ILto

October25,2006AnewMonoGCPaoloMolarolupus@novell.
com2CurrentGC:whyBoehmPortedtothemajorarchitecturesandsystemsFeaturefullVeryeasytointegrateHandlesmanagedpointersinunmanagedlandSomesupportfortypedobjectsFinalizationsemanticsmostlymatchWeakreferencesHightlytunedThread-localalloc3CurrentGC:whynotWedon'tneedorusesomeofthefeaturesIncremental(handlingofsignals.
.
.
)Resourceusage(publicAPI)FinalizationsemanticsnotthesameasneededFinalizationofobjectsinvolvedincycles(whichorder)WeakreferencesupportTrackresurrectionHeapfragmentationZeroingoverhead(atomicalloc.
.
.
)Pausetimes4NewGC:needsExactfinalizationandweakreferencesemanticsSmallheapsizesandlow-pausetime(desktopapps)Objectpinning(automaticandAPI-controlled)Fastallocation(bump-pointerandinlinableinmanagedcode)LargeobjecthandlingPrecisetypetracking5NewGC:needs(continued)InteriorpointersonthestackAppdomainunloadissues(freealltheobjectsintheappdomain)Allocationofnon-movingobjects(internedstrings.
.
.
)Nicetotheuser(nobigmemorychunk)StillalloweasyembeddingofMonoKeepingtheonlyreferencetoamanagedobjectintheCstackisfineThreadsupport6ImplementationGenerationalOldgenerationandnursery(thenewgeneration)MovingFromthenurserytotheoldgenerationCompactingoftheoldgenerationStoptheworldcollectionLargeobjectspaceCollectedwithmark&sweepduringmajorcollectionsFixedheapForinternedstrings(andmaybelaterothertypes)Collectedwithmark&sweepduringmajorcollections7GetHashCode()Objectsnowcanmove,can'tusetheobjectaddressashashcodeDatastoredinthelockwordoftheheaderWhentheobjectislocked,thehashcodeismovedinsidethefatlockstructureLoad+check+shiftinthefastpathAdditionalcheck+loadotherwiseInterlockedopwhenfirstsettingthehashmono_object_hash()forruntimehashtables8PinnedobjectsHowtocreatethem:corlibAPI:GCHandleembeddingAPI:mono_gchandle_new(obj,TRUE)fixedC#statement(PINNEDflagforlocalILvars)duringP/InvokecallsAdditionalcasesobjectsandpointersfoundintheuntypedCstackoftheruntimeandembeddingapplicationsobjectstheruntimesdoesn'twanttomove(internedstrings,Threadobjects,currentlyTypehandles)Fewoftheminpractice9PinnedobjectissuesWillfragmenttheheapWillcausemorecollectionsWillslowdownallocationsandcollectionsMakesureyoupinobjectsonlywhenreallyneededandforaslittletimeaspossibleRuntimeenhancements:TypeinformationaboutregistersandmanagedstackframeswillallowtoconsiderthereferencedobjectsnotpinnedUnmanagedstackframeswillstillpinreferencedobjectsforhassle-freeruntimecodeandmonoembedding10Pinnedobjects:findingthemGivena(possiblyinterior)pointerObjectsareallocatedsequentiallyNeedtostartfromthebeginningofanheaparea,visitingeachobjectuntilthecorrectoneisfoundDuringallocation,every4KBsavetheobjectstart–Needtoscanatmost4KBofmemorytofindobject–4KBsizeissubjecttochange–ThepinningaddressesaresortedandthelastpinnedobjectiscachedsousuallythisscanningisveryfastWhenfoundsetthePINflagoraddtopinnedobjectsarray11ThenurseryMostobjectsbornhereTypically512KB-2MBforcommonappsServersmightwantseveralMbsBump-pointerallocationstyleFastandcanbeinlinedinmanagedcodeWhenfull,triggeraminorcollectionUnlessamajorcollectionisneeded(oldgenerationisfull,too)Dividedinsmallerchunksduetofragmentationorthread-localallocation12NurseryandpinningPinnedobjectsfragmentthenurseryChunksoffreenurseryspacebetweenpinnedobjectsusedforallocationsWhenthenurseryiscompletelypinned,allocateobjectsintheoldgenerationAlternativesinclude:enlargethenursery,allocateanewnurseryPinobjectsonlywhenaboslutelyneededforshortamoutsoftimeTrackreferencespreciselyonthestack13MinorcollectionStoptheworldIdentifypinningobjectsScantheroots(includingpinnedobjects)CopytooldgenerationasyougoPlacepointertonewcopyintheoldobject'splaceScanthecopiedobjects(theyareroots,too)Checkthefinalizationandweak-reflistsPreparefornewallocsRestarttheworldPokethefinalizerthread14StoppingtheworldNeededtonotallowthemutatorstoseenon-coherentdatainobjectsforwardingpointersflagsintheobjectheadertwocopiesofthesameobjectNosupportforsafepoints(yet)Implementedwithsignalswin32/OSXhaveproperOSsupportLessthan1msforheapsupto100MBon1.
6PMParallelizemajorcollectionlater15TherootsStaticmanagedvariablesRegisteredroots(runtimeandembeddingdatastrutcures)HandletablesRememberedsetsOld-generationobjectsreferencingnurseryobjectsPinnedobjectsRuntimestackNeedtypeinfoforregistersandmanagedstackframes16HowtocopyanobjectEnsurewehaveenoughroomintheoldgenerationWeneedtoensureeveryreferencepointstothenewobjectForeachpointerfieldobj.
field=copy_object(obj.
field);Placeaforwardingpointerintheheader–It'sthepointertothenewhomeoftheobjectIfanobjecthasaforwardingpointerit'salreadycopiedPinnedobjectsreturnthesamepointerincopy_object()17ThemarkingstackWeneedtorecursivelycopytheobjectsreferencedbycopiedobjectRecursionnotasafethingintheGCTheruntimestackcouldbeverysmallandtherecursionverydeepThecopiedobjectsareaisanexplicitstackOncetherootsarescanned,scanthecopiedarea,untilnomoreobjectsarecopiedGrayobjectsaremarkedbutthefieldshavenotbeentraced18FinalizationWhentherootsarescanned,thenon-copiedobjectsaregarbageunlesstheyarepinnedorneedfinalizationCopyfinalizableobjectstotheoldgenerationsotheysurviveRecursivelycopyfromtheirfields,tooLoopuntilnomorefinalizableobjectsarefounddeadPutfinalizableobjectsinaseparatelistWeak-refsdon'tneedcopyinghandledimmediately19PreparefornewallocsCreatelistoffreefragmentsinnurserymemsetthememoryto0bytestoomuchoverheaddoingitateachobjectalloctoomuchoverheaddoingitforthewholenursery–touchestoomuchmemoryandtrashesthedatacache–doitonlyafragmentatatimeweneedunusednurseryareastobezeroedbeforecollectionsforthepinned-objectsfindingalgorithmAssignfragmentstothreadstoeachasmuchastheyneed(andresourcesallow)20MajorcollectionCurrentlycopy-basedLaterusemark/copytoreducememoryrequirementsIdentifypinnedobjectsinthewholeheapRememberedsetsareclearedScantherootsasusualLargeandfixedobjectsarejustmarkedbysettingaflagintheobjectheaderSweepunmarkedobjectsFreeunusedsectionsFinalize/Prepare/Restart21ObjectsweepingWalkthelistoflargeobjectsandfreetheunmarkedonesUnsetthemarkflagfortheothersFixedobjectsareinpinnedchunksfreedbyputtinginafree-listunmarkedaswellBoththelargeobjectlistandthenumberoffixedobjectsshouldbesmall22MemorysectionsHeapdividedinsectionsMoreflexiblethanusingabigvirtualareaEasierfortheuserExpandsasneededFairtoothercodeinthesameprogramcompetingforvirtualaddressspaceOfcourseit'shardertoimplementsomeperformanceissueslikelythedefaulton64bitsystems(hugevirtualmemoryspace)23PinnedchunksDatastructureforGC-internaldataandforfixedobjectallocationOnlyforsmallobjectsEachpagehasobjectsofsamesizeAllocationusingfreelistsPageassignedtoasizeasneededLimitedfragmentationissues(fewsizes,mostlyruntimeandGC-controlled)Mini-BoehmGCincluded24LargeobjectsExpensivetomoveCurrentlyusedforobjectsize>=64KBCollectedonlyduringmajorcollectionsusingmarkandsweepAllocatedwithmmapalreadycleared,noneedtotouchpagesNeedfortuningknobshowmuchmemorydoweallowinlargeobjectsbeforeamajorcollection25ObjectlayoutKeepreferencesclosetoeachotherBettercachelocalityBetterencodingofreferencebitmapsDegeneratecases.
.
.
Doctor,ithurtsWeshouldsetreasonablelimitsforuntrustedcodeTheJVMdoeshaveamuchbettertimewiththis.
.
.
nostructsnofixedorsequentiallayout26GCdescriptorsSeveraltypesneeded:stringsvectorsbitmapsrun-lengthencodinglargebitmapsFasthandlingofptrfreeobjectsandarraysStoredinMonoVTable27WritebarriersKeeptrackofreferencestonurseryobjectsfromtheoldgeneration:theybecomerootsforaminorcollectionComplicatedbystructsareferencecanbestoredintheheaporonthestackcopyofstructsCustomonesoptimizedforeachcasearraycopystructhandling28API/ABIchangesAfewsimplechangesneededforruntimehackersandembeddersObjectscanmove,sotheGCneedstoknowaboutallreferencesortheymustbepinnedsomeway–NomorevalidtoknowanobjectiskeptaliveandstoreareferenceinmallocmemoryFieldandarrayelementsettingmustgothroughwritebarriers–OnlyreferencesorvaluetypeswithreferencesInteriorpointerissues–ApointertotheendofanobjectisnotavalidinteriorpointerNewAPI:–MONO_OBJECT_SETREF(object,fieldname,value)–mono_array_setref(array,index,value)29TODOlistUsemark-compactforoldgenerationcollectionReducesmemoryrequiredforoldgencollectionUsemark-tablebasedwritebarriersFasterandinlineableThread-localallocationRemoveslockoverheadInlineallocationinjittedmanagedcodeNomanagedunmanagedtransitionsPrecisestackwalkReduceriskoffalsepositiveGCreferencesFixtheruntime30PreliminaryresultsStilluntuned,soexpectimprovementswritebarriers,thread-localalloc,alloccodeinliningSome(rare:-)testcases:5-7xspeedupUsually(withthread-local-alloc):10-50%speedupUsuallysmallerheap(butminimumislarger)Degeneratecases:manypinnedobjectsmanyreferencestores(writebarriers)longlinkedlists31KernelhelpThreadstart/stopneedtobeabletogetcontextofstoppedthreadmprotect(addr,size,MAP_DISCARD);dropthepagesfrommemorywillbeclearedwhenreadagainPagetablesdirtybitaccessMorecoarsewritebarriersupportNoneedforwritebarriersFixed(andpossiblytoolarge)markwindow32CompactingGarbageCollector.
33CompactingCollector34Nursery.
35NurseryandPinnedObjects.