SOCNOC reading report.docx
- 文档编号:6375332
- 上传时间:2023-01-05
- 格式:DOCX
- 页数:14
- 大小:850.17KB
SOCNOC reading report.docx
《SOCNOC reading report.docx》由会员分享,可在线阅读,更多相关《SOCNOC reading report.docx(14页珍藏版)》请在冰豆网上搜索。
SOCNOCreadingreport
SOC&NOCReadingReport
----By孙心海
TheincreasingcommunicationrequirementsoncomplexSystem-on-Chip(SoC)circuitshavequicklyforceddesignersfromaparadigmshifttoanother.Systemsbasedonasingleon-chipbushavemadewaytohierarchicalmultibusconfigurationsand,finally,toon-chipnetworks.Inadditiontocommunicationbandwidthandlatencyrequirements,therearealsodemandsforhigherimmunitytosofterrors,timingvariationsanddefectsinnanometerscaleintegration.AllthesechallengesawaittheemergingNetwork-on-Chip(NoC)paradigm.
A.MainFeaturesofNOC
1.Structure
ThefundamentalideainNetwork-on-Chipistoapplythelayeredapproachasinthetelecommunicationandcomputernetworks.
2.Circuit-switchingandpacket-switching
Circuit-switching:
theconnectionhastobeestablishedfirst,thenthetransmissionwillgooninnon-interruptiblemanner,untiltheconnectionwillbedissolved.Thisissimilartotheoldtelephonelineoperation-dialing(forningtheroutetothedestination),talking,andhangingup.
Packet-switching:
Thedataisencapsulatedintopacketswithsomeheaderinformation,passedtothenetworkfordelivery,andthesendercanforgetaboutit.Thisislikethemailsystem-addressthepacket,giveittothepostman.
3.Nodes,switchesandresources
Nodes:
canbealsocalledswitches,althoughanodecanbesomethingmorecomplicatedthanasimpleswitch.
Switches:
referstoacrossbarswitchthatcanconnectanyinputtoanyoutput.
Resources:
computationandstorageblocksthatcommunicateoverthenetwork.(e.g.,RISC,DSP)
4.Topology
Topologyorinterconnectarchitecturedetermineshowtheresourcesofthenetworkareinterconnected.Mostoftheapproachesusemeshortorustopologyasastartingpoint
.
Figure2.NoCtopologies.Mesh(a),torus(b),bi-directionalring(c),and
OctagonMP-SoC(d).
5.Routingschemes
Adaptiveroutingalgorithmscandynamicallychangetheroute,e.g.,toavoidcongestionproblems.
6.Networkperformance
Theperformancecanbeexpressedwiththroughputandlatencyofthesimulatedormeasuredtraffic.Latencyistheaveragetimeelapsingwhensendingapacketoverthenetwork,i.e.,timebetweenamessageinjectionandthereceptionoftheendofthepacket.Throughputisthemaximumloadthatthenetworkcancarry.
B.NOC-basedSystemDesignOptimization
1)EvolutionaryIPAssignmentforEfficientNoC-basedSystemDesignusingMulti-objectiveOptimization
Introduction
Astheintegrationrateofsemiconductorsincreases,morecomplexcoresforsystem-on-chip(SoC)arelaunched.AsimpleSoCisformedbyhomogeneousorheterogeneousindependentcomponentswhileacomplexSoCisformedbyinterconnectedheterogeneouscomponents.Theinterconnectionandcommunicationofthesecomponentsformanetwork-on-chip(NoC).EachcomponentofaNoCisdesignedasanintellectualproperty(IP)block.AnIPblockcanbeageneralorspecialpurposesuchasaprocessor,memoryandDSP.
Normally,aNoCisdesignedtorunaspecificapplication.Thisapplication,usually,consistsofalimitednumberoftasksthatareimplementedbyasetofIPblocks.AnIPblockcanimplementmorethanasingletaskoftheapplication.Forinstance,aprocessorIPblockcanexecutemanytasksasageneralprocessordoesbutamultiplierIPblockforfloatingpointnumberscanonlymultiplyfloatingpointnumbers.
ThenumberofIPblocksdesigners,aswellasthenumberofavailableIPblocks,isgrowingupfast.InordertoyieldanefficientNoC-baseddesignforagivenapplication,itisnecessarytochoosetheadequateminimalsetofIPblocks.BesidesIPblockscarefulassignment,itisalsonecessarytomaptheseblocksontotheNoCavailableinfra-structure,whichconsistsofasetofcorescommunicatingthroughswitches.Ofcourse,abadmappingcandegradetheNoCperformance.
Usually,theapplicationisviewedasagraphoftaskscalledtaskgraph(TG).EfficientNoC-baseddesignskeyproblem:
taskassignmentandIPblockmapping.Normally,deterministictechniquesarenotviabletosolvesuchproblemssoweusedmulti-objectiveevolutionaryalgorithms(MOEAs)withspecificoperatorsandobjectivefunctions.
HowtorealizeIPassignmentusingMulti-objectiveOptimization
WeusedatafromTheEmbeddedSystemsSynthesisbenchmarksSuite(E3S)asourIPrepository.TheE3SisacollectionofTGs,representingrealapplicationsbasedonembeddedprocessors.
WeusetwoMOEAs:
NSGA-IIandmicroGA.BothofthesealgorithmsweremodifiedaccordingtosomeprescribedNoCdesignconstraints.
PreviousStudyinIPassignmentpreview:
ØHuandMarculescuproposedabranchandboundalgorithmwhichautomaticallymapsIPs/coresintoameshbasedNoCarchitecturethatminimizesthetotalamountofconsumedpowerbyminimizingthetotalcommunicationamongtheusedcores.
ØLeiandKumarproposedatwostepgeneticalgorithmformappingtheTGintoameshbasedNoCarchitecturethatminimizestheexecutiontime.
ØMuraliandDeMicheliaddressedtheproblemunderthebandwidthconstraintwiththeaimofminimizingcommunicationdelaybyexploitingthepossibilityofsplittingtrafficamongvariouspaths.
ØZhouetsuggestedamulti-objectiveexplorationapproach,treatingthemappingproblemasatwoconflictingobjectiveoptimizationproblemthatattemptstominimizetheaveragenumberofhopsandachieveathermalbalance.
TASKGRAPHANDIPREPOSITORYMODELS
Definition1.ATaskGraph(TG)G=G(T,D)isadirectedgraphwhereeachnoderepresentsacomputationalmoduleintheapplicationreferredtoastaskti∈T.Eachdirectedarcdi,j∈D,betweentaskstiandtj,characterizeseither
dataorcontroldependencies.
ti:
ID&typeofPEarcdi:
data/controldependencies
Definition2.AnApplicationCharacterizationGraph(ACG)G=G(C,A)isadirectedgraph,whereeachvertexci∈CrepresentsaselectedIP/coreandeachdirectedarcai,jcharacterizesthecommunicationprocessfromcorecitocorecj.
Ci:
IP/coreinformationarcai:
communicationprocess
THEIPASSIGNMENTPROBLEM
ThemainobjectiveistoselectasetofIPs,fromtheIPrepository,thatminimizestheNoCconsumptionofpower,areaoccupiedandexecutiontime.
Twomainproblems:
selectingtherightsetofIPsthatoptimizetheexecutionofagivenapplicationandfindingthebestphysicalmappingoftheseIPsintotheNoCstructure.
TwoAlgorithmstoresolvetheproblem
1.NSGA-II
2.MicroGA
Thesamepoint:
Bothadoptthedominationconceptwitharankingschemaforclassification.TherankingprocessseparatessolutionsinParetofronts(帕累托最优边界,解的任何一个目标函数的值在不使其他目标函数值恶化的条件下已不可能进一步改进)whereeachfrontcorrespondstoagivenrank.
Differentpoint:
NSGA-II(Non-DominatedSortingGeneticAlgorithm)---featuresafastandelitistrankingprocessthatminimizescomputationalcomplexityandprovidesagoodspreadofsolutions.
MicroGA(Micro-geneticalgorithm)---workswithaverysmallpopulation(3to5individuals),whichmakesitveryfast.
OncetheIPassignmentiscomplete,wehaveasetofindividuals(orasingleone)withrankoneandthosearethebestindividualsfortheavailableinformationintheIPrepository.
Example:
Figure4:
Trade-offsrepresentationofthe25optimalIPassignments
Conclusion:
weproposedadecisionsupportsystembasedonMOEAstohelpNoCdesignerstoselectasetofIPsfromarepositoryofIPs.TheuseoftwodifferentMOEAsconsolidatestheobtainedresults.
2)ABidirectionalNoCArchitecturewithDynamicself-ReconfigurableChannel
Figure2-1TypicalNoCarchitectureinmeshtopology
2.1WhyneedBidirectionalNoCArchitecture
Inacity-blocktiledNoCarchitecture,neighboringroutersareconnectedviaapairofcommunicationchannels.Onechannelwillsupportout-goingtrafficandtheotherchannelwillsupportin-comingtraffic.Undervarioustrafficconditions,itisoftenobservedthattheout-goingchannelmaybefloodedwithoutgoingtrafficwhiletheincomingchannelremainsidle.Inmanymetropolitanareas,duringrushhours,thetraveldirectionsofsomehighwaylanesarereversedtemporarilytorelievecongestionsoftheopposingtrafficdirection.
WeexploresimilarideaasamechanismtorelieveintermittenttrafficcongestionintheNoCcommunicationbackbone,andhenceenhanceoverallperformance.Specifically,weproposeakeyinnovationtoreplacethepairofunidirectionalchannelsbetweenroutersbyapairofbidirectionalchannelsthatcanbedynamicallyself-reconfigured.
2.2Self-ReconfigurableRoutingScheme
2.2-a.ProblemFormulation
InaconventionalNoCarchitecture,eachpairofneighboringroutersusestwounidirectionalchannelsinoppositedirectiontopropagatedataonthenetworkasshowninFigure2.(a).Toenablethemostbandwidthutilization,datachannelsbetweeneachpairofroutersshouldbeabletotransmitdatainanydirectionateachruncycle.Thatis,fourkindsofchanneldirection
combinationsshouldbeallowedfordatatransmissionasshowninFigure2(b).
2.2-b.Example
WecandynamicallychangethedirectionofeachchannelbetweeneachpairofrouterslikethearchitectureillustratedinFigure3(c),thebandwidthutilizationwillbeimprovedandthetotalexecutiontimecanbereducedto55cycles.Figure4showsthedetailedexecutionschedules,wheretherequiredcommunicationtimebetweennodesinBiNoCisextensivelyreduced.
2.2-c.Inter-RouterTransmissionSchemeIntroduction
Figure6showsthedetailofthecontrolblocksandsignalsforonechannelbetweenthetwoadjacentrouters.Routingmoduledetectsthepacketheaderandcomputestheoutputdirection.Thechannelcontrolmoduleonthetargetchanneldecodesthisrequestandisresponsibletocommunicatewiththeotherchann
- 配套讲稿:
如PPT文件的首页显示word图标,表示该PPT已包含配套word讲稿。双击word图标可打开word文档。
- 特殊限制:
部分文档作品中含有的国旗、国徽等图片,仅作为作品整体效果示例展示,禁止商用。设计者仅对作品中独创性部分享有著作权。
- 关 键 词:
- SOCNOC reading report
![提示](https://static.bdocx.com/images/bang_tan.gif)