903min

119895

119902 is negative attribute(6)

The above formula indicates two cases if QoS attribute 119902is positive attribute the greater the value of the attribute isthe higher the satisfaction of the users is On the contrarythe negative attribute indicates that the smaller the attributevalue is the higher the user satisfaction is

In order to ensure the objectivity of the estimation resultswe use the entropy weight method to calculate entropy value119890119895 and entropy weight 119908119895 for the QoS attribute of resourcesThe formula is as follows119890119895 = minus 1ln 119899 119899sum

119894=1

119891119894119895 sdot ln119891119894119895119908119895 = 1 minus 119890119895119898 minus sum119898119895=1 119890119895119891119894119895 = 119911119894119895sum119899119894=1 119911119894119895119898sum119895=1

119908119895 = 1(7)

412 Similarity Matching Algorithm In order to reduce thematching time and improve the matching efficiency theresources are firstly classified before similarity matchingBased on theQoS attributes value requested by the users eachresource can be regarded as a point in the multidimensionalspace The distance between the resources and the needs ofusers is measured by Euclidean distance Since each user mayhave preferences for one attribute or each attribute of theresources affects the result of the measurement differently


therefore we set objective weight for each attribute and useweighted method to calculate

119889 (119894 119906119902) = radic 119898sum119895=1

119908119895 lowast (119903119902119895 minus 119906119902119895)2119889sim (119894 119906119902) = 11 + 119889 (119894 119906119902) (8)

where 119908119895 represents the weight of resource attribute 119889(119894 119906119902)indicates the distance between the 119894th resource node and theideal node (the node indicates theQoS requirements of users)in the space 119889sim(119894 119906119902) is the degree of proximity whoserange is from 0 to 1 The resource is close to the ideal nodewhen 119889(119894 119906119902) is smaller

We set the proximity threshold 120576 by computing theproximity degree between the resource node and the idealnode The range of threshold is from 0 to 1 It divides theresources into two sets1198771015840 (120576) = 119903119896 | 119889sim (119896 119906119902) ge 120576 (9)

On the basis of the resource classification we establish thematching matrix between the requested QoS of users and theQoS of the candidate resources

119875 =( 11990111 11990112 sdot sdot sdot 119901111989811990121 11990122 sdot sdot sdot 1199012119898 119901(119899+1)1 119901(119899+1)2 sdot sdot sdot 119901(119899+1)119898) (10)

where the first 119899 row represents theQoS attributes value of the119899 resources The 119899 + 1 row represents the QoS attribute valueof the expectation of usersWe use formula (6) to standardizethe matrix and calculate the matching degree between theexpectation resource of users and the resources

In this paper we use Pearson correlation coefficient tocalculate the matching degree of resources

sim (119909 119910) = Σ119903isin119877 (119902119903119909 minus 119902119909) (119902119903119910 minus 119902119910)radicΣ119903isin119877 (119902119903119909 minus 119902119909)2radicΣ119903isin119877 (119902119903119910 minus 119902119910)2 (11)

where sim(119909 119910) is the similarity between attributes 119909 and119910 119902119903119909 and 119902119903119910 represent the corresponding value of theQoS attributes 119909 and 119910 of all resources respectively 119902119909 and119902119910 represent the average value of attributes 119909 and 119910 of allresources respectively

In order to ensure that the QoS attribute value ofcandidate resources is within the expected range of userswe introduce the penalty factor and Grey incidence matrixto analyze the gap between the resource attribute valueThe penalty factor 120582 is set on the condition of resourceconstraint The smaller the penalty factor is the closer theuser expectation range is and the higher the correlation isOn the contrary when the penalty factor is larger the distance

is more away from the expectation range of users the lowerthe correlation is and the lower the matching degree is

120582119895 = 0 119903119895 lt 120575119895119903119895 minus 120575119895120574119895 minus 120575119895 120575119895 le 119903119895 le 1205741198951 119903119895 gt 120574119895 (12)

where 120575119895 and 120574119895 represent correspondingminimumvalue andmaximum value of the expectation range of users for eachattribute of the resources respectively

We introduce the penalty factor into the Grey relationalcoefficient to calculate correlation degree and correct match-ing functionThe attribute correlation coefficient is calculatedas follows 120585119894119895 = min + 120578 sdotmax120582119895 1003816100381610038161003816119901119899+1 (119895) minus 119901119894 (119895)1003816100381610038161003816 + 120578 sdotmax

min = minmin 1003816100381610038161003816119901119899+1 (119895) minus 119901119894 (119895)1003816100381610038161003816max = maxmax 1003816100381610038161003816119901119899+1 (119895) minus 119901119894 (119895)1003816100381610038161003816 (13)

The Grey incidence matrix is defined as follows

120585 = [[[[[[[12058511 12058512 sdot sdot sdot 120585111989812058521 12058522 sdot sdot sdot 1205852119898 1205851198991 1205851198992 sdot sdot sdot 120585119899119898

]]]]]]] (14)

120585119894119895 is correlation coefficient of the 119895th attribute of the 119894thresource 120578 is the distinguishing coefficient with a generalvalue of 05 120582119895 is a penalty factor of the 119895th attribute min andmax represent the maximum difference and the minimumdifference 119901119894(119895) and 119901119899+1(119895) are standardized values

The modified matching function is

sim (119903119894 119906119902) = 120572 lowast sim (119903119902119894 119906119902119894) + (1 minus 120572) lowast 1119898 119898sum119895=1

120585119894119895 (15)

where 120572 represents weight When the matching degreereaches the threshold 119896 that is sim(119903119894 119906119902) ge 119896 this indicatesa successful match The resources that successfully match areput into a candidate resource set 1198771199041015840

The specific matching algorithm is illustrated inAlgorithm 1

42 Regression-Markov Chain Prediction EstimationThrough the above analysis we select a candidate resourceset which satisfy the QoS requirements of users Becauseof the dynamic and uncertain nature of the resources theamount and value of data collected of QoS will fluctuateMoreover the inherent mobility of the IoT makes mobileusers have certain volatility when using resources Thereforewe further select resources by analyzing the load changes ofresources themselves


Input 119877119904 = 1199031 1199032 119903119899 119877 119906119902Output 1198771199041015840(1) Normalized matrix 119877(2) [119899119898] = size(119877)(3) for 119895 = 1 rarr 119898 do(4) get the 119908119895(5) end for(6) 119876 = [119877 119906119902](7) [row col] = size(119876)(8) 119877119860[] larr 0(9) 119877119861[] larr 0(10) 119906 larr 1(11) Vlarr 1(12) for 119894 = 1 rarr row-1 do(13) get the 119889(119894 119906119902) and 119889sim(119894 119906119902)(14) if 119889sim(119894 119906119902) ge 120576 then(15) 119877119860[119906] larr 119903119894(16) 119906 larr 119906 + 1(17) else(18) 119877119861[V] larr 119903119894(19) Vlarr V + 1(20) end if(21) end for(22) 1198771199041015840[] larr 0(23) 119905 larr 1(24) while resource 119903119894 isin 119877119860 do(25) get the matrix 119875119860(26) calculate the 120582119895 and 120585119894119895(27) calculate the sim(119903119894 119906119902)(28) if sim(119903119894 119906119902) ge 119896 then(29) 1198771199041015840[] larr 119903119894(30) 119905 larr 119905 + 1(31) end if(32) end while(33) return 1198771199041015840

Algorithm 1 Multiattribute QoS resource matching

Due to the randomness and volatility of the load thesingle resource predictionmethod is difficult to predict accu-rately The load has the characteristics of frequent changesin the short term [18] The future load will be affectedby the current load and the load at the next time canbe predicted by current load value Therefore this paperadopts the regression-Markov chain prediction method tobetter reflect the changing trend and stochastic fluctuationcharacteristics of resources

According to the difference of time 119905 the original datasequence of the resource load value is recorded as119883(0) (119905) = 119883(0) (1) 119883(0) (2) 119883(0) (119901) (16)

Firstly we use the linear regression method to generatethe predicted sequence119883(0) (119905) = 119883(0) (1) 119883(0) (2) 119883(0) (119901) (17)

and calculate the relative residuals of the original sequenceand the predicted sequence119883(1) (119905) = 119883(0) (119905) minus 119883(0) (119905) (18)

Then we divide the relative residual sequence 119883(1)(119905)into 119904 state intervals According to the state distribution119878 = (1 2 119904) a one-step transition probability matrix iscalculated119872119904times119904 = 119872(119904119894) (119904119895) = 119901119894119895 (119883119904+1 = 119895 | 119883119904 = 119894)

119872 = [[[[[[[11987211 11987212 sdot sdot sdot 119872111990411987221 11987222 sdot sdot sdot 1198722119904 d

1198721199041 1198721199042 sdot sdot sdot 119872119904119904]]]]]]]

(19)

If matrix 119872 satisfies the following conditions 119872119894119895 ge 0119894 119895 isin 119878 and sum119872119894119895 = 1 119894 119895 isin 119878 119872 is a random matrix119875(119899) = (119901119894119895(119899)) = 119875119899 represents 119899 step transfer matrixIt is assumed that the Markov chain reaches a stable

state after the 119899 step transition The stable state vector 119909 =[1199091 1199092 119909119904] satisfies 119909 = 119909119872119904sum119894=1

119909119894 = 1 (20)

We set the initial state to obtain the probabilities1199011 1199012 119901119904 corresponding to the error state interval bysolving the distribution probability of the stable state Accord-ing to the maximum probability principle the state corre-sponding to the maximum probability is taken as the stateof the next moment The predictive value is brought into thecorresponding state interval to solve the prediction intervalvalue On this basis we predict the availability of resourcesover a period of time and rationally select the right resource

5 Experimental Results

To test the performance of estimation method we use theperformance indicators in [13] as the estimation indicatorsThe performance indicators in this experiment are defined asfollows

Precision (Prec = |119860||119861|) is used tomeasure the accuracyof resource selection Specifically candidate resources thatsuccessfully match the user QoS make up the percentage ofall resources where 119860 represents candidate resources thatsuccessfully match the user QoS and 119861 represents all theresources

Recall (Rec = |119860||119860 cup 119862|) is used to measure theeffectiveness of resource selection Specifically the candidateresources that successfully match the user account for thepercentage of all matching successful resources where 119862represents the matching successful resources that are notwithin the candidate resources set


PrecisionRecall

02 04 06 08 10Threshold k

0

20

40

60

80

100

Rate

()

Figure 4 Variation of precision and recall under different matchingthreshold

119865-measure (119865-measure = (2 lowast Prec lowast Rec)(Prec + Rec))is the weighted average of precision and recall which is usedto reflect the overall performance The more the 119865-measuretends to 1 the more effective the estimation method is

The experimental platform adopts MATLAB to set up adifferent number of resource nodes randomly Each resourcenode includes multiple QoS attribute information The QoSattributes information is set by simulation of resourcescharacteristics on edge servers including response timeavailability and price The proximity degree 120576 and the weight120572 are 05 In order to ensure the effectiveness of the result weconduct the experiment 20 times and take the average valueas the experimental result

In order to obtain the best performance we set up amatching threshold 119896 to filter irrelevant resources by lowsimilarity Taking into account the contradictory relationbetween precision and recall we decide to use 119865-measure asthe initial standard to find the optimal threshold

Figure 4 shows the variation of precision and recallunder different matching threshold As can be seen fromFigure 4 the precision and recall are inversely related Whenthe matching degree is higher the precision is lower andthe recall is higher In order to ensure the precision andrecall within acceptable limits the threshold 119896 should bebetween 05 and 06 Figure 5 shows the change of 119865-measureunder different matching threshold It can be seen thatwith the increase of threshold the 119865-measure is significantlydecreased When the 119896 value is 05 the 119865-measure is in thehigher position so 119896 = 05 is taken as thematching threshold

Figures 6 and 7 show the precision and recall performanceof the method respectively Our method is compared withtwo methods where the first does not consider the penaltyfactor (denoted as SMNP method) and the second doesnot consider the penalty factor and Grey relational grade

01

02

03

04

05

06

07

08

09

1

F-m

easu

re

02 03 04 05 06 07 08 09 101Threshold k

Figure 5 Variation of 119865-measure under different matching thresh-old

Similarity matching algorithmSMNP methodSMNG method

05

055

06

065

07

075

08

085

09

095

1Pr

ecisi

on

300 600 900 1200 15000Numbers of resources

Figure 6 The comparison of precision of different methods underdifferent numbers of resources

(denoted as SMNGmethod) respectively As we can see ourmethod is superior to other methods on precision and staysabove 90 Since the penalty factor improves the similarityof resources it makes the resources meet the needs of usersas much as possible But the recall is lower than the other twomethods and remains at around 72 within the acceptablelevel Although the method can avoid the constraint andrelationship between the attributes of resources it onlymatches the quantity attribute of the resources

Figure 8 shows the 119865-measure under the different num-bers of resources It can be seen that our method is superior




06

065

07

075

08

Reca

ll

Figure 7 The comparison of recall of different methods underdifferent numbers of resources

Table 1 Residual state interval

State Residual interval(1) (minus20 minus10](2) (minus10 minus5](3) (minus5 0](4) (0 5](5) (5 10](6) (10 20]to other methods The 119865-measure is maintained above 80which further validates the effectiveness of our method

In this paper we take CPU utilization as resource loadindex to predict We record CPU utilization of downloadingfiles every 5 seconds to obtain time series dataWeuse the first10 timesrsquo load data to make short-term prediction of the next5 timesrsquo CPU utilization and prove the effectiveness of themethod by the error values The residual sequence is dividedinto the following 6 states by linear regression analysis asshown in Table 1

As the regression-Markov chain prediction method isinterval prediction method the prediction result is intervaldistribution We select the state interval with the maximumprobability as the prediction interval We select (minus5 0]state interval as a result by analyzing the probability of eachstate interval in the stable state

The error result of the regression-Markov chain pre-diction method is shown in Figure 9 It can be seen thatthe method has a higher prediction accuracy and less errorcompared with the linear regression prediction method Dueto the large fluctuation and randomness of the load the singleprediction method has a poor effect The linear regressionmethod can predict the changing trend of the load statebut it can not reflect the stochastic volatility The Markov


06

065

07

075

08

085

09

F-m

easu

re


Figure 8The comparison of 119865-measure of different methods underdifferent numbers of resources

Regression predictionRegression-Markov chain prediction

110 120 130 140100Time (s)

0

1

2

3

4

5

6

7

8

Erro

r (

)

Figure 9 Comparison of errors between regression prediction andregression-Markov chain prediction in the future time

chain solves the stochastic volatility problem and improvesthe accuracy of the prediction method

6 Conclusion and Future Work

With the rapid development of Internet of Things and cloudcomputing service QoS and user satisfaction become animportant challenge Local computing and storage capabili-ties of edge computing can reduce latency and improve usersatisfaction In this paper we useweighted Euclidean distancesimilarity to classify multiple QoS attribute resources We


select the appropriate resources by similarity matching andregression-Markov chain prediction method Since the QoSattributes system is extensible and the user QoS requirementsare dynamic the estimation method has certain scalabilityOn the basis of the existing work we can design a reasonablemethod of resource estimation to balance the satisfactionbetween users and service providers and improve the utiliza-tion of resources

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This work is supported by the National Natural Sci-ence Foundation of China (61672321 61771289) the Shan-dong provincial Postgraduate Education Innovation Program(SDYY14052 SDYY15049) the Shandong Provincial Spe-cialized Degree Postgraduate Teaching Case Library Con-struction Program the Shandong Provincial PostgraduateEducation Quality Curriculum Construction Program theShandongProvincialUniversity Science andTechnology PlanProject (J16LN15) and the Qufu Normal University Scienceand Technology Plan Project (xkj201525)

References

[1] L Atzori A Iera and G Morabito ldquoThe internet of things asurveyrdquoComputer Networks vol 54 no 15 pp 2787ndash2805 2010

[2] D Evans The Internet of Things How the Next Evolution of theInternet is Changing Everything Cisco Systems 2011

[3] W Shi J Cao Q Zhang Y Li and L Xu ldquoEdge computingvision and challengesrdquo IEEE Internet of Things Journal vol 3no 5 pp 637ndash646 2016

[4] F Bonomi R Milito J Zhu and S Addepalli ldquoFog computingand its role in the internet of thingsrdquo in Proceedings of the MccWorkshop onMobile Cloud Computing pp 13ndash16 ACM August2012

[5] S M Hemam and O Hioual ldquoLoad balancing issue in cloudservices selection by using MCDA and Markov Chain Modelapproachesrdquo in Proceedings of the International Conference onCloud Computing Technologies and Applications pp 163ndash169IEEE May 2016

[6] MMAl-Sayed S Khattab and F AOmara ldquoPredictionmech-anisms for monitoring state of cloud resources using Markovchainmodelrdquo Journal of Parallel andDistributedComputing vol96 pp 163ndash171 2016

[7] Y Mao J Zhang S H Song and K B Letaief ldquoStochastic jointradio and computational resource management for multi-usermobile-edge computing systemsrdquo IEEETransactions onWirelessCommunications vol 16 no 9 pp 5994ndash6009 2017

[8] S Shekhar and A Gokhale ldquoDynamic resource managementacross cloud-edge resources for performance-sensitive appli-cationsrdquo in Proceedings of the 17th IEEEACM InternationalSymposium on Cluster Cloud and Grid Computing pp 707ndash710IEEE Press May 2017

[9] N Wang B Varghese M Matthaiou and D S NikolopoulosldquoENORM a framework for edge node resource managementrdquoIEEE Transactions on Services Computing 2017

[10] M Aazam and E-N Huh ldquoFog computing micro datacenterbased dynamic resource estimation and pricing model for IoTrdquoin Proceedings of the IEEE 29th International Conference onAdvanced Information Networking and Applications (AINA rsquo15)vol 51 no 5 pp 687ndash694 IEEE Computer Society March 2015

[11] M Aazam M St-Hilaire C-H Lung and I LambadarisldquoMeFoRE QoE based resource estimation at Fog to enhanceQoS in IoTrdquo in Proceedings of the 23rd International Conferenceon Telecommunications (ICT rsquo16) pp 1ndash5 May 2016

[12] L-Y Zuo Z-B Cao and S-B Dong ldquoVirtual resource evalua-tion model based on entropy optimized and dynamic weightedin cloud computingrdquo Journal of Software vol 24 no 8 pp 1937ndash1946 2013

[13] S S Zhao Y Zhang L Yu B Cheng Y Ji and J ChenldquoA multidimensional resource model for dynamic resourcematching in internet of thingsrdquo Concurrency amp ComputationPractice amp Experience vol 27 no 8 pp 1819ndash1843 2015

[14] J Zhou M Yan X Ye and H Lu ldquoAn algorithm of resourceevaluation and selection based on Multi-QoS constraintsrdquo inProceedings of the Web Information Systems amp ApplicationsConference pp 49ndash52 IEEE Computer Society 2010

[15] W E Dong W U Nan and L I Xu ldquoQoS-oriented monitoringmodel of cloud computing resources availabilityrdquo in Proceedingsof the 5th International Conference on Computational andInformation Sciences pp 1537ndash1540 June 2013

[16] S Ding C Xia Q Cai K Zhou and S Yang ldquoQoS-awareresource matching and recommendation for cloud computingsystemsrdquo Applied Mathematics and Computation vol 247 no15 pp 941ndash950 2014

[17] J Pan and J McElhannon ldquoFuture edge cloud and edgecomputing for internet of things applicationsrdquo IEEE Internet ofThings Journal 2017

[18] L Zhang and D Y Xu ldquoMulti-step optimized GM (1 1) model-based short term resource load prediction in cloud computingrdquoComputer Engineering and Applications vol 50 no 10 pp 65ndash71 2014

International Journal of

AerospaceEngineeringHindawiwwwhindawicom Volume 2018

RoboticsJournal of

Hindawiwwwhindawicom Volume 2018


Active and Passive Electronic Components

VLSI Design



Shock and Vibration


Civil EngineeringAdvances in

Acoustics and VibrationAdvances in



Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawiwwwhindawicom

Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Control Scienceand Engineering

Journal of



Journal ofEngineeringVolume 2018

SensorsJournal of



RotatingMachinery


Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018


Chemical EngineeringInternational Journal of Antennas and

Propagation




Navigation and Observation


Hindawi

wwwhindawicom Volume 2018

Advances in

Multimedia

Submit your manuscripts atwwwhindawicom


the resource itself Finally we prove the effectiveness of theestimation method through simulation experiment

The remainder of this paper is organized as follows Therelated works are introduced in Section 2 We describe thearchitecture of the edge computing in Section 3 In Section 4we describe the method of dynamic resource estimationThe experiment results are presented in Section 5 Section 6concludes the paper

2 Related Work

Edge computing is the extension of cloud computing tonetwork edge There is no standard architecture The workof resource estimation has not been well addressed Mao etal [7] developed a joint radio and computational resourcemanagement algorithm for multiuser MEC systems with theobjective of minimizing the average weighted sum powerconsumption of the mobile devices and the MEC serverShekhar and Gokhale [8] proposed Dynamic Data DrivenCloud and Edge Systems (D3CES) as a framework for adap-tive resource management across cloud and edge resourcesto provide QoS guarantees for performance-sensitive appli-cations Wang et al [9] developed an Edge NOde ResourceManagement (ENORM) framework to manage edge nodesThey proposed amechanism for provisioning and autoscalingedge node resources Their method reduced data transmis-sion and communication frequency between the edge nodeand cloud

Aazam and Huh [10] proposed a dynamic resourceestimation and pricing model for IoTTheir work was mainlyfocused on considering different types of customers anddevices and the service-oriented relinquish probabilitiesBased on previous studies Aazam et al [11] proposed aQoE-based resource estimation model to enhance QoSTheydevised MeFoRE method on the basis of service relinquishprobability and historical records And they considered pre-vious QoE and Net Promoter Score records to enhance QoSZuo et al [12] proposed a resource estimationmodel based onentropy optimization and dynamicweightedmethod to accu-rately grasp the dynamic load and availability information ofresources They used the objective function and constraintcondition of maximum entropy to select the resources whichmeet the QoS of userThismethod achieved optimal schedul-ing and guaranteed the QoS of user Dynamic weighted loadevaluation was carried out on the filtered resources to realizeload balancing and improve system utilization

Zhao et al [13] proposed a multidimensional resourcemodel for dynamic resource matching in the IoT with themultidimensional description of the IoT resources Theyconsidered the ontology structure and ontology descriptionto calculate the similarity by matching the hardware featuresand software feature between resources Zhou et al [14]proposed a multiple QoS resource selection and estimationalgorithm They used Multiple Attribute Decision Makingand Analytic Hierarchy Process (AHP) algorithm Theirmethod mainly considered the preference information ofusers Dong et al [15] analyzed the QoS of a single resourcenode They used a local optimization algorithm to select the

appropriate cloud resources that meet the needs of usersDing et al [16] proposed QoS-aware resource matchingand recommendation method in cloud computing Theydescribed a resource matching algorithm that considers bothfunctional requirements and QoS attributes and designed aresource recommendation method for cloud computing

3 Edge Computing

Edge computing located between themobile end devices andcloud data center provides computing storage and networkservices Edge computing has the ability of decentralizedcomputation and storage compared with cloud computingThe resources near the user can support mobile devices forreal-time communication The main goal of edge computingis to preprocess data reduce the delay and serve the appli-cations of low delay and real-time response Currently thecommon edge computing architecture is a three-tier networkarchitecture as shown in Figure 1

As the underlying node the end devices not only con-sume data but also produce data In particular mobiledevices will require more resource from edge servers ratherthan cloud data center The end devices include all IoTdevices such as smart mobile phone intelligent vehicle andvirtual sensor nodes These devices are able to sense datacommunicate with the upper layer through sensor networks3G WiFi and so on and transmit the collected raw dataCompared with the servers in data centers most of the IoTdevices or sensors at the edge are somewhat limited in bothcomputing power and battery capacity [17] Edge devicesinclude some traditional network devices (routers switchesetc) and some specially deployed devices (local servers)Edge devices have the ability of computing processing stor-age and forwarding data to the cloud servers Edge serverscan be connected directly to nearby mobile devices via one-hopwireless linkMicro data center (MDC) resources includecomputing resources network resources storage resourcesand software resources On the one hand resources comefrom the cached local resources on the other hand they comefrom the data obtained by the monitoring equipment As thetop layer cloud computing layer includes data center (DC)and some cluster servers The cloud servers can receive datasent by edge devices for processing and storage

Figure 2 shows the request process of the service in theedge computing The user first submits the request to thesystem administrator through the end devices The systemadministrator stores the submitted information and transmitsthe statistical QoS requirements to the edge servers throughthe edge gateway Monitoring equipment generates statisticslog by tracking resource utilization and availability of sensorapplications and servicesThemonitoring data is transmittedto the edge servers The edge servers analyze the QoS ofrequested service for end users and process locally Theservice is divided into several tasks which are processedlocally as far as possible Each task selects the best matchingresource from the resource pool based on the estimationresultsThen the selected resource is allocated and scheduledto meet the QoS requirements


DC

Resource

MDC

Storage

Computing Network

3G4G WiFi

End device

Edge server

Cloud server

Software


Systemadministrator


Edgegateway

Cloudgateway

Monitor

Cloud data center

Edge server


Resource pool

Task 2

Serviceresponse

Request service

End user

QoSrequirements


response

Service response

Serviceresponse


Cloudcomputing

QoS

requirements

Monitoring data

Task n









Start


Resource matrix R






End

Yes

No

Yes

No



resource set Rs










119911119894119895 = 119903119894119895 minus 119903min

119895119903max119895 minus 119903min

119895


119895




119894=1


119908119895 = 1(7)




119889 (119894 119906119902) = radic 119898sum119895=1

119908119895 lowast (119903119902119895 minus 119906119902119895)2119889sim (119894 119906119902) = 11 + 119889 (119894 119906119902) (8)

















]]]]]]] (14)




120585119894119895 (15)













1198721199041 1198721199042 sdot sdot sdot 119872119904119904]]]]]]]

(19)



119909119894 = 1 (20)







PrecisionRecall


0

20

40

60

80

100

Rate

()







01

02

03

04

05

06

07

08

09

1

F-m

easu

re

02 03 04 05 06 07 08 09 101Threshold k



05

055

06

065

07

075

08

085

09

095

1Pr

ecisi

on








06

065

07

075

08

Reca

ll








06

065

07

075

08

085

09

F-m

easu

re




110 120 130 140100Time (s)

0

1

2

3

4

5

6

7

8

Erro

r (

)









Acknowledgments


References





















RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia



DC

Resource

MDC

Storage

Computing Network

3G4G WiFi

End device

Edge server

Cloud server

Software


Systemadministrator


Edgegateway

Cloudgateway

Monitor

Cloud data center

Edge server


Resource pool

Task 2

Serviceresponse

Request service

End user

QoSrequirements


response

Service response

Serviceresponse


Cloudcomputing

QoS

requirements

Monitoring data

Task n









Start


Resource matrix R






End

Yes

No

Yes

No



resource set Rs










119911119894119895 = 119903119894119895 minus 119903min

119895119903max119895 minus 119903min

119895


119895




119894=1


119908119895 = 1(7)




119889 (119894 119906119902) = radic 119898sum119895=1

119908119895 lowast (119903119902119895 minus 119906119902119895)2119889sim (119894 119906119902) = 11 + 119889 (119894 119906119902) (8)

















]]]]]]] (14)




120585119894119895 (15)













1198721199041 1198721199042 sdot sdot sdot 119872119904119904]]]]]]]

(19)



119909119894 = 1 (20)







PrecisionRecall


0

20

40

60

80

100

Rate

()







01

02

03

04

05

06

07

08

09

1

F-m

easu

re

02 03 04 05 06 07 08 09 101Threshold k



05

055

06

065

07

075

08

085

09

095

1Pr

ecisi

on








06

065

07

075

08

Reca

ll








06

065

07

075

08

085

09

F-m

easu

re




110 120 130 140100Time (s)

0

1

2

3

4

5

6

7

8

Erro

r (

)









Acknowledgments


References





















RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia



Start


Resource matrix R






End

Yes

No

Yes

No



resource set Rs










119911119894119895 = 119903119894119895 minus 119903min

119895119903max119895 minus 119903min

119895


119895




119894=1


119908119895 = 1(7)




119889 (119894 119906119902) = radic 119898sum119895=1

119908119895 lowast (119903119902119895 minus 119906119902119895)2119889sim (119894 119906119902) = 11 + 119889 (119894 119906119902) (8)

















]]]]]]] (14)




120585119894119895 (15)













1198721199041 1198721199042 sdot sdot sdot 119872119904119904]]]]]]]

(19)



119909119894 = 1 (20)







PrecisionRecall


0

20

40

60

80

100

Rate

()







01

02

03

04

05

06

07

08

09

1

F-m

easu

re

02 03 04 05 06 07 08 09 101Threshold k



05

055

06

065

07

075

08

085

09

095

1Pr

ecisi

on








06

065

07

075

08

Reca

ll








06

065

07

075

08

085

09

F-m

easu

re




110 120 130 140100Time (s)

0

1

2

3

4

5

6

7

8

Erro

r (

)









Acknowledgments


References





















RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia




119889 (119894 119906119902) = radic 119898sum119895=1

119908119895 lowast (119903119902119895 minus 119906119902119895)2119889sim (119894 119906119902) = 11 + 119889 (119894 119906119902) (8)

















]]]]]]] (14)




120585119894119895 (15)













1198721199041 1198721199042 sdot sdot sdot 119872119904119904]]]]]]]

(19)



119909119894 = 1 (20)







PrecisionRecall


0

20

40

60

80

100

Rate

()







01

02

03

04

05

06

07

08

09

1

F-m

easu

re

02 03 04 05 06 07 08 09 101Threshold k



05

055

06

065

07

075

08

085

09

095

1Pr

ecisi

on








06

065

07

075

08

Reca

ll








06

065

07

075

08

085

09

F-m

easu

re




110 120 130 140100Time (s)

0

1

2

3

4

5

6

7

8

Erro

r (

)









Acknowledgments


References





















RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia











1198721199041 1198721199042 sdot sdot sdot 119872119904119904]]]]]]]

(19)



119909119894 = 1 (20)







PrecisionRecall


0

20

40

60

80

100

Rate

()







01

02

03

04

05

06

07

08

09

1

F-m

easu

re

02 03 04 05 06 07 08 09 101Threshold k



05

055

06

065

07

075

08

085

09

095

1Pr

ecisi

on








06

065

07

075

08

Reca

ll








06

065

07

075

08

085

09

F-m

easu

re




110 120 130 140100Time (s)

0

1

2

3

4

5

6

7

8

Erro

r (

)









Acknowledgments


References





















RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia



PrecisionRecall


0

20

40

60

80

100

Rate

()







01

02

03

04

05

06

07

08

09

1

F-m

easu

re

02 03 04 05 06 07 08 09 101Threshold k



05

055

06

065

07

075

08

085

09

095

1Pr

ecisi

on








06

065

07

075

08

Reca

ll








06

065

07

075

08

085

09

F-m

easu

re




110 120 130 140100Time (s)

0

1

2

3

4

5

6

7

8

Erro

r (

)









Acknowledgments


References





















RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia





06

065

07

075

08

Reca

ll








06

065

07

075

08

085

09

F-m

easu

re




110 120 130 140100Time (s)

0

1

2

3

4

5

6

7

8

Erro

r (

)









Acknowledgments


References





















RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia






Acknowledgments


References





















RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia




RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia


Method of Resource Estimation Based on QoS in...

Documents

Transcript of Method of Resource Estimation Based on QoS in...