LITERATURE REVIEW: VARIOUS PRIORITY BASED TASK SCHEDULING ALGORITHMS IN CLOUD COMPUTING 1 MS. POOJA.P.VASANI, 2 MR. NISHANT.S. SANGHANI 1 M.Tech. [Software Systems] Student, Patel College of Scence and Technology, Rajv Gandh Proudyogk Vshwavdyalaya, Bhopal, M.P., Inda 2 M.E. [Computer Engneerng] Student, Computer Engneerng Department, V.V.P. Engneerng College, Rajkot, Gujarat, Inda poojavasan@ymal.com, nshantssn@yahoo.co.n, ABSTRACT: Cloud computng s a latest new computng paradgm where applcatons, data and IT servces are provded over the Internet. Cloud computng serves dfferent types of the resources n vrtualzed form whch can be utlsed dynamcally. There are dfferent number of ssues whch can be researched out for the proper allocaton and better utlzaton of the vrtualzed resources usng schedulng. In the process of schedulng some ntensve data or computng an ntensve applcaton, t s acknowledged that optmzng the transferrng and processng tme s crucal to an applcaton program. The Task management s the key role n cloud computng systems. Task schedulng problems are premer whch relate to the effcency of the whole cloud computng facltes. The man am of ths work s to study varous problems, ssues and types of task schedulng algorthms for cloud computng. Keywords Cloud Computng, Task Schedulng. 1. INTRODUCTION The cloud computng s a large group of nterconnected computers and cloud symbol represents a group of systems or complcated networks. Cloud computng s one way of communcaton among the varous system n the network wth the help of nternet. [9] Cloud computng s a computng paradgm, where a large pool of systems are connected n prvate or publc networks, to provde dynamcally scalable nfrastructure for applcaton, data and fle storage. Wth the advent of ths technology, the cost of computaton, applcaton hostng, content storage and delvery s reduced sgnfcantly. [8] Cloud computng s a marketng term for technologes that provde computaton, software, data access, and storage servces that do not requre enduser knowledge of the physcal locaton and confguraton of the system that delvers the servces. [1] Cloud computng s on demand resources provsonng whch means to provde the avalable resources based on the requrement of the resources. Cloud computng s a subscrpton based. Cloud computng s a pay as per usage and relable leads to an effcent network. Cloud Computng s an emergng technque and t s very successful because of the followng features lke relable, secure, fast, fault tolerance and effcent communcaton etc., among dfferent network.[9] Schedulng s used to allocate partcular resources for a certan tasks n partcular tme. Task schedulng problem s a core and challengng ssue n cloud computng. The Task executon tme cannot be predcted n cloud computng. Hence the scheduler must be dynamc. The purpose of schedulng s to ncrease the utlzaton of resources. [9] Schedulng process n cloud can be generalzed nto three stages namely Resource dscoverng and flterng- Datacenter broker dscovers the resources present n the network system and collects status nformaton related to them. Resource selecton- Target resource s selected based on certan parameters of task and resource. Ths s a decdng stage. Task submsson- Task s submtted to resource selected. [10] ISSN: 0975 6760 NOV 12 TO OCT 13 VOLUME 02, ISSUE 02 Page 298
CLOUD SERVICES Cloud computng can be thought as dfferent layers or models whch provde dfferent servces. Cloud contans three types of servces as follows. 1. Infrastructure-as-a-Servce(IaaS)- Ths type of cloud computng dstrbutes a full computer nfrastructure va the Internet. Most popular IaaS provder lke Amazon Web Servces offers vrtual server nstances wth unque IP addresses and block of storage on demand. Here customers usually use the servce provder s applcaton program nterface to start, stop, access, modfy and confgure ther vrtual servers and storage as s needed. In the enterprse, cloud computng allocates servces to a company to pay for only as much faclty as s requred, and brng more flexble tools onlne as soon as requred. [11] CLOUD DEPLOYMENT MODELS Cloud Computng technology and servces can be mplemented n dfferent ways accordng to ther purpose and characterstcs. These dfferent types of deployment of Cloud are categorzed n four ways as follows. [1-7] 1. Publc Cloud: A publc cloud s one based on the standard cloud computng model, n whch a servce provder makes resources, such as applcatons and storage, avalable to the general publc over the Internet. Publc cloud servces may be free or offered on a pay-per-usage model. 2. Communty cloud: Communty cloud shares nfrastructure between several organzatons from a specfc communty wth common concerns (securty, complance, jursdcton, etc.), whether managed nternally or by a thrd-party and hosted nternally or externally. The costs are spread over fewer users than a publc cloud (but more than a prvate cloud), so only some of the benefts of cloud computng are realzed. 3. Hybrd cloud: Hybrd cloud s a composton of two or more clouds (prvate, communty, or publc) that reman unque enttes but are bound together, offerng the benefts of multple deployment models. It can also be defned as multple cloud systems that are connected n a way that allows programs and data to be moved easly from one deployment system to another. Fg1. Cloud Servces [11] 2. Platform-as-a-Servce(PaaS)- Ths type of cloud computng offers a product development tool or envronment that users can access and utlze onlne, even n collaboraton wth others and hosted on the provder's nfrastructure. In PaaS, developers create applcatons on the servce provder's platform over the Internet. PaaS servce provders may use Applcaton Program Interfaces (APIs), gateway software or webste portals nstalled on the customer's premses. [11] 3. Software-as-a-Servce (SaaS) - Ths type of cloud computng model offers users the hardware nfrastructure, the software product and nterrelates wth the users through a front-end gateway or portal. Here a provder authorzes an applcaton to clents ether as a servce on demand n a "pay-as-you-go" model or at no charge by a subscrpton. Lke we can say that we have a sky drve n Hotmal. There we can save our word, PowerPont document and when we need to edt them we can easly do t. [11] 4. Prvate cloud: Prvate cloud s nfrastructure operated solely for a sngle organzaton, whether managed nternally or by a thrd-party and hosted nternally or externally. Fg2. Cloud deployment models [1-7] 2. ALGORITHMS AND METRICS A. Tasks Schedulng optmzaton for the Cloud Computng Systems ISSN: 0975 6760 NOV 12 TO OCT 13 VOLUME 02, ISSUE 02 Page 299
Task schedulng algorthm s an NPcompleteness problem whch play key role n cloud computng. In Hadoop, the open-source mplementaton of MapReduce, many schedulng polces such as FIFO schedulng s used by the master node to dstrbute watng tasks to computng nodes. So t can stll acheve more mprovement n the task schedulng process. Therefore, n ths paper, we purpose an optmzed algorthm based on the Fuzzy-GA optmzaton whch makes a schedulng decson by evaluatng the entre group of task n the job queue. [12] Here, the model works on the followng aspects Task Model and Related Propertes Suppose that a cloud computng system consst of heterogeneous process unt possess of m (m>1) unts. The tasks n ths system features are as follows: a) Tasks are aperodc;.e., task arrval tmes are not known a pror. Every task has the attrbutes arrval tme (a), worst case computaton tme (c), and deadlne (d). The ready tme of a task s equal to ts arrval tme. b) Tasks are nonpre-emptve; each of them s ndependent. c) Each task has two ways of access to a process unt:(1)exclusve access, n whch case no other task can use the resource wth t, or (2) shared access, n whch case t can share the resource wth another task (the other task also should be wllng to share the resource). Predcted Executon Tme Mode The statstcal modelng approach - Kernel Canoncal Correlaton Analyss (KCCA), whch representng each Hadoop job as a feature vector of job characterstcs and a correspondng vector of performance metrcs. From the Hadoop job logs, performance feature vectors, such as map tme, reduce tme, total executon tme, map output bytes, HDFS bytes wrtten, and locally wrtten bytes, could be extracted. These results convncngly demonstrate the effectveness of ths approach whch can also be used for our mechansm of MapReduce schedulng optmzaton. System Model The system model descrbes the data store and computng cluster that jobs could be assgned to the cluster ncludes machnes arranged n a general treeshaped swtched network. Reschedule When a task cannot be completed due to a dsk falure, processor falure, or other problems, the job may never get complete. A tme out Tout could be set for ths stuaton and the non-completed tasks could be rescheduled n the next calculaton. Other than falures, when new jobs arrve, the new jobs also need to be scheduled wth uncompleted tasks. Objectve Functon The objectve functon for our algorthm s the latest completon tme of the task schedule, referred as Makespan. The Makespan s calculated n objectve functon. Where t represents the tme that processor wll have fnshed the prevously assgned jobs and E[t][] s the predcted executon tme that task t s processed on processor. Schedulng Optmzer A master processor unt n cloud, collectng all tasks, wll take charge of dspatchng them to other process unts. Each process unt has ts own dspatch queue (DQ). The master unt communcates wth other process unts through these dspatch queues. Ths organzaton ensures that processor unts always fnd some tasks n the dspatch queue when t fnshes the executon of ther current task. The master unt works n parallel wth other unts, schedulng the newly arrved tasks, and perodcally updatng the dspatch queues. Tasks are sorted ascendng by the value of deadlnes. Reasons to choose GA as an optmzaton algorthm s smplcty of operaton and power of effect. It s sutable to some NP-hard problems. o Fuzzfcaton The Fuzzfcaton comprses the process of transformng values nto grades of membershp for lngustc terms of fuzzy sets. The membershp functon s used to assocate a grade to each lngustc term. o Genome representaton We use drect representaton. For job schedulng problem, a drect representaton s obtaned as follows. Feasble solutons are encoded n a vector, called schedule, of sze n tasks, where the numbers ndcates the slot where task s assgned by the schedule. Thus, the values of ths vector are natural numbers ncluded n the range [1; m slots]. o Genetc Operators In the ntalzaton process, the slot number (node number) whch has the data for the task s chosen as the ntal schedule. If there are several choces for one task, the slot number s randomly selected among these slots. Algorthm I. Get new tasks to be scheduled. The tasks to be scheduled nclude the uncompleted task and new jobs. But f jobs arrve n dynamcally and make too many jobs watng to be assgned at one tme, the sldng-wndow technque could be used as an opton. The wndow sze s fxed. Tasks fall nto the sldng wndow wll get scheduled. II. Generatng E matrx for the job Usng KCCA technque to predct the executon tme of any ndvdual task assgned to every node. III. Get the current state of the system. IV. Fuzzfcaton of all above parameter to get optmzed task schedule. ISSN: 0975 6760 NOV 12 TO OCT 13 VOLUME 02, ISSUE 02 Page 300
V. The Fuzzfy parameter Map n GA to get optmzed. a. Generate an ntal populaton of chromosomes randomly. b. Evaluate the ftness of each chromosome n the populaton. Evaluate P accordng to nformaton n E c. Create a new populaton by repeatng the followng steps untl the new populaton s complete, Selecton: Select two parent chromosomes from a populaton accordng to ther ftness. (The better the ftness, the hgher s the chance for gettng selected). Crossover Wth a crossover probablty, do cross over operatons on the parents to form new offsprng. If no crossover s performed, offsprng s the exact copy of the parents. Mutaton wth a mutaton probablty, mutate new offsprng at each locus (Poston n chromosome) Acceptance Place the new offsprng n the new populaton. d. Usng the newly generated populaton for a further sum of the algorthm. e. If the test condton s satsfed, stop and return the best soluton n the current populaton. f. Repeat Step c untl the target s met. VI. Fnally obtan the optmal soluton. B. Task Schedulng Optmzaton n Cloud Computng Based on Heurstc Algorthm In ths paper n order to mnmze the cost of the processng a model for task schedulng and propose a partcle swarm optmzaton (PSO) algorthm whch s based on small poston value rule s formulated. By vrtue of comparng PSO algorthm wth the PSO algorthm embedded n crossover and mutaton and n the local research, the experment results show the PSO algorthm not only converges faster but also runs faster than the other two algorthms n a large scale. The experment results prove that the PSO algorthm s more sutable to cloud computng. [14] The PSO Algorthm The algorthm begns wth k random partcle vector and each partcle s n n dmensons. Every partcle vector s a canddate soluton of the underlyng problem. The partcles are the task to be assgned and the dmensons of the partcle are the number of the specal tasks n a workflow. Then, each partcle moves by the drecton on the pbest and gbest untl the maxmal number of teratons. When the algorthm executes over, the gbest and ftness value are the correspondng task schedulng and the mnmal cost of the optmal strategy. PSO algorthm 1. Intalze partcle poston vector and velocty vector randomly. The vector's dmenson equal to the sze of the specal tasks. 2. Convert the contnuous poston vector (x k = [x 1, x 2,, x n ]) to dscrete vector (s k = [s 1, s 2,, s n ]) n lght of SPV rule. Then, transform the (s k = [s 1, s 2,,s n ]) to processor's vector (p k = [p 1, p 2,, p n ]). 3. If one partcle's ftness value s better than current, settng current value replace prevous pbest and as the new pbest. 4. Selectng the best partcle from the entre partcle as the gbest. 5. For all partcles update ther poston and velocty. 6. If reachng to the maxmum teraton or gettng the deal result stops, otherwse repeatng from Step 2. C. Comparson of load balancng algorthms n a Cloud Here the jobs are submtted by the clents to the computng system. As the submtted jobs arrve to the cloud they are queued n the stack. The cloud manager estmates the job sze and checks for the avalablty of the vrtual machne and also the capacty of the vrtual machne. Once the job sze and the avalable resource (vrtual machne) sze match, the job scheduler mmedately allocates the dentfed resource to the job n queue. The mpact of the ESCE algorthm s that there s an mprovement n response tme and the processng tme. The jobs are equally spread, the complete computng system s load balanced and no vrtual machnes are underutlzed. Due to ths advantage, there s reducton n the vrtual machne cost and the data transfer cost. [13] ESCE LOAD ALGORITHM ACTIVE VM LOAD BALANCER [START] Step1:- fnd the next avalable VM Step2:-check for all current allocaton count s less than max length of VM lst allocate the VM Step3:- f avalable VM s not allocated create a new one n Step 4:- count the actve load on each VM Step5:- return the d of those VM whch s havng least load [END] D. A Dynamc Optmzaton Algorthm for Task Schedulng n Cloud Envronment Ths paper proposes a schedulng algorthm whch addresses these major challenges of task ISSN: 0975 6760 NOV 12 TO OCT 13 VOLUME 02, ISSUE 02 Page 301
schedulng n cloud. The ncomng tasks are grouped on the bass of task requrement lke mnmum executon tme or mnmum cost and prortzed. Resource selecton s done on the bass of task constrants usng a greedy approach. [10] Algorthm An optmum schedulng algorthm s proposed and Implemented n ths secton. The proposed algorthm works as follows 1. Incomng tasks to the broker are grouped on the bass of ther type deadlne constraned or low cost requrement. 2. After ntal groupng they are prortzed accordng to deadlne or proft. Ths s requred because the tasks wth shorter deadlne need to be scheduled frst and smlarly the tasks resultng n more proft should be scheduled on lost cost machnes. Thus, the prortzng parameter s dfferent based on the nature or type of task. 3. A. For each prortzed task n deadlne constraned group ) Turnaround tme at each resource s calculated takng followng parameters nto account. Watng tme Task length Processng Power of vrtual machne ) The vrtual machne wth mnmum turnaround tme that s capable to execute the task s selected and task s scheduled for executon on that machne. ) Watng tme and resource capacty of selected machnes are updated accordngly. B. For cost based group ) Vrtual Machnes are selected on the bass of processng power of machne and ts cost ) For each vrtual machne cloudlets from the group are scheduled tll the resource capacty s permtted. ) Resource capacty and watng tme are updated accordngly. 3. CONCLUSION Cloud task schedulng algorthms are desgned to support large scale applcatons and to support cloud nfrastructure. Varous algorthms are proposed n order to acheve proper schedulng n cloud computng envronment through the better resource allocaton and optmzaton. Better performance and effcency can be acheved by applyng varous dfferent proposed technques dscussed above. In ths paper, analyss and overvew of varous dfferent task schedulng algorthms s carred out whch uses dfferent parameters lke tme, cost, and make span, speed, scalablty, resource allocaton, schedulng and sophstcated tools. 4. REFERENCES [1] Cloud computng resources. http://www.cloud9s.net/cloudcomputngresources.ht ml [2] http://www.thecloudcomputng.org/2011 http://salsahpc.ndana.edu/cloudcom2010/papers.ht ml [3] http://www.tworld.com/nternet/69141/5-coolcloud-computng-research-projects [4] http://en.wkpeda.org/wk/cloud_computng [5] http://www.thecloudcomputng.org/ 2012/hstory.html [6] http://www.wheresdoc.co [7] Torry hars CLOUD COMPUTING An Overvew [8] V. Venkatesa Kumar and K. Dnesh Job Schedulng Usng Fuzzy Neural Network Algorthm n Cloud Envronment Bonfrng Internatonal Journal of Man Machne Interface, Vol. 2, No. 1, March 2012 [9] Monka Choudhary and Sateesh Kumar Peddoju A Dynamc Optmzaton Algorthm for Task Schedulng n Cloud Envronment Internatonal Journal of Engneerng Research and Applcatons (IJERA), Vol. 2, Issue 3, May-Jun 2012, pp.2564-2568 [10] Sudhr Sngh Performance Optmzaton n Gang Schedulng In Cloud Computng IOSR Journal of Computer Engneerng (IOSRJCE), ISSN: 2278-0661 Vol. 2, Issue 4, July-Aug. 2012, pp. 49-52 [11] Sandeep Tayal Tasks Schedulng optmzaton for the Cloud Computng Systems Internatonal Journal Of Advanced Engneerng Scences And Technologes, Vol No. 5, Issue No. 2, pp. 111 115 [12] Jaspreet kaur Comparson of load balancng algorthms n a Cloud Internatonal Journal of Engneerng Research and Applcatons, Vol. 2, Issue 3, May-Jun 2012, pp.1169-1173 [13] Lzheng Guo, Shuguang Zhao, Shgen Shen, Changyuan Jang Task Schedulng Optmzaton n Cloud Computng Based on Heurstc Algorthm, JOURNAL OF NETWORKS, VOL. 7, NO. 3, MARCH 2012 ISSN: 0975 6760 NOV 12 TO OCT 13 VOLUME 02, ISSUE 02 Page 302