《云计算平台虚拟机.docx》由会员分享，可在线阅读，更多相关《云计算平台虚拟机.docx（19页珍藏版）》请在三一文库上搜索。
1、云计算平台虚拟机 - 让每个人平等地提升自我 1 An Evaluation of KVM for Use in Cloud Computing,Clemson UniversityIn this paper we describe a virtual cluster based on the Kernel-based VirtualMachine(KVM as an alternative to VMWare and cally we show how the virtual cluster is built and tailored to t virtual technique prese
2、nted in this paper,known as the Virtual Organization Cluster Model,shows great potential for cloud our implementation, we used a minimalist installation of Slackware Linux12on14compute nodes,ensuring minimal host prototype Virtual Organization Cluster is composed of28virtual computes nodes,each runn
3、ing Condor,and of testing the prototype were encouraging,with the exception of network performance.Categories and Subject Descriptors: Computer Systems Organization:Performance of Systems Design studies;Performance attributes; Net-works:Distributed SystemsGeneral Terms:Design,Experimentation,Measure
4、ment,PerformanceAdditional Key Words and Phrases:Virtual Machines,KVM,High-Performance ComputingA current problem in scienti c computing is that the expertise needed to deploy and maintain a cluster remains scarce despite recent declines in hardware costs. Smaller research groups may not be able to
5、a ord to have their own clusters,and purchasing time on an existing cluster raises concerns such as vendor lock-in and data security.In this paper,we present a cloud computing model which de nes a minimum speci cation for a compute clusters sole purpose is to host other com-pute clusters through - 让
6、每个人平等地提升自我 this way,a Virtualization Service Provider (VSPcan sell compute power without having to directly maintain each end-users particular application.Similarly,Virtual Organizations(VOscan purchase compute power from VSPs without having to worry about hardware or software VO is free to develop
7、a model cluster locally,perhaps even on a personal workstation,test it,andAuthorsaddress:School of Computing,Clemson University,Clemson,SC29634-0974,USA.This material is based upon work supported under a National Science Foundation Graduate Re-search Fellowship.Permission to make digital/hard copy o
8、f all or part of this material without fee for personal or classroom use provided that the copies are not made or distributed for pro t or commercial advantage,the ACM copyright/server notice,the title of the publication,and its date appear,and notice is given that copying is by permission of the AC
9、M, copy otherwise,to republish, to post on servers,or to redistribute to lists requires prior speci c permission and/or a fee.c 20YY ACM1529-3785/20YY/0700-0001$ACM Transactions on Computational Logic,Month20YY,Pages1 0?.2 et al.2- 让每个人平等地提升自我3 (aType1(bType2deploy it to a VSPs hardware with reasona
10、ble assurances that the operatingenvironment will be fully compatible.We will rst provide a brief overview of virtualization technologies,followed by a description of our model virtual ,we will de ne the infrastructure for which an VSP would be ,we will present some results of a model cluster constr
11、ucted at the Cyberinfrastructure Research Laboratory at Clemson University.MODELIn essence,virtualization is making one computer appear to be multiple computers. Jones2006Virtualization is accomplished with a program called a hypervisor, while systems running under a hypervisor are known as virtual
12、machines(VMs. There are two basic types of hypervisorIBM2005:Type1hypervisors directly interface with the system operating systems run inside a virtual is usually a special,privileged virtual machine that can manage the is an example of this type of hypervisor. Type2hypervisors run as a normal progr
13、am inside a normal operating system. This OS is known as the guest OS runs as a process in the host OS. These processes can be manipulated just like any other and KVM are examples of this type of hypervisor. - 让每个人平等地提升自我 4 See Figure1for a comparison of Type1and2hypervisors.A strict Type2hypervisor
14、 requires that all I/O devices be emulated completely insoftware,resulting in added overhead for I/O allows the virtual machine to make calls directly to the hypervisor,resulting in potentially increased e requires modi cations to the guest kernel. IBM2005See Table I for a comparison of KVM and Xen.
15、ACM Transactions on Computational Logic,Month20YY.An Evaluation of KVM for Use in Cloud Computing3TableKV M XenT ype1Hypervisor T ype2HypervisorHost is a privileged guest Host directly on hardwareUnprivileged guests Guests have user privilegesx86ring abstraction UNIX process abstractionP aravirtuali
16、zed guests Unmodified guestsVirtual Machine(KVMThe Kernel-based Virtual Machine(KVMis a Type2hypervisor maintained by Qumranet,IncHabib2008Qumranet2006.KVM is based on the QEMU emu-lator and derives all its management tools from main focus of KVM development is to use thex86VT extensions,which allow
17、 virtual machines to make system callsvan - 让每个人平等地提升自我 Doorn2006.KVM versions newer than KVM-62have support for paravirtualized Linux guests,but we did not utilize this capability in our initial prototype.KVM uses a set of Linux kernel modules to provide VT can run on a stock Linux kernel that is:(
18、anew enough and(bhas had the KVM modules built for contrast,Xen requires a heavily patched Linux kernel,on which development lags behind the mainline kernel.KVM supports the QEMU Copy-on-write(QCOWdisk image format,allowing it to support a snapshot mode for its disk I/O snapshot mode, all disk write
19、s are directed to a temporary le,and changes are not persisted to the original disk image VMs can be run from one disk image, somewhat mitigating the huge storage requirements associated with hosting a grid of VMsKeahey et .Destroying a virtual cluster is as simple as sending SIGKILL to each hypervi
20、sor and deleting the image from disk.KVM supports the standard Linux TUN/TAP model for Ethernet using this model,each VM gets its own networking resources,making it indistin-guishable from a physical machine.Compute NodesCentral to the Virtual Organization Cluster Model(VOCMis the Virtual Organi-zat
21、ion Cluster(VOC,which is composed of Virtual Compute Nodes Virtual Organization(VOthat wishes to utilize the compute facilities provided by a Virtualization Service Provider(VSPmust provide a VM image or set of VM images,along with some general con guration each image will potentially be used to spa
22、wn multiple VMs,the con guration of each image must not make any assumptions about the type ofnetworking(hardware interface, hostname,or system-speci c con guration ,dynamic networking con guration should be a hostname has been obtained,dynamic con- guration based upon the hostname is allowed.5 - 让每
23、个人平等地提升自我 6 Our model VOC was built from two VCNs,each with . CentOS provides substantialout-of-the-box support for cluster computing appli-cations and,along with its cousin,Red Hat Enterprise Linux,is widely supported in the scienti c and high-performance computing two VCNs were:ACM Transactions on
24、 Computational Logic,Month20YY.4 et al.(1A virtual head node,which was con gured with the Condor central managerand submit daemons(condor_collector,condor_negotiator,condor_schedd, Ganglia monitoring daemon(gmond,and Ganglia meta daemon(gmetad.(2A virtual compute element,which was con gured with the
25、 Condor job starter(condor_startd,MPICH2,ATLAS(tuned for the virtual CPU,and Ganglia monitoring daemon(gmond.Our model VOC was designed as an Open Science Grid(OSGcompute element. The virtual head node used Condor to distribute incoming OSG jobs to the virtual compute elements.MODELPreparing the phy
26、sical(as opposed to virtualcluster for VOC support required con guring the host OS,setting up support services,con guring networking ser-vices,and con guring storage our prototype implementation,support services included a Lightweight Directory Access Protocol(LDAPserver for cen-tralized administrat
27、ion of hosts and physical user accounts,a Dynamic Host Con- guration Protocol(DHCPserver for assigning IPv4addresses to nodes,and a Domain Name Server(DNSfor host resolution.OS con guration - 让每个人平等地提升自我 7 When providing virtualization services,the host OS should be minimalist in order toreserve as
28、many resources as possible for the this end,Slackware Linux 12was chosen as the host custom kernel was compiled to enable support for KVM and additional network unnecessary hardware drivers and other features were left out of the kernel to minimize its memory driver modules also were built for the c
29、ustom kernel.For maintainability reasons,all the Slackware nodes were installed via PXE boot and a custom automated install allowed the whole cluster to be re-created quickly in case of added nodes,hardware failure,or administrator additional software was maintained in the form of Slackware packages
30、 to allow for rapid deployment across the entire cluster.Support ServicesCon guration information for each VCN was stored in an LDAP database to pro-vide a centralized administration VCN was represented as an LDAP entry with the hostname,IP address,and MAC address MAC address was generated as a loca
31、lly-administered address as to avoid con icts with any other MACs on the LDAP-aware,batch,remote administration tool was also written to aid the systems assist in the dynamic con guration of VCNs,the physical cluster provided DHCP and DNS services to the order to maintain a single con guration sourc
32、e,DHCP and DNS con guration les were generated from the LDAP database by a custom utility. These utilities were required to be run whenever the host information in LDAP is VCN images were maintained on an NFS export that was mounted on each physical compute images were accessed in a read-only mode s
33、ince KVMs snapshot mode was employed to start several VMs from the same ACM Transactions on Computational Logic,Month20YY.An Evaluation of KVM for Use in Cloud Computing5 - 让每个人平等地提升自我 8 TableP rocess Grid (P xQ14x 2P roblem Size 77000CentOS GF LOP S GF LOP S Advantage%TableP rocess Grid (P xQ 1x 11
34、4x 27x 4P roblem Size 00P hysical GF LOP S OC GF LOP S irtualization P enalty % image writes to the virtual disk were made to a temporary le on each physicalcompute KVM exited,these les were .RESULTSTwo di erent tests were performed:the standard High Performance Linpack bench-mark and measuring the
35、VCN boot Performance Linpack (HPLThe following HPL parameters were optimized for our cluster and remained con-stant throughout these tests:block size (NB,process mapping (PMAP,threshold,panel fact (PFACT,recursive stopping criterium (NBMIN,panels in recursion (NDIVs,recursive panel fact.(RFACTs,broa
36、dcast (BCAST,lookahead depth (DEPTH,SWAP,swapping threshold,L1form ,U form,Equilibration,and mem-ory problem size (Nwas derived with the formulaN =nDU where n was the number of nodes tested,D was the number of doubles that can t into a single nodes memory (bytes of node memory /8,and U was the ratio
37、 of memory available for user processes to total memory (80%is a good rule of tests were run with - 让每个人平等地提升自我 ATLAS separately for physical nodes and VCNsand II is a comparison of low-overhead Slackware Linux on the physical nodes to tests were run with 56GiB of memory over14nodes.Table III is a c
38、omparison of HPL run on the physical nodes to HPL run on the tests were run with 28GiB of memory over 14physical nodes and each physical node is dual-core,two VCNs were run on TimesBooting the physical hardware was accomplished in under three minutes for each physical head node required 79seconds fr
39、om power-on to completion of the boot this time,the rst 43seconds were occupied by the Power-On Self Test (POSTprocedure and menu time-outs present for human physical compute node booted in a range of time from 160seconds to this time,up to 117seconds were occupied by the POST procedure,ACM Transact
40、ions on Computational Logic,Month 20YY.6 M. Fenn et al. menu time-outs, and a PXE boot time-out, leaving approximately 45 seconds for the actual Linux boot procedure. Some variation in recorded times was likely due to human error in timing the boot procedure, which was done by means of a stopwatch.
41、After the machines were booted, approximately 75 processes were found to be running on the physical head node, with approximately 65 processes running on each physical compute node. The 29 CentOS virtual machines (28 compute nodes and 1 head node booted in an average of 58 seconds from initiation of
42、 the KVM process. A minimum boot time of 52 seconds, and a maximum boot time of 63 seconds, were observed. Approximately 5 seconds can be attributed to a boot-loader time-out in the virtual machine conguration, giving an average Linux boot procedure time of 47 seconds. Following the boot procedure,
43、approximately 100 processes were found to be running on the virtual head node, with approximately 75 processes running on each VCN. 5. CONCLUSIONS Our choice of a low-overhead host OS seemed to be well-supported by the % (Table II speed advantage of Slackware Linux 12 over CentOS on identical9 - 让每个
44、人平等地提升自我 hardware. Since the Slackware system had such a small memory footprint, larger problem sizes could be used while still avoiding swapping. Boot times for virtual machines were comparable to the actual post-boot-loadertime-out portion of the physical machine boot processes, even though more p
45、rocesses were started in the VMs at boot time. Since there were no hardware timeouts present in the VM boot procedure, the eective wall time for booting a VCN was substantially less than the time required for booting a physical compute node. Furthermore, the similar times for the actual Linux boot p
46、rocesses indicates that KVM disk I/O overhead is negligable in the context of basic system operations. Thus, performing maintenance tasks that require rebooting the system should result in less downtime for a Virtual Organization Cluster when compared to an equivalent physical cluster. As shown in T
47、able III, KVM was ecient when network I/O is not a factor. Unfortunately, poor network performance was observed, as shown by the large (7578% overheads present in MPI jobs. Additional testing will be needed to determine the cause of this performance loss and resolve the underlying issue. Poor networ
48、k performance could be attributed to several factors, which will require more testing to dierentiate: Since the Spanning Tree Protocol (STP was not enabled on our TUN/TAP bridges, routing loops might have been present, introducing a large amount of network latency Each VCN on a physical node had a 1Gbps link to the bridge, but each physical node only has 1Gbps to the switch, collisions might have occurred. Such collisions could have caused binary-exponential backo, thus crippling latency and bandwidth under heavy load. The code for the emulated Gigabit Ethernet NIC in QE