您的当前位置:首页正文

Virtual resource monitoring in cloud computing

2021-07-11 来源:易榕旅网
J Shanghai Univ(Engl Ed),2011,15(5):381—385 Digital Object Identiifer(DOI):10.1007/s11741—011—0755一l Virtual resource monitoring in cloud computing HAN Fang—fang(韩芳芳) ,PENG Jun-jie(彭俊杰) ,ZHANG wu(张武) ,LI Qing(李LI Jian—dun(李建敦) ,JIANG Qin—long(江钦龙) ,YUAN Qin(袁勤) 1.School of Computer Engineering and Science,Shanghai Universityj Shanghai 200072,P.R.China 2.Key Laboratory of Computer System and Architecture,Institute of Computing Technology,Chinese Academy of Science 青) , Beijing 100190,P.R.China ⑥Shanghai University and Springer—Verlag Berlin Heidelberg 2011 Abstract Cloud computing is a new computing mode1.The resource monitoring tools are immature compared to traditional distributed computing and grid computing.In order to better monitor the virtual resource in cloud computing.a periodically and event driven push fPEP)monitoring model is proposed.Taking advantage of the push and event—driven mechanism, the model can provide comparatively adequate information about usage and status of the resources. It can simplify the communication between Master and、^ rk Nodes without missing the important issues happened during the push interva1. Besides.we develop“mon”to make up for the deficiency of Libvirt in monitoring of virtual CPU and memory. Keywords virtual resource monitoring,cloud computing,virtualization Introduction Cloud computing is a new information technology trend that moves computing and data away from desl(_ tops and portable PCs into large data centers.The basic principle of cloud computing is to deliver applications as services over the Internet as well as infrastructure[ . Virtualization technology is one of the keY bases in computing pricing and is the premise of other techniques in cloud computing as load balancing and task schedul— lng. However,since cloud computing is still under the budding stage while the virtual machine technology is still under development.the monitoring techniques or clfoud computing based on virtual machine resource monitoring are for from mature.From the upper per— cloud computing,and it provides cloud computing a ba- sic underlying platform[ .First introduced in 1960’s. spective,there are sound monitoring systems designed orf grid computing such as Ganglia[4]Nagios[引net— ,,the virtualization technology booms as the fast devel— work weather service fNWS)【6l and these are mainly or distributed and grifd computing systems.They are structured to monitor physical resources in above men- tioned systems and not capable of monitoring the highly opment of computer hardware.Modern computers are sufficiently powerfu1 to use virtualization to present the illusion of many smaller virtual machines fVMs),each running a separate operating system instance.This has virtualized cloud computing resources.Another prob— lem is about the feature of cloud computing with no led to a resurgence of in[crest in VM technologyt ̄J. This paper mainly focuses on the virtual resource monitoring in cloud computing.Cloud computing is not adequate monitoring data and no accurate pricing,and this is unacceptable for both the provider and the user. Then from the virtual machine perspective,the moni— iust a kind of new technology,and its main concept is to provide users with service through“pay—as—you—gO” manner.Thus how to get the accurate monitoring data of the resources users consumed is really of importance. toring data submitted from the virtual machine monitor (VMM)are far less than accurate and su佑cient. In this paper,we explore a new resource monitoring The resource monitoring leads to an effective feedback model from the platform level to the virtualization leve1. 0n the platform level,push and event—driven schema are used to optimize monitoring result.On the virtu— for the system to the discovery reallocation or adiust— ment of resources accompanied with the variation of re. source usage rate.It provides a rational basis for cloud Received Apr.23,2011;Revised May 26,2011 Project supported by the Shanghai Leading Academic Discipline Project(Grant No.J50103),the Ph D Programs Foundation of Ministry of Education of China(Grant No.200802800007),the Key Laboratory of Computer System and Architecture fInstitute of Computing Technology,Chinese Academy of Sciences),and the Innovation Project of Shanghai Municipal Education Commission(Grant No.1 lYZ09) Corresponding author PENG Jun-jie,Ph D,Prof,E mail:jjie.peng@shu.edu.an J Shanghai Univ(Engl Ed),2011,15(5):381—385 However we can introduce this concept into cloud CON— puting with modification. As to cloud computing,in the pull mode,the work nodes are in a passive position.They need the master node to tell them the time that they have to send the monitoring data.While in push mode,the work nodes will send the monitoring information to the master node without being foretold,they are active in this situation. There are also two modes for estimating when to push or pull:periodically or by event—driven.Thus combined the push and pull with these two modes,four kinds of combination are available:periodical pull,pe— riodica1 push.event—driven pull,and event—driven push. Each has its adva ntages and disadvantages. In periodical pull mode the choosing of time interval matters a lot:too frequent pull aggravates network load; too long interval cannot assure the freshness of the data, and may lose the capture of information during the pull message response time.Meanwhile,the transmission of pull request consumes time and bandwidth. In periodical push mode,the same problems still bother except for the last one. However when some emergency happens.it would not send it immediately to the master node 妇a£is what we do not expect to see. In event—driven pull mode.obviously it has worse co— herency.There is difficulty in the master node to iden— tify the“event”. Lastly,in event—driven push mode.it has a quick re— sponse to the“event”.but the real—time coherency can not be guaranteed. In view of the all above in our model,we adopt pe— riodical push and event—driven push modes both which are complementary to each other.Here,we define this model as periodically and event-driven push(PEP).By means of it.we can assure the coherency as well as the event response time. Figure 2 is the architecture of our virtual resource monitoring mode1.In the master node,data collector is responsible for the collection of overall monitoring infor. mation.and GUI module displays the monitoring situ. ation to users. In each work node within the VMM,we choose the widely used Libvirt tool to monitor the status of vir— tual resources.Meanwhile,as the defects of Libvirt we have mentioned above.we developed“mon’’to make up ofr getting more comprehensive monitoring information. The collections of monitoring information are transmit. ted to data collector of work node.The push module transfers them to master node under two circumstances: It pushes periodically according to the time interval we set before;in case of emergencies,by means of the event trigger module,it can also push the urging information to the master node immediately. Fig.2 Architecture of the monitoring model 2.1 Algorithm of PEP model As can be seen from Fig.3,there are two key points in this algorithm:the value of push—interval and the definition of“event”.The f0rmer influenees the moni— toring status directly:a long push—interval may result in the“freshness”of data.while a short push_interval may aggravate the data transmission pressure. Reference f1 21 discussed the value of interval;through experiment they found 1 s is an appropriate value.Thus,as many researchers have chosen.we also set the push—interval as】s. Fig.3 PEP algorithm The purpose of“event—trigger”is to provide mas— ter node with an alerting mechanism.Here we choose event as a threshold which represents the load condition of CPU and memory.We define nodes’resource load vector as R(h)=( pu(^), 。 ( )) (1) Here R(h)is the node’s resource load vector, pu( ),Umem( )represent the resource utilization rate of CPU and memory respectively. The threshold of event trigger can be set manually. J 'el(!Xalnplc,if tile CPU utilization rate is over 90%or t 1le uleulory utilization rate is over 90%.or both of them i re I(?SS thaii 5%.it will trigger the event.and send this info,’nlatioIl to ntaster node regardless of the push in— t(、I va1.Thel1 the monitor in master node wil1 instantly know t he using status of the resource.No matter the w()rk no(]o is too busy or too idle.monitor will feedback t}lese inf0I’11lation to other parts like load balancing to optinfize ttle load. 2.2 Collect ion of virt ual resource Within VMM except for the widely used Libvirt to cat)ttire the monitoring information,we also use seine t o(:hniques to make up for the inefiectiveness of Libvirt in virtual resonrce monitoring.、^,e use Xen as the VM iii tlifs pat)er. Through LibviI’t we can get the CPU utilization I÷tt( .1llenlory allocation status,number of VCPU.CPU sl at1is ( tc. h】order to get the VCPU utilization rate.we COnl— pule it as follows: Ⅵ. (£): ×100% \vhel’c CT depicts the VCPU running time up to now C"1 1 the VCPU rlinning tilne t seconds before. Fig.4 Collection of virtual resource McIIlO1 Y utilization rate:each domain pursues their {)w11 1ltelnorv nsin ̄status frOIn the/proc file system. ;1n(1 sent t;heln to Dolnain().As to the network and I/O loa‘1 Donla in()(;all acquire from its back—end. Ftle data collect;or acquires both static and dynamic lll{ )rmarion.It gets static data such as CPU number, 1 H)( .freqlmncy.physical nlelIlOl’Y,etc.from the/proc ifle svstel11.These data i8 recorded the nloment the start of system a nd only need record once In addition it gets the dynamic data such as the CPU utilization rate, inelnory utilization rate,CPU running status,VCPU n1lnlber.etc..base(i on our PEP model from Libvirt in一 rf}lce aiId“Inoi1”ruedule. J Shanghai Univ(Engl Ed).201 1.15(5j:381—385 3 Model analysis The overall design goals of'ally system are scalabil— ity.efifciency.and accuracy.Ⅵ analyze the propose(1 monitoring model frOUl these perspectives nOWr Fr0in perspective of the scalabi一 鼙 lity our PEP mode1 is inherently scalable.Nowadays,the structure ot’cloud computing is heterogeneous. Business giants construct their own cloud computin ̄illfrastructures.There a r(] nunlerous categories of cloud computing:public,pri— rate.hybrid etc.However.the fllndamental technol~ ogy of cloud computing is virtualization.Om model roots in the virtualizatioI1 leve1.screening the hetero— geneity of systems.We monitor tile status not solely under specific configuration but can apply it to general platforms. Conling to the emciency.PEP model 1)alances the workload and freshness of data.The appropriate choos— ing of push interval can guarantee the obtaining of。 monitoring data without ireposing t0()n]uch burden to the system.The event trigger mechanism using siiI1一 ply threshold control method can instantly react with the monitor.With the consideration 0f ef}iciency PEP adopts simply threshold control since it will not occnpy the computing resource of the CPU. To obtain inore adequate monitoring data is ell( ̄t)j the basic roles of any monitoring systems.Libvirt gallis an upper hand in the scalability since it can be use(1 under Xen.KVM or QEMU.Howevei’ the monitoring data it gathers are not adequate enough especially tlLt、 virtual resource using status.Xel1.itself.provides SO111t tools for the collection of monitoring data.such as tll(、 xrn top.Xenstore.XenMon[13j.These tools work on ttlc lower layer.can get more accurate data.How-evei’the problem is that it can only be used in Xen.1et aloll( the other kinds of cloud computing architectures.PEP model llses the Libvirt interface.so it carries f(】rwar(1 the scalability.Meanwhile.it also nlakes up the ineapa— bility of Libvirt in virtual resource usage monitoring. Future work includes evaluation of tile proposed SO— lutions.T0 evaluate the monitoring proposal,a proto— type is being developed as a proof'of concept and it s performance and usability wil1 be tested. 4 Conclusions and fut ure work The flourishing cloud conlputing concept has a lot of new problems to be solved.The virtual resoui’ce lUO1li— toring i8 one of the challenging one.Ill order to Inonitol’ the virtual resource monitoring in cloud compllting bel— ter and provide comparatively sound monitorillg infbl’一 mation to users and other colnponents in cloud con1put— ink we put forward PEP monitoring mode1.Ttie nlodcl takes advantage of the push and event—driven inecha— nism.and simplify the colnnmnication between¨laStei’ J Shanghai Univ(Engl Ed),2011,15(5):381—385 and work nodes without missing the important issues happened during the push interva1.In addition we de— velop“mon”to make up for the deficiency of Libvirt in monitoring of virtua1 CPU and memory.Our model is able to capture the virtual resources dynamic usage sta— tus well,compared to other methods 1ike using Libvirt only.It does not take up too much computing resource and provide more adequate information. 0ur further aim is to be able to provide these infor— mation not only to users but also to the other significant components in cloud computing,such as load balancing, pricing,etc.Based on these monitoring data,we can develop easy interoperating ways to reallocate resources and ease the overload nodes. References KANG M,KANG D—I,YUN M,PARK G L,LEE J Design for run—time monitor on cloud computing[J] Communications in Computer and Information Science 2010,DOI:10.1007/978—3—624—16444—6—36. JONES M T.Cloud computing with Linux cloud pom— puting platforms and applications[EB/OL].(2008一 O9—10)[2011—05—26].http://www.ibm.com/developer— works/linux/libr—ary/1一cloud—computing. [3]BARHAM P,]-)RAGOVIE B,FRASER K,HAND S,HAR— RIS T,HO A,NEUGEBAUER R,PRATT I,WARFIELD A. Xen and the art of virtualization[c]//Proceedings of the nineteenth ACM symposium on operating systems principles.New York.USA.2003:164—177. [4]MASSIE M L,CHUN B N,CULLER D E.The gan— glia distributed monitoring system:Design,implemen— 385 ta ̄ion,and experience[J1.Parallel Computing,2004 30(7):817 840. [5]Nagios【EB/OL].(2011—01—29)[2011—03—24].http:// www.nagios.org. f61 POLADIAN V,GARLAN D,SHAW M,SATYANARAYANAN M,SCHMERL B R,PEDRO J.Leveraging resource pre- diction for anticipatory dynamic configuration『C]// Proceedings of the First International Conference on Self-Adaptive and Self-Organizing Systems,Washing- ton DC,USA.2007:214-223. [7]Virt—manager[EB/OL].(2011—02—03)【2011—04-18] http://www.virt—manager.org. [8]Libvirt:The virtualization API[EB/OL].(2010—12—11) [2011—04—09].http://libvirt.org. [9 KVM9](Kernel—based Virtual Machine)[EB/OL]. (2010—11—12)[2011—11—02].http://kvm.qumranet.com/ kvmwiki. 【10】BELLARD F.QEMU,a fsat and portable dynamic trans— lator[c]//Proceedings of Usenix Annual Technical Conference,Berkeley,CA,USA.2005:41—46. [1 1]ZANIKOLAS S,SAKELLARIOU R.A taxonomy of grid monitoring systems[J].Future Generation Computer Systems,2005,21(1):163—188. [12] YANG Gang,SUI Yu—lei.Adaptive approach to monitor resource for cloud computing platform[J]_Computer Engineering and Applications,2009,45(29):14—17(in Chinese). 【13】GUPTA D,GARDNER R,CHERKASOVA L.XenMon QoS monitoring and performance profiling tool[R] California,USA:Technical Report of HP Lab,2005. (Editor CHEN Hai—qing) 

因篇幅问题不能全部显示,请点此查看更多更全内容