bringphysicalmemory

physicalmemory  时间:2021-05-19  阅读:()
Lecture20:CacheHierarchies,VirtualMemoryToday'stopics:CachehierarchiesVirtualmemoryReminder:Assignment8willbepostedsoon(dueTue11/21)ExampleAccessPattern8-bytewords101000Direct-mappedcache:eachaddressmapstoauniqueaddressByteaddressTagCompareDataarrayTagarrayAssumethataddressesare8bitslongHowmanyofthefollowingaddressrequestsarehits/misses4,7,10,13,16,68,73,78,83,88,4,7,10…IncreasingLineSize32-bytecachelinesizeorblocksize10100000ByteaddressTagDataarrayTagarrayOffsetAlargecachelinesizesmallertagarray,fewermissesbecauseofspatiallocalityAssociativity10100000ByteaddressTagDataarrayTagarraySetassociativityfewerconflicts;wastedpowerbecausemultipledataandtagsarereadWay-1Way-2CompareAssociativity10100000ByteaddressTagDataarrayTagarrayHowmanyoffset/index/tagbitsifthecachehas64sets,eachsethas64bytes,4waysWay-1Way-2CompareExample32KB4-wayset-associativedatacachearraywith32bytelinesizesHowmanysetsHowmanyindexbits,offsetbits,tagbitsHowlargeisthetagarrayCacheMissesOnawritemiss,youmayeitherchoosetobringtheblockintothecache(write-allocate)ornot(write-no-allocate)Onareadmiss,youalwaysbringtheblockin(spatialandtemporallocality)–butwhichblockdoyoureplacenochoiceforadirect-mappedcacherandomlypickoneofthewaystoreplacereplacethewaythatwasleast-recentlyused(LRU)FIFOreplacement(round-robin)WritesWhenyouwriteintoablock,doyoualsoupdatethecopyinL2write-through:everywritetoL1writetoL2write-back:marktheblockasdirty,whentheblockgetsreplacedfromL1,writeittoL2WritebackcoalescesmultiplewritestoanL1blockintooneL2writeWritethroughsimplifiescoherencyprotocolsinamultiprocessorsystemastheL2alwayshasacurrentcopyofdataTypesofCacheMissesCompulsorymisses:happensthefirsttimeamemorywordisaccessed–themissesforaninfinitecacheCapacitymisses:happensbecausetheprogramtouchedmanyotherwordsbeforere-touchingthesameword–themissesforafully-associativecacheConflictmisses:happensbecausetwowordsmaptothesamelocationinthecache–themissesgeneratedwhilemovingfromafully-associativetoadirect-mappedcacheVirtualMemoryProcessesdealwithvirtualmemory–theyhavetheillusionthataverylargeaddressspaceisavailabletothemThereisonlyalimitedamountofphysicalmemorythatissharedbyallprocesses–aprocessplacespartofitsvirtualmemoryinthisphysicalmemoryandtherestisstoredondisk(calledswapspace)Thankstolocality,diskaccessislikelytobeuncommonThehardwareensuresthatoneprocesscannotaccessthememoryofadifferentprocessAddressTranslationThevirtualandphysicalmemoryarebrokenupintopagesVirtualaddress8KBpagesizepageoffsetvirtualpagenumberTranslatedtophysicalpagenumberPhysicaladdress13MemoryHierarchyPropertiesAvirtualmemorypagecanbeplacedanywhereinphysicalmemory(fully-associative)ReplacementisusuallyLRU(sincethemisspenaltyishuge,wecaninvestsomeefforttominimizemisses)Apagetable(indexedbyvirtualpagenumber)isusedfortranslatingvirtualtophysicalpagenumberThepagetableisitselfinmemoryTLBSincethenumberofpagesisveryhigh,thepagetablecapacityistoolargetofitonchipAtranslationlookasidebuffer(TLB)cachesthevirtualtophysicalpagenumbertranslationforrecentaccessesATLBmissrequiresustoaccessthepagetable,whichmaynotevenbefoundinthecache–twoexpensivememorylook-upstoaccessonewordofdata!
AlargepagesizecanincreasethecoverageoftheTLBandreducethecapacityofthepagetable,butalsoincreasesmemorywastageTLBandCacheIsthecacheindexedwithvirtualorphysicaladdressToindexwithaphysicaladdress,wewillhavetofirstlookuptheTLB,thenthecachelongeraccesstimeMultiplevirtualaddressescanmaptothesamephysicaladdress–mustensurethatthesedifferentvirtualaddresseswillmaptothesamelocationincache–else,therewillbetwodifferentcopiesofthesamephysicalmemorywordDoesthetagarraystorevirtualorphysicaladdressesSincemultiplevirtualaddressescanmaptothesamephysicaladdress,avirtualtagcomparisoncanflagamissevenifthecorrectphysicalmemorywordispresentCacheandTLBPipelineTLBVirtualaddressTagarrayDataarrayPhysicaltagcomparionVirtualpagenumberVirtualindexOffsetPhysicalpagenumberPhysicaltagVirtuallyIndexed;PhysicallyTaggedCacheBadEventsConsiderthelongestlatencypossibleforaloadinstruction:TLBmiss:mustlookuppagetabletofindtranslationforv.
pagePCalculatethevirtualmemoryaddressforthepagetableentrythathasthetranslationforpageP–let'ssay,thisisv.
pageQTLBmissforv.
pageQ:willrequirenavigationofahierarchicalpagetable(let'signorethiscasefornowandassumewehavesucceededinfindingthephysicalmemorylocation(R)forpageQ)AccessmemorylocationR(findthiseitherinL1,L2,ormemory)Wenowhavethetranslationforv.
pageP–putthisintotheTLBWenowhaveaTLBhitandknowthephysicalpagenumber–thisallowsustodotagcomparisonandchecktheL1cacheforahitIfthere'samissinL1,checkL2–ifthatmisses,checkinmemoryAtanypoint,ifthepagetableentryclaimsthatthepageisondisk,flagapagefault–theOSthencopiesthepagefromdisktomemoryandthehardwareresumeswhatitwasdoingbeforethepagefault…phew!
TitleBullet

A400互联1H/1G/10M/300G流量37.8元/季

A400互联是一家成立于2020年的商家,本次给大家带来的是,全新上线的香港节点,cmi+cn2线路,全场香港产品7折优惠,优惠码0711,A400互联,只为给你提供更快,更稳,更实惠的套餐。目前,商家推出香港cn2节点+cmi线路云主机,1H/1G/10M/300G流量,37.8元/季,云上日子,你我共享。A400互联优惠码:七折优惠码:0711A400互联优惠方案:适合建站,个人开发爱好者配置...

百星数据(60元/月,600元/年)日本/韩国/香港cn2 gia云服务器,2核2G/40G/5M带宽

百星数据(baixidc),2012年开始运作至今,主要提供境外自营云服务器和独立服务器出租业务,根据网络线路的不同划分为:美国cera 9929、美国cn2 gia、香港cn2 gia、韩国cn2 gia、日本cn2 gia等云服务器及物理服务器业务。目前,百星数据 推出的日本、韩国、香港cn2 gia云服务器,2核2G/40G/5M带宽低至60元/月,600元/年。百星数据优惠码:优惠码:30...

95IDC香港特价物理机服务器月付299元起,5个ip/BGP+CN2线路;美国CERA服务器仅499元/月起

95idc是一家香港公司,主要产品香港GIA线路沙田CN2线路独服,美国CERA高防服务器,日本CN2直连服务器,即日起,购买香港/日本云主机,在今年3月份,95IDC推出来一款香港物理机/香港多ip站群服务器,BGP+CN2线路终身7折,月付350元起。不过今天,推荐一个价格更美的香港物理机,5个ip,BGP+CN2线路,月付299元起,有需要的,可以关注一下。95idc优惠码:优惠码:596J...

physicalmemory为你推荐
签约xp清华大学经济管理学院previouslybit支持ipad支持ipad支持ipadwindows键是哪个Windows键是哪个键啊?ipadwifiIpad怎么用移动无线上网win10关闭445端口win10家庭版怎么禁用445端口联通版iphone4s怎么区分iphone4s电信版和联通版
中文域名申请 东莞电信局 漂亮qq空间 安云加速器 私人服务器 softbank官网 lamp配置 免费smtp服务器 英文站群 100x100头像 我爱水煮鱼 什么是刀片服务器 qq云端 卡巴斯基免费试用 东莞服务器托管 中国linux lamp的音标 登陆qq空间 godaddy空间 阿里云邮箱个人版 更多