assist迈拓kvm
迈拓kvm 时间:2021-01-23 阅读:()
ExtendingKVMModelsTowardHigh-PerformanceNFVJunNakajima,JamesTsai,MesutErgin,YangZhang,andWeiWang14October2014LegalDisclaimerINFORMATIONINTHISDOCUMENTISPROVIDEDINCONNECTIONWITHINTELPRODUCTS.
NOLICENSE,EXPRESSORIMPLIED,BYESTOPPELOROTHERWISE,TOANYINTELLECTUALPROPERTYRIGHTSISGRANTEDBYTHISDOCUMENT.
EXCEPTASPROVIDEDININTEL'STERMSANDCONDITIONSOFSALEFORSUCHPRODUCTS,INTELASSUMESNOLIABILITYWHATSOEVER,ANDINTELDISCLAIMSANYEXPRESSORIMPLIEDWARRANTY,RELATINGTOSALEAND/ORUSEOFINTELPRODUCTSINCLUDINGLIABILITYORWARRANTIESRELATINGTOFITNESSFORAPARTICULARPURPOSE,MERCHANTABILITY,ORINFRINGEMENTOFANYPATENT,COPYRIGHTOROTHERINTELLECTUALPROPERTYRIGHT.
INTELPRODUCTSARENOTINTENDEDFORUSEINMEDICAL,LIFESAVING,ORLIFESUSTAININGAPPLICATIONS.
Intelmaymakechangestospecificationsandproductdescriptionsatanytime,withoutnotice.
Allproducts,dates,andfiguresspecifiedarepreliminarybasedoncurrentexpectations,andaresubjecttochangewithoutnotice.
Intel,processors,chipsets,anddesktopboardsmaycontaindesigndefectsorerrorsknownaserrata,whichmaycausetheproducttodeviatefrompublishedspecifications.
Currentcharacterizederrataareavailableonrequest.
IntelandtheIntellogoaretrademarksorregisteredtrademarksofIntelCorporationoritssubsidiariesintheUnitedStatesandothercountries.
*Othernamesandbrandsmaybeclaimedasthepropertyofothers.
Copyright2014IntelCorporation.
TheChallengeArchitectureProposalsforNFVforKVMCurrentStatusandSummary3Agenda4NFVVisionfromETSISource:http://portal.
etsi.
org/nfv/nfv_white_paper2.
pdfHighperformanceacrossallpacketsizes,includingsmallpackets(e.
g.
64B)Real-timeprocessing,includinglowlatencyandjitterRASSecurity.
.
.
5New/DifferentRequirementsforNFVComparedwithConventionalVirtualizationFocusonPerformanceTopicsTodayTheChallenge020,000,00040,000,00060,000,00080,000,000100,000,000120,000,000140,000,000160,000,000641281922563203844485125766407047688328969601024108811521216128013441408147210GbEPacketsPerSecond40GbEPacketsPerSecond100GbEPacketsPerSecond6.
72ns16.
8ns67.
2nsSaturationLineRate(MPPS)Source:DPDKSummit,VenkyVenkatesan,"ApplicationPerformanceTuningandFutureOptimizationsinDPDK",September8,2014https://www.
youtube.
com/watchv=qpfwDySweUADisclaimer:SoftwareandworkloadsusedinperformancetestsmayhavebeenoptimizedforperformanceonlyonIntelmicroprocessors.
Performancetests,suchasSYSmarkandMobileMark,aremeasuredusingspecificcomputersystems,components,software,operationsandfunctions.
Anychangetoanyofthosefactorsmaycausetheresultstovary.
Youshouldconsultotherinformationandperformanceteststoassistyouinfullyevaluatingyourcontemplatedpurchases,includingtheperformanceofthatproductwhencombinedwithotherproducts.
425593164.
925505010015020025030020092010201120122013SystemLevelL3Performance(MPPS)*Othernamesandbrandsmaybeclaimedasthepropertyofothers.
Disclaimer:SoftwareandworkloadsusedinperformancetestsmayhavebeenoptimizedforperformanceonlyonIntelmicroprocessors.
Performancetests,suchasSYSmarkandMobileMark,aremeasuredusingspecificcomputersystems,components,software,operationsandfunctions.
Anychangetoanyofthosefactorsmaycausetheresultstovary.
Youshouldconsultotherinformationandperformanceteststoassistyouinfullyevaluatingyourcontemplatedpurchases,includingtheperformanceofthatproductwhencombinedwithotherproducts.
IntelDPDKPerformanceAsnapshotofondifferentarchitecturesIntegratedMemoryControllerPCI-EGen2DataDirectI/OIntegratedPCI-EGen3AVX(integer,128-bit)4x10GbENICsPlatformFeaturesSource:DPDKSummit,VenkyVenkatesan,"ApplicationPerformanceTuningandFutureOptimizationsinDPDK",September8,2014https://www.
youtube.
com/watchv=qpfwDySweUALinuxKernelVMorUserProcessKernel(virt.
I/O)VM1VT-d,SR-IOVMiddleBox(e.
g.
virtualswitch)KVMKernel(virt.
I/O)VM2.
.
.
8FocusAreasforNFVPerformanceonKVMRecall67.
2ns,16.
8ns,…FastandEfficientInter-VMCommunicationGeneric:NetworkI/O,NUMA,NUMA-I/O,Caching,Affinity,…MorecoresMoremiddleboxespersocket,perserverServicechainingonserverLowerlatencyInter-VM(i.
e.
intra-node)vs.
Inter-nodeHigherBandwidthMemory(orcache)vs.
PCIebus9WhyInter-VMCommunicationFigure1.
TheIntelXeonprocessorE5-2600V2productfamilyMicroarchitectureSource(Figure1.
):https://software.
intel.
com/en-us/articles/intel-xeon-processor-e5-2600-v2-product-family-technical-overviewNotificationsforqueuecontrolKick,DoorBellVirtualSwitchPacketTransmissionCopy,etc.
TransitionsUser-KernelGuest-Host10Inter-VMCommunicationonKVMHostOSRxVMtapTxVMtapswitchappapphypervisorStackdrvdevdevStackdrvSwitchingpathcanbeabigperformancebottleneckXYX0.
712Mpps*Y0.
717Mpps**Intelinternalmeasurements64Bpackets,virtio-net+vhost-netTSCCycles(Haswell3.
2GHz),RoundTrip*:UserKernel(SystemCall)inVM(onKVM)E.
g.
getppid():1300(≈400ns)GuestHost(HyperCall)E.
g.
NullHypercall:1500-1600(≈500ns)ToreachSaturationLineRate(10GbE):Ifsystemcall/Hypercallisusedforeach64Bpackettransmission,wewouldneed:>6-7Cores**40GbE:>24-28Cores11CostofTransitions/IsolationPerspectiveofCPUCycles**:400/67.
2=5.
9,500/67.
2=7.
4*IntelinternalmeasurementsDisclaimer:SoftwareandworkloadsusedinperformancetestsmayhavebeenoptimizedforperformanceonlyonIntelmicroprocessors.
Performancetests,suchasSYSmarkandMobileMark,aremeasuredusingspecificcomputersystems,components,software,operationsandfunctions.
Anychangetoanyofthosefactorsmaycausetheresultstovary.
Youshouldconsultotherinformationandperformanceteststoassistyouinfullyevaluatingyourcontemplatedpurchases,includingtheperformanceofthatproductwhencombinedwithotherproducts.
Practically,thoseareratherlowerboundsbecausebatchingislimitedandactualpacketprocessinginhypercallsoverturnsgainofbatching.
TheChallengeArchitectureProposalsforNFVforKVMCurrentStatusandSummary12Agenda1.
Moveknowledgeandcontrolforinter-VMcommunicationtoVMs2.
AllowVMstoaccessotherVMstoshareoraccessmemoryinasafewayProvideVMswith"ProtectedMemoryView"Mappingitselfisprovidedbythehypervisor3.
AllowVMstouselow-latencynotificationmechanismsw/oVMexitsorinterruptsE.
g.
MONITOR/MWAIT,PostedInterrupt13Solutions:EmpowerGuestsinaSafeWayAvoidhypervisorinterventionsMotivation:WhydoesakernelmoduleneedtoknowaboutdatastructuresforPVdriversinguestsBecausewetrustkernelorkernelmodulesonly.
Whatifwetrustspecific(partof)guests…Vhost-netinguestcanavoidhypercallsifitcandirectlyaccessdestinationguests(virtqueue,etc.
)14Example:vhost-netFunctionalityinGuestsvhost-userisalreadythere15High-LevelArchitectureforFastInter-VMCommunication(w/oVT-d,SR-IOV)LinuxKernelKVMVM1Kernelvirto-netVM2Kernelvirto-netSharedmemoryforsynchronizationFastPathFastPathProtectedMemoryViewLow-LatencyNotification1.
DataTransmission2.
NotificationDirectAccesstoGuestsInProtectedMemoryViewVhost-netAPIFastPathcanworkwithvirtio-netorindependentlyvirtio-netvirtio-netLinuxKernelKVMVM1Kernelvirtio-netVM2Kernelvirtio-netSharedmemoryforsynchronizationFastPathFastPathMiddleBox(e.
g.
virtualswitch)VM0VT-d,SR-IOVFastPacketTransmissionSharedmemoryforsynchronization16High-LevelArchitectureforFastInter-VMCommunication(withVT-d,SR-IOV)FastPacketTransmissioncanbeinuser-levelVMFUNCinstructionwithEAX=0ValueinECXselectsanentryfromtheEPTP(Extended-Page-TablePointer)listAvailableinRing0-3,executedinguestNoVMexitCanbevirtualizedifnotavailable17IntroducingVMFunction0:EPTP*Switching…EPTP…ECX(index)EPTPlist(4KB)VMCS(perVCPU)*:Extended-Page-TablePointerEPTPSwitchingandTrampolineCodeGuestPhysicalPagesProtectedView(code,data)---DefaultEPTEPT:HostPhysicalPagesXWR18NoAccessTrampolinecodeforVMFUNCXWR-WRX-RX-RVMFUNCexecutedoutsideTrampolineCodewillcauseEPTviolationatnextinstructionHypervisorneedstorestoreDefaultEPTtodelivervirtualinterruptsEPTPswitchingDefaultViewAlternateViewstart_xmit(*skb,*dev){.
.
.
send(packets);}send(*packet){.
.
.
VMFUNC#0,EPTP;Tx(packets);VMFUNC#0,0}PageBoundaryTx(*packet){move_data();notify();}XWRstart_xmit(*skb,*dev){.
.
.
send(packets);}send(*packet){.
.
.
VMFUNC#0,EPTP;Tx(packets);VMFUNC#0,0}Tx(*packet){move_data();notify();}----WR-WRXWRXWRX-RX-REPTPerm.
ModifyqueuedescriptorsModifyqueuedescriptorsMoveDatabyTx()19MoreDetails:TransmittingPacketsDestinationVMSourceVMTrampolineCodeProtectedView24315DefaultViewAlternateViewPostedInterruptDelivervirtualinterruptsondestinationguestsw/oVMexits.
AlreadysupportedbyKVMStillrequiresVMexitonsourceguestMONITOR/MWAIT(Energy-EfficientPolling)betweenguestsThefeatureisnotadvertisedonKVMtodayUsevariablesonsharedmemorybetweensourceanddestinationPAUSELoop(Polling)betweenguestsLowestlatency,butnotenergyefficientInpractice,combineInterruptandPolling(likeNAPI)20Low-LatencyNotificationKnownmethodsMinimizeimpactofTLBmisses,cachemisses:Largepages(bothguest,EPT,VT-d),NUMA,IO-NUMA,DataDirectI/OE.
g.
LIFOmemorypoolZero-copyE.
g.
AddsourcebuffersmappingtoEPTofdestinationIfEPTPTEswerenotvalid,noINVEPTisrequired21PracticesforPerformanceGeneralDisclaimer:SoftwareandworkloadsusedinperformancetestsmayhavebeenoptimizedforperformanceonlyonIntelmicroprocessors.
Performancetests,suchasSYSmarkandMobileMark,aremeasuredusingspecificcomputersystems,components,software,operationsandfunctions.
Anychangetoanyofthosefactorsmaycausetheresultstovary.
Youshouldconsultotherinformationandperformanceteststoassistyouinfullyevaluatingyourcontemplatedpurchases,includingtheperformanceofthatproductwhencombinedwithotherproducts.
FrequencyofVMFUNCoperation:CostofVMFUNCisabout150TSCcycles(Haswell,3.
2GHz)*Around50ns,andsensitivetoTLB,cachesRecall67.
2ns,16.
8ns,…ToreachSaturationLineRate(10GbE):IfVMFUNCiscalledforeach64Bpackettransmission,we>1-2Cores(100nsforround-trip)40GbE:>4-8CoresThecostofVMFUNCwouldberelativelysmall,anditwouldprovidescalableperformance22PracticeforPerformanceEPTPSwitching*IntelinternalmeasurementsDisclaimer:SoftwareandworkloadsusedinperformancetestsmayhavebeenoptimizedforperformanceonlyonIntelmicroprocessors.
Performancetests,suchasSYSmarkandMobileMark,aremeasuredusingspecificcomputersystems,components,software,operationsandfunctions.
Anychangetoanyofthosefactorsmaycausetheresultstovary.
Youshouldconsultotherinformationandperformanceteststoassistyouinfullyevaluatingyourcontemplatedpurchases,includingtheperformanceofthatproductwhencombinedwithotherproducts.
getppid()inVM:1300(≈400ns)NullHypercall:1500-1600(≈500ns)Practically,thoseareratherlowerboundsbecausebatchingislimitedandactualpacketprocessingoverturnsgainofbatching.
TrampolineCodeisloadedbytheguest,buttheEPTpermission(X-R)issetbyKVMShouldbesignedtogetherwiththecodeintheProtectedViewinadvanceThesetofpages(inDestinationVM)accessedbycodeinProtectedViewneedtobecheckedandaddedbyKVMInaway,codeinProtectedViewisanextensionoftheKVM/hypervisorrunningincontrolledenvironment(stillinVXMnon-rootmode)23SecurityConsiderationTheChallengeArchitectureProposalsforNFVforKVMCurrentStatusandSummary24AgendaPoCinprogress:MeasuredcostofVMFUNC,memorybandwidthEnabledandmeasuredlatencyofMONITOR/MWAITinguestsMeasuringpathAWorkingonpathB25CurrentStatusPoCLinuxKernelKVMVM1Kernelvirto-netVM2Kernelvirto-netSharedmemoryforsynchronizationFastPathFastPathABvirtio-netvirtio-netBenefitsoftheArchitecture:ContainknowledgeandcontrolforInter-VMcommunicationinguestsAllowKVMtoenablemoreoptimizationandcustomizationforgueststohandlehighnetworkloadsefficientlyMoreefficientandscalablethanexistingonesWorkwithdirectI/OassignmentaswellNextStep:CompletePoCandgetmoredata26Summary27BackupCanoccuronlyinguest(vector20)SomeEPTviolationscangenerate#VEinsteadofVMexits(controlledbyhypervisor)Canvirtualizedifnotavailable#VE:VirtualizationException
NOLICENSE,EXPRESSORIMPLIED,BYESTOPPELOROTHERWISE,TOANYINTELLECTUALPROPERTYRIGHTSISGRANTEDBYTHISDOCUMENT.
EXCEPTASPROVIDEDININTEL'STERMSANDCONDITIONSOFSALEFORSUCHPRODUCTS,INTELASSUMESNOLIABILITYWHATSOEVER,ANDINTELDISCLAIMSANYEXPRESSORIMPLIEDWARRANTY,RELATINGTOSALEAND/ORUSEOFINTELPRODUCTSINCLUDINGLIABILITYORWARRANTIESRELATINGTOFITNESSFORAPARTICULARPURPOSE,MERCHANTABILITY,ORINFRINGEMENTOFANYPATENT,COPYRIGHTOROTHERINTELLECTUALPROPERTYRIGHT.
INTELPRODUCTSARENOTINTENDEDFORUSEINMEDICAL,LIFESAVING,ORLIFESUSTAININGAPPLICATIONS.
Intelmaymakechangestospecificationsandproductdescriptionsatanytime,withoutnotice.
Allproducts,dates,andfiguresspecifiedarepreliminarybasedoncurrentexpectations,andaresubjecttochangewithoutnotice.
Intel,processors,chipsets,anddesktopboardsmaycontaindesigndefectsorerrorsknownaserrata,whichmaycausetheproducttodeviatefrompublishedspecifications.
Currentcharacterizederrataareavailableonrequest.
IntelandtheIntellogoaretrademarksorregisteredtrademarksofIntelCorporationoritssubsidiariesintheUnitedStatesandothercountries.
*Othernamesandbrandsmaybeclaimedasthepropertyofothers.
Copyright2014IntelCorporation.
TheChallengeArchitectureProposalsforNFVforKVMCurrentStatusandSummary3Agenda4NFVVisionfromETSISource:http://portal.
etsi.
org/nfv/nfv_white_paper2.
pdfHighperformanceacrossallpacketsizes,includingsmallpackets(e.
g.
64B)Real-timeprocessing,includinglowlatencyandjitterRASSecurity.
.
.
5New/DifferentRequirementsforNFVComparedwithConventionalVirtualizationFocusonPerformanceTopicsTodayTheChallenge020,000,00040,000,00060,000,00080,000,000100,000,000120,000,000140,000,000160,000,000641281922563203844485125766407047688328969601024108811521216128013441408147210GbEPacketsPerSecond40GbEPacketsPerSecond100GbEPacketsPerSecond6.
72ns16.
8ns67.
2nsSaturationLineRate(MPPS)Source:DPDKSummit,VenkyVenkatesan,"ApplicationPerformanceTuningandFutureOptimizationsinDPDK",September8,2014https://www.
youtube.
com/watchv=qpfwDySweUADisclaimer:SoftwareandworkloadsusedinperformancetestsmayhavebeenoptimizedforperformanceonlyonIntelmicroprocessors.
Performancetests,suchasSYSmarkandMobileMark,aremeasuredusingspecificcomputersystems,components,software,operationsandfunctions.
Anychangetoanyofthosefactorsmaycausetheresultstovary.
Youshouldconsultotherinformationandperformanceteststoassistyouinfullyevaluatingyourcontemplatedpurchases,includingtheperformanceofthatproductwhencombinedwithotherproducts.
425593164.
925505010015020025030020092010201120122013SystemLevelL3Performance(MPPS)*Othernamesandbrandsmaybeclaimedasthepropertyofothers.
Disclaimer:SoftwareandworkloadsusedinperformancetestsmayhavebeenoptimizedforperformanceonlyonIntelmicroprocessors.
Performancetests,suchasSYSmarkandMobileMark,aremeasuredusingspecificcomputersystems,components,software,operationsandfunctions.
Anychangetoanyofthosefactorsmaycausetheresultstovary.
Youshouldconsultotherinformationandperformanceteststoassistyouinfullyevaluatingyourcontemplatedpurchases,includingtheperformanceofthatproductwhencombinedwithotherproducts.
IntelDPDKPerformanceAsnapshotofondifferentarchitecturesIntegratedMemoryControllerPCI-EGen2DataDirectI/OIntegratedPCI-EGen3AVX(integer,128-bit)4x10GbENICsPlatformFeaturesSource:DPDKSummit,VenkyVenkatesan,"ApplicationPerformanceTuningandFutureOptimizationsinDPDK",September8,2014https://www.
youtube.
com/watchv=qpfwDySweUALinuxKernelVMorUserProcessKernel(virt.
I/O)VM1VT-d,SR-IOVMiddleBox(e.
g.
virtualswitch)KVMKernel(virt.
I/O)VM2.
.
.
8FocusAreasforNFVPerformanceonKVMRecall67.
2ns,16.
8ns,…FastandEfficientInter-VMCommunicationGeneric:NetworkI/O,NUMA,NUMA-I/O,Caching,Affinity,…MorecoresMoremiddleboxespersocket,perserverServicechainingonserverLowerlatencyInter-VM(i.
e.
intra-node)vs.
Inter-nodeHigherBandwidthMemory(orcache)vs.
PCIebus9WhyInter-VMCommunicationFigure1.
TheIntelXeonprocessorE5-2600V2productfamilyMicroarchitectureSource(Figure1.
):https://software.
intel.
com/en-us/articles/intel-xeon-processor-e5-2600-v2-product-family-technical-overviewNotificationsforqueuecontrolKick,DoorBellVirtualSwitchPacketTransmissionCopy,etc.
TransitionsUser-KernelGuest-Host10Inter-VMCommunicationonKVMHostOSRxVMtapTxVMtapswitchappapphypervisorStackdrvdevdevStackdrvSwitchingpathcanbeabigperformancebottleneckXYX0.
712Mpps*Y0.
717Mpps**Intelinternalmeasurements64Bpackets,virtio-net+vhost-netTSCCycles(Haswell3.
2GHz),RoundTrip*:UserKernel(SystemCall)inVM(onKVM)E.
g.
getppid():1300(≈400ns)GuestHost(HyperCall)E.
g.
NullHypercall:1500-1600(≈500ns)ToreachSaturationLineRate(10GbE):Ifsystemcall/Hypercallisusedforeach64Bpackettransmission,wewouldneed:>6-7Cores**40GbE:>24-28Cores11CostofTransitions/IsolationPerspectiveofCPUCycles**:400/67.
2=5.
9,500/67.
2=7.
4*IntelinternalmeasurementsDisclaimer:SoftwareandworkloadsusedinperformancetestsmayhavebeenoptimizedforperformanceonlyonIntelmicroprocessors.
Performancetests,suchasSYSmarkandMobileMark,aremeasuredusingspecificcomputersystems,components,software,operationsandfunctions.
Anychangetoanyofthosefactorsmaycausetheresultstovary.
Youshouldconsultotherinformationandperformanceteststoassistyouinfullyevaluatingyourcontemplatedpurchases,includingtheperformanceofthatproductwhencombinedwithotherproducts.
Practically,thoseareratherlowerboundsbecausebatchingislimitedandactualpacketprocessinginhypercallsoverturnsgainofbatching.
TheChallengeArchitectureProposalsforNFVforKVMCurrentStatusandSummary12Agenda1.
Moveknowledgeandcontrolforinter-VMcommunicationtoVMs2.
AllowVMstoaccessotherVMstoshareoraccessmemoryinasafewayProvideVMswith"ProtectedMemoryView"Mappingitselfisprovidedbythehypervisor3.
AllowVMstouselow-latencynotificationmechanismsw/oVMexitsorinterruptsE.
g.
MONITOR/MWAIT,PostedInterrupt13Solutions:EmpowerGuestsinaSafeWayAvoidhypervisorinterventionsMotivation:WhydoesakernelmoduleneedtoknowaboutdatastructuresforPVdriversinguestsBecausewetrustkernelorkernelmodulesonly.
Whatifwetrustspecific(partof)guests…Vhost-netinguestcanavoidhypercallsifitcandirectlyaccessdestinationguests(virtqueue,etc.
)14Example:vhost-netFunctionalityinGuestsvhost-userisalreadythere15High-LevelArchitectureforFastInter-VMCommunication(w/oVT-d,SR-IOV)LinuxKernelKVMVM1Kernelvirto-netVM2Kernelvirto-netSharedmemoryforsynchronizationFastPathFastPathProtectedMemoryViewLow-LatencyNotification1.
DataTransmission2.
NotificationDirectAccesstoGuestsInProtectedMemoryViewVhost-netAPIFastPathcanworkwithvirtio-netorindependentlyvirtio-netvirtio-netLinuxKernelKVMVM1Kernelvirtio-netVM2Kernelvirtio-netSharedmemoryforsynchronizationFastPathFastPathMiddleBox(e.
g.
virtualswitch)VM0VT-d,SR-IOVFastPacketTransmissionSharedmemoryforsynchronization16High-LevelArchitectureforFastInter-VMCommunication(withVT-d,SR-IOV)FastPacketTransmissioncanbeinuser-levelVMFUNCinstructionwithEAX=0ValueinECXselectsanentryfromtheEPTP(Extended-Page-TablePointer)listAvailableinRing0-3,executedinguestNoVMexitCanbevirtualizedifnotavailable17IntroducingVMFunction0:EPTP*Switching…EPTP…ECX(index)EPTPlist(4KB)VMCS(perVCPU)*:Extended-Page-TablePointerEPTPSwitchingandTrampolineCodeGuestPhysicalPagesProtectedView(code,data)---DefaultEPTEPT:HostPhysicalPagesXWR18NoAccessTrampolinecodeforVMFUNCXWR-WRX-RX-RVMFUNCexecutedoutsideTrampolineCodewillcauseEPTviolationatnextinstructionHypervisorneedstorestoreDefaultEPTtodelivervirtualinterruptsEPTPswitchingDefaultViewAlternateViewstart_xmit(*skb,*dev){.
.
.
send(packets);}send(*packet){.
.
.
VMFUNC#0,EPTP;Tx(packets);VMFUNC#0,0}PageBoundaryTx(*packet){move_data();notify();}XWRstart_xmit(*skb,*dev){.
.
.
send(packets);}send(*packet){.
.
.
VMFUNC#0,EPTP;Tx(packets);VMFUNC#0,0}Tx(*packet){move_data();notify();}----WR-WRXWRXWRX-RX-REPTPerm.
ModifyqueuedescriptorsModifyqueuedescriptorsMoveDatabyTx()19MoreDetails:TransmittingPacketsDestinationVMSourceVMTrampolineCodeProtectedView24315DefaultViewAlternateViewPostedInterruptDelivervirtualinterruptsondestinationguestsw/oVMexits.
AlreadysupportedbyKVMStillrequiresVMexitonsourceguestMONITOR/MWAIT(Energy-EfficientPolling)betweenguestsThefeatureisnotadvertisedonKVMtodayUsevariablesonsharedmemorybetweensourceanddestinationPAUSELoop(Polling)betweenguestsLowestlatency,butnotenergyefficientInpractice,combineInterruptandPolling(likeNAPI)20Low-LatencyNotificationKnownmethodsMinimizeimpactofTLBmisses,cachemisses:Largepages(bothguest,EPT,VT-d),NUMA,IO-NUMA,DataDirectI/OE.
g.
LIFOmemorypoolZero-copyE.
g.
AddsourcebuffersmappingtoEPTofdestinationIfEPTPTEswerenotvalid,noINVEPTisrequired21PracticesforPerformanceGeneralDisclaimer:SoftwareandworkloadsusedinperformancetestsmayhavebeenoptimizedforperformanceonlyonIntelmicroprocessors.
Performancetests,suchasSYSmarkandMobileMark,aremeasuredusingspecificcomputersystems,components,software,operationsandfunctions.
Anychangetoanyofthosefactorsmaycausetheresultstovary.
Youshouldconsultotherinformationandperformanceteststoassistyouinfullyevaluatingyourcontemplatedpurchases,includingtheperformanceofthatproductwhencombinedwithotherproducts.
FrequencyofVMFUNCoperation:CostofVMFUNCisabout150TSCcycles(Haswell,3.
2GHz)*Around50ns,andsensitivetoTLB,cachesRecall67.
2ns,16.
8ns,…ToreachSaturationLineRate(10GbE):IfVMFUNCiscalledforeach64Bpackettransmission,we>1-2Cores(100nsforround-trip)40GbE:>4-8CoresThecostofVMFUNCwouldberelativelysmall,anditwouldprovidescalableperformance22PracticeforPerformanceEPTPSwitching*IntelinternalmeasurementsDisclaimer:SoftwareandworkloadsusedinperformancetestsmayhavebeenoptimizedforperformanceonlyonIntelmicroprocessors.
Performancetests,suchasSYSmarkandMobileMark,aremeasuredusingspecificcomputersystems,components,software,operationsandfunctions.
Anychangetoanyofthosefactorsmaycausetheresultstovary.
Youshouldconsultotherinformationandperformanceteststoassistyouinfullyevaluatingyourcontemplatedpurchases,includingtheperformanceofthatproductwhencombinedwithotherproducts.
getppid()inVM:1300(≈400ns)NullHypercall:1500-1600(≈500ns)Practically,thoseareratherlowerboundsbecausebatchingislimitedandactualpacketprocessingoverturnsgainofbatching.
TrampolineCodeisloadedbytheguest,buttheEPTpermission(X-R)issetbyKVMShouldbesignedtogetherwiththecodeintheProtectedViewinadvanceThesetofpages(inDestinationVM)accessedbycodeinProtectedViewneedtobecheckedandaddedbyKVMInaway,codeinProtectedViewisanextensionoftheKVM/hypervisorrunningincontrolledenvironment(stillinVXMnon-rootmode)23SecurityConsiderationTheChallengeArchitectureProposalsforNFVforKVMCurrentStatusandSummary24AgendaPoCinprogress:MeasuredcostofVMFUNC,memorybandwidthEnabledandmeasuredlatencyofMONITOR/MWAITinguestsMeasuringpathAWorkingonpathB25CurrentStatusPoCLinuxKernelKVMVM1Kernelvirto-netVM2Kernelvirto-netSharedmemoryforsynchronizationFastPathFastPathABvirtio-netvirtio-netBenefitsoftheArchitecture:ContainknowledgeandcontrolforInter-VMcommunicationinguestsAllowKVMtoenablemoreoptimizationandcustomizationforgueststohandlehighnetworkloadsefficientlyMoreefficientandscalablethanexistingonesWorkwithdirectI/OassignmentaswellNextStep:CompletePoCandgetmoredata26Summary27BackupCanoccuronlyinguest(vector20)SomeEPTviolationscangenerate#VEinsteadofVMexits(controlledbyhypervisor)Canvirtualizedifnotavailable#VE:VirtualizationException
- assist迈拓kvm相关文档
- Size迈拓kvm
- manage迈拓kvm
- VGA迈拓kvm
- Memory迈拓kvm
- 文件迈拓kvm
- Console迈拓kvm
物语云-VPS-美国洛杉矶VPS无限流量云windows大带宽100M不限流量 26/月起
物语云计算怎么样？物语云计算（MonogatariCloud）是一家成立于2016年的老牌国人商家，主营国内游戏高防独服业务，拥有多家机房资源，产品质量过硬，颇有一定口碑。本次带来的是特惠活动为美国洛杉矶Cera机房的不限流量大带宽VPS，去程直连回程4837，支持免费安装Windows系统。值得注意的是，物语云采用的虚拟化技术为Hyper-v，因此并不会超售超开。一、物语云官网点击此处进入物语云...
licloud：$39/月，香港物理服务器，30M带宽，e3-1230v3/16G内存/1T硬盘
licloud官方消息：当前对香港机房的接近100台物理机(香港服务器)进行打折处理，30Mbps带宽，低至不到40美元/月，速度快，性价比高，跑绝大多数项目都是绰绰有余了。该款香港服务器自带启动、关闭、一键重装功能，正常工作日内30~60分钟交货（不包括非工作日）。 官方网站：https://licloud.io 特价香港物理服务器 CPU：e3-1230v2（4核心、8线程、3.3GH...
美国服务器20G防御 50G防御 688元CN2回国
全球领先的IDC服务商华纳云“美国服务器”正式发售啦~~~~此次上线的美国服务器包含美国云服务器、美国服务器、美国高防服务器以及美国高防云服务器。针对此次美国服务器新品上线，华纳云也推出了史无前例的超低活动力度。美国云服务器低至3折，1核1G5M低至24元/月，20G DDos防御的美国服务器低至688元/月，年付再送2个月，两年送4个月，三年送6个月，且永久续费同价，更多款高性价比配置供您选择。...
迈拓kvm为你推荐
-
国内免备案服务器哪里的免备案服务器到国内访问速度快？聚酯纤维和棉哪个好聚酯纤维棉和羽丝绵哪个好传奇类手游哪个好什么传奇类手游好玩不坑钱？手机炒股软件哪个好手机炒股哪个软件好 要免费的雅思和托福哪个好考雅思和托福哪个好考?51空间登录怎样登51个人空间360云盘登录360云盘登陆账号360云盘共享群360云盘共享群怎么没有了360云盘36t现在360云盘不能免费领取36T了吗？活动到期了？怎么领取更多空间？详细步骤和地址。广东联通网上营业厅广州联通营业厅地址在哪？