电子邮件服务器容量规划

Exchange电子邮件服务器容量规划叶晓彬高级顾问微软顾问咨询部门议题容量规划过程Exchange服务器容量规划和设计基本监控总结What’sOutThere?Benchmarks基准…不是容量规划最好的数据来源Why?单服务器不符合实际的配置单一的负荷压力没有额外的软件和冗余考虑用来对比服务器的类型，而不是提供容量规划的指南其它方法?压力测试工具LoadSim2003,JetStress,ESPSizingtoolsVaryinglevelsRiskofGIGOassolutiongrowsWhitepapersKBArticles技术顾问，系统架构师6步容量规划的方法1.获得基础数据a)充分理解SLA对于设计一个成功的项目非常重要2.建立自己的benchmarks3.获得基线数据4.确定瓶颈5.平衡系统资源6.实施，然后主动监控获得数据:使用场景Userprofiles转换为可计量的数据IOPS/uMegacycles/mailboxNetworkbytes/mailbox利用峰值作容量规划确定瓶颈反应时间Server’sreactiontoloadOnlypartofthepicture吞吐量Messages/secTransactions/sec瓶颈在哪里?High-demandresourceWorkload-dependentNeverreallyeliminate;justmove#OFUSERSTPUTRATE#OFUSERSRESPTIME平衡系统资源CPUMemoryPhysicalandVirtualI/ODisksNetworkScaleuporscaleout?议题容量规划过程Exchange服务器容量规划和设计服务器角色邮箱服务器容量规划I/O规划工具基本监控总结Exchange2003服务器角色不同的角色不同的需求后端邮箱服务器前端服务器连接器InboundSMTPgateways/SMTPBridgeheadX.400/legacyconnectorserversDLExpansionservers公共文件夹Free/busy其它(e.g.,OABGenerator)选择服务器采用下面的服务器类型:BladeserversFEservers,BEserverswithSAN/iSCSIDenseserverHighlyavailable,scalableBE,FEserversHigh-capacityserversBridgeheadservers,replicatedsites/departmentalVirtualserversLab&hardware-independentQA考虑:PowerconsumptionDirect-attachedversusnetworkstorageHigh-performanceserversHigh-availabilityoptionsExchange容量:CPU最佳选择:仍然是4processorsFamily-dependentYesterday:PentiumIIIXeonToday:Xeon/XeonMP/OpteronGoal:Burn80%(achallenge?)扩展性的因素Hyperthreading/dual-coreClockfrequency(3.6GHz)L2,L3cache(upto8MBL3cache)Front-sidebus/HyperTransportx64extensions处理器比较PIIIXeonXeonMPOpteronXeonEM64TXeonMPEM64TMax#procs4(32)4(32)4+(singleanddual-core)2;44Clockspeed(max)900MHz3.0GHz2.6GHz3.6GHz3.3GHzCache2MBL24MBL31MBL21-2MBL28MBL3x64?NoNoYesYesYesFSBspeed133MHz400MHzProcspeed800MHz667MHzdualHyper-threadingNoYesDual-coreYesYesMemorysupportedPC133PC1600PC3200PC2-3200RPC2-3200RTopMMB3N/A9,500(4P)10,200(4Psingle-core)8,208(2P)10,520(4P)TimeFrequency,L3Cache,FSBComparisonsTotalCPU66.4%54.51%49.65%40.95%39.43%0%10%20%30%40%50%60%70%%CPU2.8GHz/4002.8GHz/5333.06GHz/512KB3.06GHz/1MB3.2GHz/1MBFront-SideBus400MHz533MHzClockspeedboostwithoutL3cacheAdditionofL3cache(1MB)ClockspeedboostwithL3cachePlatform:ProLiantDL360G3;2P,Exchange2003RTM;identicalconfigurationWorkload:3,000MMB3.ResponseTimerelativelyconsistentHyper-ThreadingResultsHyper-Threadingcomparison14%30%47%66%97%7%17%28%46%73%0%10%20%30%40%50%60%70%80%90%100%10002000300040005000MMB3012345678910H-TDisabledH-TEnabledProcQueue(H-TOff)Procqueue(H-TOn)Platform:ProLiantDL580G2;4P2.0GHz/2Mcache,Exchange2003RTMWorkload:MMB315–25%boostExchange容量:内存物理内存和虚拟内存4GBonhigh-endback-endservers(2,000mailboxesandabove)/3GB(BEandPFservers)with1GB其它减少虚拟内存碎片的方法HeapDecommitFreeBlockThresholdWatchFreeSystemPageTableEntries:/USERVA:3030ESECachetuning不要在FE,SMTP上配置CheckKBs815372,810371Exchange容量:网络后端服务器Typically,100MbitfullduplexsufficientConsiderGigabitandNICteamingif:Networkbackup/restoreiSCSIorNASstorage–TCPOffloadEnginesupportHighconcentrationofOWA/POP/IMAPusersConsiderNICswithIPSecoffloadMAPIcompression/bufferpacking30–40%improvementincachedmodesynchronization,standardMAPItrafficoverE2K前端服务器Dualswitched100MbitorGbitdefaultExchange容量:网络基本考虑点SizeforI/O*and*capacity考虑SLAArchitectforhighavailability,backup…don’t“add-on”VSSorreplication将数据库和log分放在不同的PerformanceanddisasterrecoveryconsiderationsRAID10everywhere优化延迟Database:fastreads(20ms)Logs:fastwrites(10ms)BurstyrandomI/OpatternsDesignforcontrollercacheavoidanceDesignforMondaymorningpeakDiskIOPsIOper-spindledependenciesRAIDtypeLatencyrequirementsAccesspatternRead/writemixSpindleperformance(rotationalspeed)基本原则(FCandU320SCSI)10Kdisks:120IOPSperdisk@20ms15Kdisks:160IOPSperdisk@20msConservativeuseof100IOPSperdiskfordesignallowsroomforpeakperiodsFigureareRAID1+0;RAID5IOPSdropsto~90/10KdiskTransactionlogsrequire~10%ofSGIOPS磁盘瓶颈分析MSA1000;16disks/database;1DB/LUN/RAID0+1Array;increasingloadtosaturationgeneratedbyJetStressReactiontoLoadSaturation1384.31600.61855.31944.01855.3513.4603.4695.8725.7692.1020040060080010001200140016001800200056101620#JetStressThreads0510152025r/secw/secdiskqueuemsec/rmsec/wMB/s~20ms/write~10ms/readSaturation存储设计:前端服务器SMTPCreateasinglepartition;~30smallmsgs/sec/spindleRAID0+1recommended;100%write-backcacheX.400andlegacyconnectorsSeparateMTA(RAID5),logs(RAID1)andpagefile,ifpossibleOtherFEserversNotI/O-intensive存储设计:后端服务器很多需要考虑的因素…Array/LUNdesignJournaling/ArchivingVSS-basedplexes(snapshot,clones)Temp/TMPdrivePartitionAlignmentwithphysicalsectors(Diskpar/DiskpartwithSP1)…tomentionafewDatabaseIOPs/uRatesRoughestimates(MAPIcorporateusers)ISP/“light”user:.3IOPs/uCorporate:.7–1.2IOPs/uMicrosoft:.9–3IOPs/uRatesdonotstayconstantasload,mailboxsizesincrease501001502002503003504004505005506006507007508008509009501000AverageMailboxsize(MB)500users1000users1500users2000users2500users3000users3500users4000users4500users5000usersRAID级别比较RAID1+0,RAID5,andRAIDADGPerformanceComparison0100020003000400050006000515253545JetstressThreadsDiskTransfers/secRAID1+0RAID5RAIDADGRAID1+0,RAID5,andRAIDADGLatencyComparison0510152025515253545JetstressThreadsAvg.Di