research article optimization approach for...

Research ArticleOptimization Approach for Resource Allocation on CloudComputing for IoT

Yeongho Choi1 and Yujin Lim2

1Department of Computer Science University of Suwon San 2-2 Wau-ri Bongdam-eup HwaseongGyeonggi-do 445-743 Republic of Korea2Department of Information Technology Engineering Sookmyung Womenrsquos University Cheongpa-ro 47-gil 100Yongsan-gu Seoul 04310 Republic of Korea

Correspondence should be addressed to Yujin Lim yujin91sookmyungackr

Received 28 December 2015 Accepted 28 January 2016

Academic Editor Fan Wu

Copyright copy 2016 Y Choi and Y Lim 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

Combinatorial auction is a popular approach for resource allocation in cloud computing One of the challenges in resourceallocation is that QoS (Quality of Service) constraints are satisfied and providerrsquos profit is maximized In order to increase theprofit the penalty cost for SLA (Service Level Agreement) violations needs to be reduced We consider execution time constraintas SLA constraint in combinatorial auction system In the system we determine winners at each bidding round according to thejobrsquos urgency based on execution time deadline in order to efficiently allocate resources and reduce the penalty cost To analyze theperformance of our mechanism we compare the providerrsquos profit and success rate of job completion with conventional mechanismusing real workload data

1 Introduction

Over recent years the growth of the IoT (Internet of Things)is expected to create a pervasive connection of things such asembedded devices sensors and actuatorsThiswill inevitablyresult in the generation of enormous amount of datawhich have to be autonomously stored processed accessedand managed Cloud computing has been recognized as aparadigm for the big data problem [1] Cloud computingallows the sensing data to be stored and used intelligently forsmart applications

One of challenges which arose when IoT meets cloudis resource allocation by which a cloud provider efficientlyallocates its resources to cloud users with SLA constraintsThe dominating performance factors in resource allocationinclude providerrsquos and userrsquos profit resource utilization andQoS (Quality of Service) [2] For resource allocation in cloudauction-based model is a popular approach in resource allo-cation and pricing [3] In particular combinatorial auction ispreferred in cloud computing because it allows a user to buy apackage of resources rather than an individual resource Since

the auction is done on group of resources it provides efficientresource allocation and helps to improve the profit to both aprovider and a user

The other issue of resource allocation in cloud computingis meeting SLA (Service Level Agreement) established witha user Before a provider provisions a service to a user aprovider and a user need to establish SLA contract The SLAis an agreement that specifies QoS between a provider and auser [4] We define two-levels of SLA to represent differentobjectives class-based SLA and job-based SLA [5] In class-based SLA for each job class QoS is measured based onperformance metrics In job-based SLA QoS is measuredusing the metrics of individual jobs Any single job with apoor service quality immediately affects the measured QoSand incurs some SLA penalty cost We believe the job-basedSLA is a more robust type of SLAs from the usersrsquo perspectivethan providersrsquo perspective and we focus on this In generalSLA is defined in terms of various performance metricssuch as service latency throughput consistency and securityIn our paper SLA is defined in terms of deadline alongwith execution time of each job The deadline violation is

Hindawi Publishing CorporationInternational Journal of Distributed Sensor NetworksVolume 2016 Article ID 3479247 6 pageshttpdxdoiorg10115520163479247

2 International Journal of Distributed Sensor Networks

important from QoS guarantees perspective It causes lossof profit for a provider due to incompletion to execute thecertain jobs and SLA penalty cost [6]

Considering SLA guarantee we propose a resourceallocation mechanism in combinatorial auction system forcloud computing For efficient resource allocation winnerdetermination problem in the auction should be solvedThe conventional winner determination mechanisms areproposed to maximize resource utilization and a providerrsquosprofit for each time interval However the profit can bemaximized by reducing the penalty cost for SLA violationTo reduce SLA violations the deadline constraints need tobe considered when winner is determined in the auctionThus we propose a winner determination mechanism withconsideration for the deadline constraints to maximize theproviderrsquos profit

The rest of the paper is organized as follows In Section 2we discuss the related work In Section 3 we present ourwinner determination mechanism In Sections 4 and 5 weevaluate the performance of ourmechanism and conclude thepaper with the future research plan

2 Related Work

We investigate resource allocation strategies for the SLA-based cloud computing framework Several proposals inthe literature are based on utility functions to dynamicallyallocate resources In [7] a two-tier resource managementapproach based on utility functions is presented It defines theVM (Virtual Machine) utility function as a linear function ofthe CPU resources In [8] an autonomic resource manageris presented to control the virtualized environment whichdecouples the provisioning of resources from the dynamicplacement of VMs It finds an optimal number and sizeof VMs to allocate CPU resources to applications whilemaximizing the utility function

There are several proposals that dynamically manageVMs by optimizing objective functions Reference [9] deter-mines dynamic placement of VMs based on minimizing anobjective cost function using linear programming The func-tion is to provide an economic model between infrastructureproviders and users The pMapper [10] system includes apower-aware application placement controller to optimize acost-performance function It uses the bin-packing algorithmbased on a modified version of FFD (First-Fit Decreasing) toplace the VMs on the servers while trying to meet the targetutilization

Researchers have investigated maximizing the profitunder SLA constraints In [11] a multilevel generalizedassignment problem is formulated To solve the problem thefirst-fit heuristic is used for maximizing the profit under SLAand the power budget constraint In [12] a combinatorialoptimization problem is defined to maximize the providerrsquosprofit from SLA compliant placement The problem is pre-sented as a multiunit combinatorial auction To solve theproblem a column generation method is presented to obtainnear optimal solutions

A combinatorial auction allows bidders to bid for acombination of multiple resources In other words bids aresubmitted for the whole bundle as a single unit After the auc-tion each bidder receives either the whole bundle or nothingIn the auction the goal of a cloud provider is to efficientlyallocate the available resources to users and to maximallygenerate its revenue In [13] CA-provision is proposed as anallocation mechanism to maximize the revenue for a cloudprovider as well as maximize the resource utilization Theproviderrsquos revenue has three elements the revenue of theresource the cost of the VM instances that are allocated tothe users and the cost of keeping the remaining resourcesidle The purpose of considering the two costs is to reducethe cloud providerrsquos losses Reference [14] proposes a modelcalled ABRA (Auction-Based Resource Co-Allocation) tosolve the resource coallocation problem It imposes penaltycosts on unallocated resources after an auction in order toimprove the resource utilization First a new neighborhoodstructure is proposed to define an ordering as a permutationof the bids in the auctionThen the search space is defined asthe set of all possible orderings to find the optimum solutionFinally various neighbor selection methods are used to findthe solution

The conventional resource allocation mechanisms incombinatorial auction system present winner determinationalgorithms to reduce the penalty cost in each bidding roundHowever since they do not consider the probability of SLAviolation when deadline gets to end the performance islimited We present a new winner determination algorithmwith consideration for deadline constraints of jobs (ieurgency) to reduce the penalty cost for SLA violation andmaximize the profit of the provider

3 Proposed Mechanism

A cloud provider offers computing services to users through119898 different types of VM instances VM

1VM2 VM

119898

The computing power of a VM instance of type VM119894 119894 =

1 119898 is 119908119894 where 119908

1= 1 119908

1lt 1199082

lt sdot sdot sdot lt 119908119898 and

119908 = (1199081 1199082 119908

119898) [13] We consider that 119908 is determined

mainly by the number of cores like Amazon EC2 [15] Let 119896119894

be the number of VM119894instances provisioned by the provider

Let119872 be themaximumnumber of VM instances provisionedby the provider The provider provisions a combination ofinstances given by the vector (119896

1 1198962 119896

119898) as long as

sum119898

119894=1119908119894119896119894le 119872 We consider 119899 users 119906

1 119906

119899who request

computing resources from the provider specified as bundlesof VM instances A user 119906

119895requests VM instances for its job

job119895by submitting a bid 119861

119895= (119903119895

1 119903

119895

119898 V119895) to the provider

where 119903119895

119894is the number of requested instances of type VM

119894

and V119895is the price user 119906

119895is willing to pay to use the requested

bundle of VMs for a unit of time In addition let us denote by119901119895the amount paid by user 119906

119895for using its requested bundle

of VMs and 119901119895and V119895can be different (usually 119901

119895le V119895) We

assume that the users are single minded which means a userbids for only one bundle

The provider runs auction mechanism periodically toallocate the VM instances Thus users bid for obtaining

International Journal of Distributed Sensor Networks 3

the VM bundles for a unit of time If the userrsquos job requiresa bundle for more than one unit of time the user has tobid again in the next round of the auction separately Auser bids until its application is completed or its deadline isexceeded Conventional winner determination mechanismsdo not consider that each job has different urgency In otherwords the job which has impending deadline needs to havea weight to the job with the time left before the deadline inthe competitive bidding To maximize the providerrsquos profitby reducing the penalty cost for SLA violations we use theprobability of deadline violations by considering the jobrsquosurgency when winners are determined [16]

In our paper SLA is defined in terms of completiondeadline 119889

119895for job job

119895of user 119906

119895 For simplicityrsquos sake we

assume that a user has one job at a time Our problem is that119870119895 119870119895le 119889119895biddings should be succeeded to complete job

119895

before 119889119895gets to endThe current bidding round 120579

119895is the sum

of the number of successful biddings 119896119895 119896119895

lt 119870119895and the

number of bidding failures 119891119895 in other words 120579

119895= 119896119895+ 119891119895

and 120579119895le 119889119895 We define the problem as the combination of 119889

119895

taken119870119895

119889119895119862119870119895

(1)

subject to119889119895ge 119870119895 (2)

The success probability of119870119895biddings before 119889

119895is as follows

1

119889119895119862119870119895

(3)

We identify that the probability of successful job completionneeds to be increased as 119891

119895or 120579119895increases At the current

round 120579119895with 119896

119895successful bidding to complete the job

before 119889119895 119870119895minus 119896119895successful biddings need more The

remaining rounds before 119889119895are 119889119895minus120579119895= 119889119895minus119891119895minus119896119895 Thus to

get the probability of job completion before119889119895 (3) is rewritten

as follows

Pr119895=

1

(119889119895minus119891119895minus119896119895)119862

(119870119895minus119896119895)

=

(119870119895minus 119896119895)

(119889119895minus119891119895minus119896119895)119875

(119870119895minus119896119895)

(4)

Using (4) we calculate the expected value of the providerrsquosprofit for each user to determine the winners at a biddinground The profit is divided into two cases When the jobis completed before deadline the profit is the differencebetween revenue and running cost 119862

119877of the VM instances

to be allocated to the user When the job is not completedbefore deadline and SLA is violated the provider should paythe penalty cost119862penalty

119895With the probability of SLA violation

in (4) we define 119864profit119895

the expected value of the providerrsquosprofit for 119906

119895as follows

119864profit119895

= Pr119895(V119895minus 119862119877

119898

sum

119894=1

119908119894119903119895

119894)

+ (1 minus Pr119895)(V119895minus 119862119877

119898

sum

119894=1

119908119894119903119895

119894minus 119862

penalty119895

)

(5)

0 2 4 6 8 10Number of biddings

Number of bidding successes (kj)Number of bidding failures (fj)

00

02

04

06

08

10

Prj

Figure 1 Pr119895with varying the number of biddings

In (6) we normalize 119864profit119895

with the maximum value of 119864profit119895

to remove the dependence on V119895 The maximum value of

119864profit119895

indicates the expected value when there is no SLAviolation

119864profit119895

V119895minus 119862119877sum119898

119894=1119908119894119903119895

119894

(6)

Using (6) winners are determined at each roundFigure 1 shows that Pr

119895increases as 119896

119895or 119891119895increases in

our mechanism We set 119889119895and 119870

119895to 15 and 10 respectively

Thus when 119891119895gt 5 job

119895is not completed before the deadline

In the figure Pr119895gets to 1 in order to complete the job before

the deadline when 119891119895= 5 Besides as you can see 119891

119895has

the bigger impact than 119896119895on Pr119895 Because 119864profit

119895increases as

Pr119895increases the probability that 119906

119895is to be a winner also

increases Thus the jobs with impending deadlines are likelyto be determined as winners Through the determinationdeadline violation can be reduced and the profit of theprovider can be improved

4 Performance Analysis

To evaluate the performance of our mechanism we con-duct simulation with real workload data and compare theperformance of the conventional mechanism (CA-provision[13]) In preliminary experiments we use UniLu-Gaia-2014workload logs from the ParallelWorkloadArchive [17] In thelogs the average number of submitted jobs per round is set toabout 24 and the average number of processors required perjob is 997 In the experiments we set 119862

119877to 05 and 119862

penalty119895

to10 percent of V

119895 119870119895is randomly selected from (1 20) and the

average execution time of a job is multiplying bidding timeinterval Δ119905 with 119870

119895

Figure 2 shows the probability of bidding success atcurrent round 120579

119895when the remaining bidding rounds before

119889119895vary In CA-provision the probability does not show


OurCA-provision

0 5 10 15 20 2500

02

04

06

08

10

Prob

abili

ty o

f bid

ding

succ

ess

at cu

rren

t rou

nd120579j

Remaining bidding rounds before dj

Figure 2 The probability of bidding success at a current round

OurCA-provision

00

02

04

Succ

ess r

ate o

f job

com

plet

ion

06

08

10

10 20 30 400User ID

Figure 3 The success rate of job completion for users

the big change In our mechanism the probability increasesas the remaining round decreases By considering the jobrsquosurgency the probability gets to 1 as the deadline gets toend and the jobs with impending deadlines are likely to bedetermined as winners

Figure 3 shows the success rate of job completion foreach user The success rate of job completion indicates thenumber of jobs completed before the deadline to the numberof jobs submitted by user 119906

119895in the total simulation time The

success rate of ours is about 274 higher than that of CA-provision In Figure 4 the deadline of a job is determined bymultiplying a deadline factor 120591

119895with 119870

119895 thus the deadline

is 119889119895= 120591119895119870119895 Here 120591

119895is chosen from 15 20 25 and 30

The figure shows the success rate of job completion withvarying the deadline factors As you can see the success rateincreases as the deadline increases The success rate of ourmechanism is about 2 133 198 and 274 higher thanthat of CA-provision respectively Figures 3 and 4 show thatourmechanism effectively reduces the deadline violation andincreases the success rate of job completion

Our CA

Our

CA

Our

CA

Our

CA

Deadline factors

OurCA-provision

00

02

04

06

08

Succ

ess r

ate o

f job

com

plet

ion

10

times15 times20 times25 times30

Figure 4 The success rate of job completion with varying thedeadline factors

OurCA

OurCA

OurCA

OurCA

0

5000

10000

15000

20000Pr

ofit o

f clo

ud p

rovi

der

Deadline factors

OurCA-provision


Figure 5 The providerrsquos profit with varying the deadline factors

Figure 5 shows the profit of a cloud provider with varyingthe deadline factors The profit of our mechanism is about105 12 10 and 13 higher than that of CA-provisionrespectively This figure shows that through the winnerdetermination by considering the jobrsquos urgency deadlineviolation decreases and the profit of the provider increases

To evaluate the performance in different system scenar-ios we use eight workload logs from the Parallel WorkloadArchive [17] In Table 1 we show a brief description ofthe workload files The table describes the log file namethe average execution time of the jobs (119879wl) the averagenumber of jobs submitted for a unit of time (119869wl) theaverage number of processors required per job (119875wl) andthe total number of processors in the system (119872wl) Fromthe real workload data we use several data such as jobnumber execution time the number of allocated processorsaverage CPU time used and user ID Some records in a logfile are not specified because the original files had missinginformation So if a record data is missing we randomly


Table 1 Real workload data

Log file 119879wl (hours) 119869wl (jobshr) 119875wl 119872wl

SDSC-DS 13 1026 6236 8192LLNL-Thunder 5 3362 4236 4008LLNL-Atlas 8 76 401 9216RICC 5 11987 1645 4096PIK-IPLEX 40 2046 3449 2560LCG 1 26116 3452 4096LLNL-uBGL 7 2234 576 2048UniLu-Gaia 3 2406 997 64

Our

CAOur

CA

Our

CA

OurCAOur

CAOur

CA

Our

CA

Our

CA

Succ

ess r

ate o

f job

com

plet

ion

00

02

04

06

08

10

per p

roce

ssor

-rou

nd

OurCA-provision

Workload file (normalized load)

SDSC

-DS(021

)

LLN

L-Th

unde

r(053

)

LLN

L-At

las(082

)

PIK-

IPLE

X(237

)

LCG

(765

)

LLN

L-uB

GL(785

)

Uni

Lu-G

aia998400 (1

49

)

RICC

998400(158

)

Figure 6 The success rate of job completion with varying theworkloads

generate the record data within the average range of theother records using a uniform distribution In RICC andUniLu-Gaia of the eight workloads we decrease119872wl by about70 to make competitive environment Thus we mark thetwo workloads with 119877119868119862119862

1015840 and 119880119899119894119871119906-1198661198861198941198861015840 in the figures

Since theworkloads are heterogeneous in several dimensionswe need normalization for workload logs To normalizeworkload logs we use normalized load in a workload wl120578wl = (119869wltimes119879wltimes119875wl)119872wlThe normalized loadmeasures theaverage amount of load per processor When analyzing theresults we use the normalized load to rank the heterogeneouslog files

Figure 6 shows the success rate of job completion indifferent workloads In the experiments we set 120591

119895to 20 Our

mechanism has 182 more success rate than CA-provisionFigure 7 shows the profit of a cloud provider Since theworkloads are generated for different durations of time forsystems with different number of processors we scale theprofit with respect to the total simulation hours and the

Our

CA

Our

CA Our

CA

Our

CA

Our

CA

Our

CAOur

CA OurCA

Profi

t of c

loud

pro

vide

r

OurCA-provision


00

02

04

06

08

10

12

14

per p

roce

ssor

-rou

nd

SDSC

-DS(021

)

LLN

L-Th

unde

r(053

)

LLN

L-At

las(082

)

PIK-

IPLE

X(237

)

LCG

(765

)

LLN

L-uB

GL(785

)

Uni

Lu-G

aia998400 (1

49

)

RICC

998400(158

)

Figure 7 The providerrsquos profit with varying the workloads

number of processors We define the profit per processor-hour as Π

prwl = Πwl(119872wl times 119877wl) where Πwl is the sum

of profits in all bidding round and 119877wl is the number ofbiddings provided in each workload [13] As you can see ourmechanism has about 268 more profit than CA-provisionAs a result our mechanism shows the better performancewith respect to the success rate and the profit than CA-provision in various workload scenarios

5 Conclusion

To increase the providerrsquos profit the penalty cost for SLAviolations needs to be considered To reduce the cost weconsider the jobrsquos urgency based on the deadline constraintwhen winners are determined in the combinatorial auctionTaking the urgency into consideration we calculate theprobability of deadline violation for each jobThen using theprobability we calculate the expected value of the providerrsquosprofit when the corresponding user is selected as a winnerat a bidding round The user with the larger expected valueis likely to be determined as a winner Thus the penaltycost decreases by the decrease of SLA violation and theproviderrsquos profit increasesThe experimental results show thatour mechanism has higher profit and success rate of jobcompletion than these of the conventional mechanism Welook forward to compare the other winner determinationmechanisms in the combinatorial auction and demonstrateeffectiveness of our mechanism

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper


Acknowledgment

This research was supported by Basic Science ResearchProgram through the National Research Foundation ofKorea (NRF) funded by the Ministry of Education (NRF-2015R1D1A1A09057141)

References

[1] H-L Truong and S Dustdar ldquoPrinciples for engineering IoTcloud systemsrdquo IEEE Cloud Computing vol 2 no 2 pp 68ndash762015

[2] U Lampe M Siebenhaar A Papageorgiou D Schuller and RSteinmetz ldquoMaximizing cloud provider profit from equilibriumprice auctionsrdquo in Proceedings of the IEEE 5th InternationalConference on Cloud Computing (CLOUD rsquo12) pp 83ndash90Honolulu Hawaii USA June 2012

[3] S R Shirley and P Karthikeyan ldquoA survey on auction basedresource allocation in cloud environmentrdquo International Journalof Research in Computer Applications and Robotics vol 1 no 9pp 96ndash102 2013

[4] S Son G Jung and S C Jun ldquoAn SLA-based cloud computingthat facilitates resource allocation in the distributed data centersof a cloud providerrdquo Journal of Supercomputing vol 64 no 2pp 606ndash637 2013

[5] H J Moon Y Chi and H Hacıgumus ldquoSLA-aware profitoptimization in cloud services via resource schedulingrdquo in Pro-ceedings of the 6th International Conference on World Congresson Services (SERVICES rsquo10) pp 152ndash153 IEEEMiami Fla USAJuly 2010

[6] M Alrokayan A V Dastjerdi and R Buyya ldquoSLA-awareprovisioning and scheduling of cloud resources for big data ana-lyticsrdquo in Proceedings of the 3rd IEEE International Conferenceon Cloud Computing for Emerging Markets (CCEM rsquo14) pp 1ndash8Bangalore India October 2014

[7] D Minarolli and B Freisleben ldquoUtility-based resource alloca-tion for virtual machines in cloud computingrdquo in Proceeding ofthe 16th IEEE Symposium on Computers and Communications(ISCC rsquo11) pp 410ndash417 Kerkyra Greece July 2011

[8] H N Van F D Tran and J-M Menaud ldquoAutonomic virtualresource management for service hosting platformsrdquo in Pro-ceedings of the ICSE Workshop on Software Engineering Chal-lenges of Cloud Computing (CLOUD rsquo09) pp 1ndash8 VancouverCanada May 2009

[9] J-G Park J-M KimH Choi and Y-CWoo ldquoVirtualmachinemigration in self-managing virtualized server environmentsrdquoin Proceeding of the 11th IEEE International Conference onAdvanced Communication Technology (ICACT rsquo09) pp 2077ndash2083 February 2009

[10] A Verma P Ahuja and A Neogi ldquopMapper power andmigration cost aware application placement in virtualizedsystemsrdquo in Middleware 2008 ACMIFIPUSENIX 9th Inter-national Middleware Conference Leuven Belgium December 1ndash5 2008 Proceedings vol 5346 of Lecture Notes in ComputerScience pp 243ndash264 Springer Berlin Germany 2008

[11] W Shi and B Hong ldquoTowards profitable virtual machineplacement in the data centerrdquo in Proceedings of the 4th IEEEInternational Conference on Cloud and Utility Computing (UCCrsquo11) pp 138ndash145 IEEE Victoria Australia December 2011

[12] D Breitgand and A Epstein ldquoSLA-aware placement of multi-virtual machine elastic services in compute cloudsrdquo in Proceed-ings of the IFIPIEEE International Symposium on Integrated

NetworkManagement (IM rsquo11) pp 161ndash168Dublin IrelandMay2011

[13] S Zaman and D Grosu ldquoA combinatorial auction-based mech-anism for dynamic VM provisioning and allocation in cloudsrdquoIEEETransactions onCloudComputing vol 1 no 2 pp 129ndash1412013

[14] A H Ozer and C Ozturan ldquoAn auction based mathematicalmodel and heuristics for resource co-allocation problem ingrids and cloudsrdquo in Proceedings of the 5th International Confer-ence on Soft Computing Computing with Words and Perceptionsin SystemAnalysis Decision and Control (ICSCCW rsquo09) pp 1ndash4Famagusta Cyprus September 2009

[15] Amazon EC2 Spot Instances httpawsamazoncomec2spot-instances

[16] Y Choi andY Lim ldquoResourcemanagementmechanism for SLAprovisioning on cloud computing for IoTrdquo in Proceedings of theInternational Conference on Information and CommunicationTechnology Convergence (ICTC rsquo15) pp 500ndash502 IEEE JejuIsland South Korea October 2015

[17] D G Feitelson ldquoParallel Workloads Archiverdquo httpwwwcshujiacillabsparallelworkloadlogshtml

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Active and Passive Electronic Components

Control Scienceand Engineering

Journal of



RotatingMachinery


Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design



Shock and Vibration


Civil EngineeringAdvances in

Acoustics and VibrationAdvances in



Electrical and Computer Engineering

Journal of

Advances inOptoElectronics


Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of


Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Chemical EngineeringInternational Journal of Antennas and

Propagation




Navigation and Observation



DistributedSensor Networks



important from QoS guarantees perspective It causes lossof profit for a provider due to incompletion to execute thecertain jobs and SLA penalty cost [6]

Considering SLA guarantee we propose a resourceallocation mechanism in combinatorial auction system forcloud computing For efficient resource allocation winnerdetermination problem in the auction should be solvedThe conventional winner determination mechanisms areproposed to maximize resource utilization and a providerrsquosprofit for each time interval However the profit can bemaximized by reducing the penalty cost for SLA violationTo reduce SLA violations the deadline constraints need tobe considered when winner is determined in the auctionThus we propose a winner determination mechanism withconsideration for the deadline constraints to maximize theproviderrsquos profit

The rest of the paper is organized as follows In Section 2we discuss the related work In Section 3 we present ourwinner determination mechanism In Sections 4 and 5 weevaluate the performance of ourmechanism and conclude thepaper with the future research plan

2 Related Work

We investigate resource allocation strategies for the SLA-based cloud computing framework Several proposals inthe literature are based on utility functions to dynamicallyallocate resources In [7] a two-tier resource managementapproach based on utility functions is presented It defines theVM (Virtual Machine) utility function as a linear function ofthe CPU resources In [8] an autonomic resource manageris presented to control the virtualized environment whichdecouples the provisioning of resources from the dynamicplacement of VMs It finds an optimal number and sizeof VMs to allocate CPU resources to applications whilemaximizing the utility function

There are several proposals that dynamically manageVMs by optimizing objective functions Reference [9] deter-mines dynamic placement of VMs based on minimizing anobjective cost function using linear programming The func-tion is to provide an economic model between infrastructureproviders and users The pMapper [10] system includes apower-aware application placement controller to optimize acost-performance function It uses the bin-packing algorithmbased on a modified version of FFD (First-Fit Decreasing) toplace the VMs on the servers while trying to meet the targetutilization

Researchers have investigated maximizing the profitunder SLA constraints In [11] a multilevel generalizedassignment problem is formulated To solve the problem thefirst-fit heuristic is used for maximizing the profit under SLAand the power budget constraint In [12] a combinatorialoptimization problem is defined to maximize the providerrsquosprofit from SLA compliant placement The problem is pre-sented as a multiunit combinatorial auction To solve theproblem a column generation method is presented to obtainnear optimal solutions

A combinatorial auction allows bidders to bid for acombination of multiple resources In other words bids aresubmitted for the whole bundle as a single unit After the auc-tion each bidder receives either the whole bundle or nothingIn the auction the goal of a cloud provider is to efficientlyallocate the available resources to users and to maximallygenerate its revenue In [13] CA-provision is proposed as anallocation mechanism to maximize the revenue for a cloudprovider as well as maximize the resource utilization Theproviderrsquos revenue has three elements the revenue of theresource the cost of the VM instances that are allocated tothe users and the cost of keeping the remaining resourcesidle The purpose of considering the two costs is to reducethe cloud providerrsquos losses Reference [14] proposes a modelcalled ABRA (Auction-Based Resource Co-Allocation) tosolve the resource coallocation problem It imposes penaltycosts on unallocated resources after an auction in order toimprove the resource utilization First a new neighborhoodstructure is proposed to define an ordering as a permutationof the bids in the auctionThen the search space is defined asthe set of all possible orderings to find the optimum solutionFinally various neighbor selection methods are used to findthe solution

The conventional resource allocation mechanisms incombinatorial auction system present winner determinationalgorithms to reduce the penalty cost in each bidding roundHowever since they do not consider the probability of SLAviolation when deadline gets to end the performance islimited We present a new winner determination algorithmwith consideration for deadline constraints of jobs (ieurgency) to reduce the penalty cost for SLA violation andmaximize the profit of the provider

3 Proposed Mechanism

A cloud provider offers computing services to users through119898 different types of VM instances VM

1VM2 VM

119898

The computing power of a VM instance of type VM119894 119894 =

1 119898 is 119908119894 where 119908

1= 1 119908

1lt 1199082

lt sdot sdot sdot lt 119908119898 and

119908 = (1199081 1199082 119908

119898) [13] We consider that 119908 is determined

mainly by the number of cores like Amazon EC2 [15] Let 119896119894

be the number of VM119894instances provisioned by the provider

Let119872 be themaximumnumber of VM instances provisionedby the provider The provider provisions a combination ofinstances given by the vector (119896

1 1198962 119896

119898) as long as

sum119898

119894=1119908119894119896119894le 119872 We consider 119899 users 119906

1 119906

119899who request

computing resources from the provider specified as bundlesof VM instances A user 119906

119895requests VM instances for its job

job119895by submitting a bid 119861

119895= (119903119895

1 119903

119895

119898 V119895) to the provider

where 119903119895

119894is the number of requested instances of type VM

119894

and V119895is the price user 119906

119895is willing to pay to use the requested

bundle of VMs for a unit of time In addition let us denote by119901119895the amount paid by user 119906

119895for using its requested bundle

of VMs and 119901119895and V119895can be different (usually 119901

119895le V119895) We

assume that the users are single minded which means a userbids for only one bundle

The provider runs auction mechanism periodically toallocate the VM instances Thus users bid for obtaining




119895for job job

119895of user 119906



119895


119895is the sum


lt 119870119895and the


119895= 119896119895+ 119891119895


119895

taken119870119895

119889119895119862119870119895

(1)

subject to119889119895ge 119870119895 (2)


119895is as follows

1

119889119895119862119870119895

(3)



round 120579119895with 119896





as follows

Pr119895=

1

(119889119895minus119891119895minus119896119895)119862

(119870119895minus119896119895)

=

(119870119895minus 119896119895)

(119889119895minus119891119895minus119896119895)119875

(119870119895minus119896119895)

(4)







119895as follows

119864profit119895

= Pr119895(V119895minus 119862119877

119898

sum

119894=1

119908119894119903119895

119894)

+ (1 minus Pr119895)(V119895minus 119862119877

119898

sum

119894=1

119908119894119903119895

119894minus 119862

penalty119895

)

(5)



00

02

04

06

08

10

Prj





119864profit119895


119864profit119895

V119895minus 119862119877sum119898

119894=1119908119894119903119895

119894

(6)



119895or 119891119895increases in







119895has


119895increases as






119877to 05 and 119862

penalty119895

to10 percent of V



119895





OurCA-provision

0 5 10 15 20 2500

02

04

06

08

10

Prob

abili

ty o

f bid

ding

succ

ess

at cu

rren

t rou

nd120579j



OurCA-provision

00

02

04

Succ

ess r

ate o

f job

com

plet

ion

06

08

10

10 20 30 400User ID






119895with 119870


is 119889119895= 120591119895119870119895 Here 120591



Our CA

Our

CA

Our

CA

Our

CA

Deadline factors

OurCA-provision

00

02

04

06

08

Succ

ess r

ate o

f job

com

plet

ion

10



OurCA

OurCA

OurCA

OurCA

0

5000

10000

15000

20000Pr

ofit o

f clo

ud p

rovi

der

Deadline factors

OurCA-provision









Our

CAOur

CA

Our

CA

OurCAOur

CAOur

CA

Our

CA

Our

CA

Succ

ess r

ate o

f job

com

plet

ion

00

02

04

06

08

10

per p

roce

ssor

-rou

nd

OurCA-provision


SDSC

-DS(021

)

LLN

L-Th

unde

r(053

)

LLN

L-At

las(082

)

PIK-

IPLE

X(237

)

LCG

(765

)

LLN

L-uB

GL(785

)

Uni

Lu-G

aia998400 (1

49

)

RICC

998400(158

)






119895to 20 Our


Our

CA

Our

CA Our

CA

Our

CA

Our

CA

Our

CAOur

CA OurCA

Profi

t of c

loud

pro

vide

r

OurCA-provision


00

02

04

06

08

10

12

14

per p

roce

ssor

-rou

nd

SDSC

-DS(021

)

LLN

L-Th

unde

r(053

)

LLN

L-At

las(082

)

PIK-

IPLE

X(237

)

LCG

(765

)

LLN

L-uB

GL(785

)

Uni

Lu-G

aia998400 (1

49

)

RICC

998400(158

)





5 Conclusion





Acknowledgment


References





















RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation












119895for job job

119895of user 119906



119895


119895is the sum


lt 119870119895and the


119895= 119896119895+ 119891119895


119895

taken119870119895

119889119895119862119870119895

(1)

subject to119889119895ge 119870119895 (2)


119895is as follows

1

119889119895119862119870119895

(3)



round 120579119895with 119896





as follows

Pr119895=

1

(119889119895minus119891119895minus119896119895)119862

(119870119895minus119896119895)

=

(119870119895minus 119896119895)

(119889119895minus119891119895minus119896119895)119875

(119870119895minus119896119895)

(4)







119895as follows

119864profit119895

= Pr119895(V119895minus 119862119877

119898

sum

119894=1

119908119894119903119895

119894)

+ (1 minus Pr119895)(V119895minus 119862119877

119898

sum

119894=1

119908119894119903119895

119894minus 119862

penalty119895

)

(5)



00

02

04

06

08

10

Prj





119864profit119895


119864profit119895

V119895minus 119862119877sum119898

119894=1119908119894119903119895

119894

(6)



119895or 119891119895increases in







119895has


119895increases as






119877to 05 and 119862

penalty119895

to10 percent of V



119895





OurCA-provision

0 5 10 15 20 2500

02

04

06

08

10

Prob

abili

ty o

f bid

ding

succ

ess

at cu

rren

t rou

nd120579j



OurCA-provision

00

02

04

Succ

ess r

ate o

f job

com

plet

ion

06

08

10

10 20 30 400User ID






119895with 119870


is 119889119895= 120591119895119870119895 Here 120591



Our CA

Our

CA

Our

CA

Our

CA

Deadline factors

OurCA-provision

00

02

04

06

08

Succ

ess r

ate o

f job

com

plet

ion

10



OurCA

OurCA

OurCA

OurCA

0

5000

10000

15000

20000Pr

ofit o

f clo

ud p

rovi

der

Deadline factors

OurCA-provision









Our

CAOur

CA

Our

CA

OurCAOur

CAOur

CA

Our

CA

Our

CA

Succ

ess r

ate o

f job

com

plet

ion

00

02

04

06

08

10

per p

roce

ssor

-rou

nd

OurCA-provision


SDSC

-DS(021

)

LLN

L-Th

unde

r(053

)

LLN

L-At

las(082

)

PIK-

IPLE

X(237

)

LCG

(765

)

LLN

L-uB

GL(785

)

Uni

Lu-G

aia998400 (1

49

)

RICC

998400(158

)






119895to 20 Our


Our

CA

Our

CA Our

CA

Our

CA

Our

CA

Our

CAOur

CA OurCA

Profi

t of c

loud

pro

vide

r

OurCA-provision


00

02

04

06

08

10

12

14

per p

roce

ssor

-rou

nd

SDSC

-DS(021

)

LLN

L-Th

unde

r(053

)

LLN

L-At

las(082

)

PIK-

IPLE

X(237

)

LCG

(765

)

LLN

L-uB

GL(785

)

Uni

Lu-G

aia998400 (1

49

)

RICC

998400(158

)





5 Conclusion





Acknowledgment


References





















RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










OurCA-provision

0 5 10 15 20 2500

02

04

06

08

10

Prob

abili

ty o

f bid

ding

succ

ess

at cu

rren

t rou

nd120579j



OurCA-provision

00

02

04

Succ

ess r

ate o

f job

com

plet

ion

06

08

10

10 20 30 400User ID






119895with 119870


is 119889119895= 120591119895119870119895 Here 120591



Our CA

Our

CA

Our

CA

Our

CA

Deadline factors

OurCA-provision

00

02

04

06

08

Succ

ess r

ate o

f job

com

plet

ion

10



OurCA

OurCA

OurCA

OurCA

0

5000

10000

15000

20000Pr

ofit o

f clo

ud p

rovi

der

Deadline factors

OurCA-provision









Our

CAOur

CA

Our

CA

OurCAOur

CAOur

CA

Our

CA

Our

CA

Succ

ess r

ate o

f job

com

plet

ion

00

02

04

06

08

10

per p

roce

ssor

-rou

nd

OurCA-provision


SDSC

-DS(021

)

LLN

L-Th

unde

r(053

)

LLN

L-At

las(082

)

PIK-

IPLE

X(237

)

LCG

(765

)

LLN

L-uB

GL(785

)

Uni

Lu-G

aia998400 (1

49

)

RICC

998400(158

)






119895to 20 Our


Our

CA

Our

CA Our

CA

Our

CA

Our

CA

Our

CAOur

CA OurCA

Profi

t of c

loud

pro

vide

r

OurCA-provision


00

02

04

06

08

10

12

14

per p

roce

ssor

-rou

nd

SDSC

-DS(021

)

LLN

L-Th

unde

r(053

)

LLN

L-At

las(082

)

PIK-

IPLE

X(237

)

LCG

(765

)

LLN

L-uB

GL(785

)

Uni

Lu-G

aia998400 (1

49

)

RICC

998400(158

)





5 Conclusion





Acknowledgment


References





















RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation













Our

CAOur

CA

Our

CA

OurCAOur

CAOur

CA

Our

CA

Our

CA

Succ

ess r

ate o

f job

com

plet

ion

00

02

04

06

08

10

per p

roce

ssor

-rou

nd

OurCA-provision


SDSC

-DS(021

)

LLN

L-Th

unde

r(053

)

LLN

L-At

las(082

)

PIK-

IPLE

X(237

)

LCG

(765

)

LLN

L-uB

GL(785

)

Uni

Lu-G

aia998400 (1

49

)

RICC

998400(158

)






119895to 20 Our


Our

CA

Our

CA Our

CA

Our

CA

Our

CA

Our

CAOur

CA OurCA

Profi

t of c

loud

pro

vide

r

OurCA-provision


00

02

04

06

08

10

12

14

per p

roce

ssor

-rou

nd

SDSC

-DS(021

)

LLN

L-Th

unde

r(053

)

LLN

L-At

las(082

)

PIK-

IPLE

X(237

)

LCG

(765

)

LLN

L-uB

GL(785

)

Uni

Lu-G

aia998400 (1

49

)

RICC

998400(158

)





5 Conclusion





Acknowledgment


References





















RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










Acknowledgment


References





















RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation











RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation









research article optimization approach for...

Documents