Send Odes fo Repnts to epnts@benthascence.ae 334 The Open Cybenetcs & Systecs Jounal, 205, 9, 334-340 Open Access Reseach on Cloud Coputng Load Balancng Based on Vtual Machne Mgaton Lu Kun,*, Xu Gaochao and Chen Jngxa 2 College of Copute Scence and Technology, Jln Unvesty, Jln, Changchun, 3002, P.R. Chna; 2 College of Infoaton Scence & Technology, Beng Unvesty of Checal Technology, Beng, 0029, P.R. Chna Abstact: Ths atcle s desgned to solve the poble of syste load balance and low effcency ntoduced by assve paallel tasks unnng on soe heavy nodes n the Cloud Coputng envonent. Ths study poposes a load balancng algoth based on vtual achne gaton whch ncludes load balancng achtectue, load gatheng, load ontong, load foecastng, gaton tgge te, souce achne selectng and etc. Expeental esults based on the CloudS tool eveal that the algoth can guaantee load balancng and pove the pefoance of the syste. Keywods: Cloud coputng, load balancng, gaton, vtual achne.. INTRODUCTION Cloud coputng eeges as a new coputng and busness odel. Uses can use the cloud coputng platfo fo tepoay opeaton nstead of puchasng expensve pesonal coputes, seves o physcal devces. Entepses also only need to buy cheap, easly extended cloud platfo nstead of buldng the huge vtual seve clustes. Clustes load wll becoe unbalancng fo the uses vaous o dynac deand and seves heteogeneous esouces. Ths poble dectly affects the clustes esouces effcency and bngs about the waste of a lage nube of cloud esouces. So how to pove the cloud data cente s pefoance by applyng sutable schedulng algoth s an potant poble of the Cloud Coputng. Whethe the pefoance of the load balancng algoth s good o bad s vey potant to the whole cluste []. Theefoe, the evaluaton of a load balancng algoth ust pay attenton to both the algoth tself and the occason of use of the algoth. Scholas at hoe and aboad have done a lot of eseach on the load balancng poble of cloud coputng. In pape [2], the autho balanced each node accodng to the next oent load whch s coputed by the pedcted load based on Eo Back Popagaton Tanng(BP) algoth. In pape [3], the autho udged the next oent load on the bass of pedcted load coputed by the sulated annealng algoth. In pape [4], the autho poposed a weghted sequence dynac algoth. In pape [5], the autho dscussed the addtonal oveload poduced by the balancng pocedue and poposed a patcle sze foula to avod addtonal ovehead. Bono [6] adusted each seve s load afte pedctng the load at the next oent. Ths pape pesents a dynac load balancng algoth based on esouce schedulng, whch anly ncludes the followng thee aspects of wok: poposng a load balancng achtectue, eseachng on load balancng tgge te, and eseachng on the selecton of load balancng souce achne. 2. DEFINITION AND ARCHITECTURE FOR LOAD BALANCING 2.. Defnton Defnton : The nube of vtual achne on node s n. Defnton 2: The CPU utlzaton ate of node s CUR, CUR MUR, n C n. C s the CPU utlzaton ate of vtual. Defnton 3: The eoy utlzaton ate of node s MUR n M MT. M s the eoy use of vtual. MT s the total eoy aount of node. Defnton 4: The bandwdth utlzaton ate of node s NBUR, NBUR n NB NBT. NB s the bandwdth use of vtual. NBT s the total bandwdth aount of node. Defnton 5: The aveage CPU utlzaton of the cluste CUR s CA: CA. Lette denotes the nube of nodes n the cluste. 874-0X/5 205 Bentha Open
Reseach on Cloud Coputng Load Balancng The Open Cybenetcs & Systecs Jounal, 205, Volue 9 335 Defnton 6: The aveage eoy utlzaton of the clus- MUR te s MA: MA. Lette denotes the nube of nodes n the cluste. Defnton 7: The aveage bandwdth utlzaton of the NBUR cluste s NBA: NBA. Lette denotes the nube of nodes n the cluste. Defnton 8: Hgh theshold(h th ) s a value. If the load of the node exceeds ths value, the node s called oveload node. Defnton 9: Adaptve theshold( th ) s a value whch udge the oveload node cobned wth H th. Defnton 0: Oveload Matx s expessed as hgh. CUR MUR NBUR RL NUM CUR2 MUR2 NBUR2 RL2 NUM 2 hgh CURn MURn NBURn RLn NUM n Values of each ow ae the paaete values of each node. The fst thee coluns epesent CUR, MUR and NBUR. The fouth colun RL of 0 ndcates that CUR value s hgh, whle the value of ndcates that MUR value s hgh, and the value of 2 ndcates NBUR value s hgh. The last colun NUM epesent the node's nube. 2.2. Achtectue A dynac gaton algoth ust consde [7] thee pobles: when the dynac gaton s executed, whch vtual achne wll be gated, and whee does the vtual achne gate to. Ths study popose a load balancng odel achtectue daga, as shown n Fg. (). Fg. (). Achtectue daga. Gatheng odel gathes the load of each node and stoes the load data to database. Montong odel decdes whethe tgge the gaton. Foecastng odel foecasts futue load value and the gaton wll not be tggeed when the load value s a nstantaneous peak value. Selectng ode ncludes souce achne selectng and taget achne selectng. Mgaton odel executes vtual achne gaton. 3. LOAD BALANCING ALGORITHM 3.. Gatheng Model and Montong Model The cental node collects CUR, MUR, NBUR of each node egulaly. If the nteval te s too shot, the cental node wll be too busy, and data tanssson wll occupy the bandwdth. If the nteval te s too long, the heavy load nodes wll not be handled tely whle those low load nodes wll be handled. Most papes set the nteval te as 0 to 20 seconds. In ths pape, the cental ead each node evey 5 seconds and the node whose load value fluctuate oe than 0% send ts load to the cental. Montong odel analyzes the load value eceved fo the gatheng odule and detenes whch node wll be balanced. Effcent peequste fo vtual achne gaton s to foulate a easonable tgge condtons (theshold). We use paaetes Hth and th to detene whethe a node s oveload. The algoth s shown n Fg. (2). 3.2. Foecastng Model Vtual achne wll be gated when the load exceeds the theshold n tadtonal load balancng algoths. In fact ost nodes have nstantaneous load peak value and the load value wll etun to noal afte ths peod. Fo tadtonal algoth, the nstantaneous peak value wll tgge load balancng, whch wll cause unnecessay gaton. At pesent, soe scholas have eseached on the pedcton on physcal hosts and found that load changes has self slaty and long-ange dependence [8]. In ths pape, the foecastng odule based on exponental soothng ethod s desgned to pedct the load value at the next oent, and then decde whethe to tgge gaton. ' ' yt+ * yt + ( ) * yt s the foula of exponental ' soothng ethod. y t+ s the pedcton value at oent y ' t+, t s the actual value at oent t, y t s the pedcton value at oent t and s the sooth coeffcent. [ 0,]. When a node s load exceeds theshold n ontong odel, the fst d-(d s an uncetan nube) oent data of ths node ae passed to foecastng odel to get p(p s an uncetan nube) foecastng values. If s(s s an uncetan nube) values exceed theshold, the load balancng wll be tggeed. 3.3 Souce Machne Selecton Most authos use weght to calculate the total value, such as w*cur+w2*mur+w3*nbur, w+w2+w3. In ths foula, the factos that have the heavy weght wll oe affect the fnal total value.
336 The Open Cybenetcs & Systecs Jounal, 205, Volue 9 Kun et al. _ th ; _ th _ th _th Fg. (2). Flow chat of coputng atx hgh. Fo exaple, thee ae two nodes whose CUR, MUR, NBUR ae espectvely <0.9,0.3,0.2> and <0.5,0.5,0.2>. It s obvously that the fst node has geat CPU load and the second load value sn t hgh. So the fst node ust be fst to be balanced. If the thee weghts value ae w0.2, w20.5, w30.3, total load of the two nodes ae 0.39, 0.6, so the second node wll be balanced. Ths shows that the choce of dffeent weght values wll affect the balancng algoth. In ths pape, the nfoaton entopy s used to detene the weght value. The basc deas of entopy ethod s to detene the weght obectvely accodng to the vaablty. Geneally speakng, the nfoaton entopy value beng salle shows that the degee of vaablty s uch geate, the aount of nfoaton povded s uch oe, and so the weght s geat. On the contay, the nfoaton entopy value beng uch geate shows that the degee of vaablty s uch salle, the aount of nfoaton povded s uch salle, and so the weght s sall. Total load coputng ethods ae gven below: () The fst thee coluns of oveload atx hgh ae chosen fo ontong odel to fo a new atx D, CUR MUR NBUR CUR2 MUR2 NBUR2. D CURn MURn NBURn (2) Matx D s standadzed to R, 2 R n CUR, MUR, NBUR 2 n. 3 n CUR MUR NBUR n (3) The entopy s coputed as, 2, 3. (4) F s set as F -E. w (5) Each weght s set as F 2, 2 F + F2 + F3 w F 3. 3 F + F2 + F3 E 2 22 n2 3 23, n 3 n ln, ln n w F, F + F2 + F3 (6) Total load of node s Load w CUR + w * MUR w * NBUR. * 2 + 3 (7) Accodng to the Load n descendng ode, a queue s foed as Q{q, q 2,, q n }. Each eleent has two values as node nube and flag RL, that s q (RL k, NUM k ). (8) The fst achne n the queue Q s selected as the souce achne. The vtual whose CUR /MUR s the bggest on the achne s selected as the gated vtual because eoy gaton s the ost dffcult.
Reseach on Cloud Coputng Load Balancng The Open Cybenetcs & Systecs Jounal, 205, Volue 9 337 Table. Expeental confguaton table. Paaete Nae Unt Value Physcal achne nube 50 CPU pocessng capacty of each physcal achne MIPS {000,800,2600,3000} Physcal achne eoy value G { 2 4 8} Physcal achne bandwdth value Mb/S {500,700,000} Vtual achne nube 200 CPU pocessng capacty of each vtual achne MIPS {200,500,000,500,2500} Vtual achne eoy value G {0.5,, 2, 3} Vtual achne bandwdth value Mb/S {00, 200, 500} 4. EXPERIMENTS We choose CloudS [9, 0] whch developed by Ruyya et al. fo Melboune Unvesty as the sulaton tool. The expeents selected 50 heteogeneous physcal hosts, and 3 to 5 vtual achnes wee espectvely aanged on each host. The confguaton as shown n Table. In ode to vefy whethe the foecast odule s effectve, we have done any expeents on the CPU utlzaton ate. It s sla as eoy utlzaton ate and bandwdth utlzaton ate. The paaetes value ae shown n Table 2. The value s set as 0.7. Table 2. Paaetes value table. Paaetes Nae Value H th 70% th 0% Focast peod(p) 5 Judgng nube(s) 4 Dataset Sze(d) 50 The expeental esults ae shown n Fg. (3), whee vetcal axs epesents the CPU utlzaton ate, hozontal axs epesents the nube of ontong data ponts, the dotted lne wth a ccula ndcates the actual value of CPU utlzaton ate, the sold lne wth a sta ndcates the pedctve values, and H th value 0.7. () The actual values exceed theshold at the oent t4 and t35. The foecastng value doesn t exceed theshold, So the balancng wll not tggeed. (2) The actual values exceed theshold at oent t6 to t20, and the foecastng values exceed theshold at oent t7 to t2. Thee ae fou values exceed theshold n the fve values, so the balancng wll be tggeed. Afte balancng, the load value dops at oent t2. The balancng wll be tggeed when the actual values exceed theshold f not steeped n foecast algoth, fo exaple, at oent t4, t27, t28 o t35. Usng the foecastng algoth effectvely avods the balancng be tggeed by nstantaneous peak value. Fo Fgs. (4-9), we can fnd the dstbuton of CUR, MUR and NBUR s vey uneven. Afte seveal ounds of load balancng, the thee values ae obvously be balanced. Fg. (3). Chat of foecastng.
338 The Open Cybenetcs & Systecs Jounal, 205, Volue 9 Kun et al. Fg. (4). Chat of CUR befoe balancng. Fg. (8). Chat of NBUR befoe balancng. Fg. (5). Chat of CUR afte balancng. Fg. (6). Chat of MUR befoe balancng. Fg. (7). Chat of MUR afte balancng. Fg. (9). Chat of NBUR afte balancng. Lots of expeents on the algoth n y pape and the algoth of cetan weght(heenafte efeed to as the algoth A) have been done. In algoth A, the load foula s defned as Load w * CUR + w2 * MUR + w w 3 * NBUR and the values of w, w 2, w 3 ae 0.7,0.5, 0.5, naely the CPU utlzaton ate accounts fo the axu weght. In Fg. (0) to 2, hozontal axs epesents each physcal achne, and vetcal axs epesents the utlzaton ate devaton whch s the dffeence between the actual value and the aveage value. Sall devaton suggests the syste s oe balance. Fo Fgs. (0-2), CUR of algoth A s slghtly bette than ths pape s algoth, whle MUR and NBUR of ths pape s algoth ae oe bette than algoth A s. The weght of CPU utlzaton ate s hghe than the othe two n algoth A, so the esult on CPU utlzaton s bette n algoth A. Although ths pape s algoth s wose than algoth A on CUR expeent, ths pape s algoth akes the oveall pefoance on CUR, MUR and NBUR uch bette. 5. CONCLUSION Ths study poposed a vtual gaton algoth to balance the cluste s load ncludng gatheng odel, ontong odel, foecastng odel, selectng odel, gaton odel. A load udgng atx based on hgh theshold and adaptve theshold was desgned. The exponental soothng ethod was ntoduced to avod nstantaneous peak value,
Reseach on Cloud Coputng Load Balancng The Open Cybenetcs & Systecs Jounal, 205, Volue 9 339 Fg. (0). Devaton of NBUR. Fg. (). Devaton of CUR. Fg. (2). Devaton of MUR. and the souce achne selectng algoth based on nfoaton entopy was poposed to solve the poble of uneasonable weght value. Fnally, we caed out ultple expeents and poved he effectveness of the algoth n ths study. CONFLICT OF INTEREST The authos conf that ths atcle content has no conflct of nteest. ACKNOWLEDGEMENTS The study s subsdzed by Fundng Poect of Copetence Developent Poga fo Beng VET Teaches and Beng Cty Boad of Educaton Scence and Technology Poect (KM20247008, KM2054700) and Beng Hghe Educaton YoungElte Teache Poect (YETP767). REFERENCES [] S. Zhu, A new way fo dynac load balancng algoth desgn, Copute Engneeng and Desgn, vol. 6, no. 3, pp. 25-30, 994.
340 The Open Cybenetcs & Systecs Jounal, 205, Volue 9 Kun et al. [2] S. Zhang, Dynac BP Algoth Based on Load Balance Foecast, Henan:Henan Unvesty, 2009. [3] H. Y, Reseach on Netwok Load Balancng Algoth Based on Sulated Annealng Genetc Algoth, WuHan: WuHan Unvesty of Technology, 203. [4] Y. Hu, and Y. C. Ou, A esouce load balacng ethod to educe the enegy cosupton n the cloud envonent, Copute Engneeng, vol. 38, no. 5, pp. 53-55, 202. [5] D. L, and H. Sh, Study on geneal odel of load balancng schedulng pobles, Copute Engneeng and Applcaton, vol. 43, no. 8, pp. 2-25, 2007. [6] R. Bono, P. J. Fleng, and P. Stenbeg, An adaptve on the based queue ule fo load shang on dstbuted copute syste, In: Poceedngs of the 28 th Confeence On Decson and Contol, Tapa, Fla, 989, pp. 2554-2559. [7] K. Yang, J. Gu, T. Zhao, and G. Sun, An optzed contol stategy fo load balancng based on lve gaton of vtual achnef, 20 6 th Annual Chna Gd Confeence, 203, pp. 4-46. [8] P. A. Dnda, The statstcal popetes of host loa, Scentfc Pogang, vol. 7, no. 3-4, pp. 2-229, 999. [9] R.N. Calheos, R. Ranans, and A. Beloglazov, CloudS: a toolkt fo odelng and sulaton of cloud coputng envonents and evaluaton of esouce povsonng algoths, Softwae: Pactce and Expeence, vol. 4, no., pp. 23-50, 20. [0] SCloud Platfo, http://scloud.co/ Receved: June 0, 205 Revsed: July 29, 205 Accepted: August 5, 205 Kun et al.; Lcensee Bentha Open. Ths s an open access atcle lcensed unde the tes of the Ceatve Coons Attbuton Non-Coecal Lcense (http://ceatvecoons.og/- lcenses/by-nc/3.0/) whch pets unestcted, non-coecal use, dstbuton and epoducton n any edu, povded the wok s popely cted.