Eﬀec%ve Soware - cvut.cz · 2018. 2. 25. · » JVM is used to implement also other languages...

David Šišlák [email protected]

Effec%veSo*ware

Lecture2:Virtualmachine,JVM,bytecode,(de-)compilers,disassembler,profiling

26thFebruary2018 ESW–Lecture2 2

Introduc%on–VirtualMachine

»  Virtualmachinemodel(.NET,JVM–Scala,Jython,JRuby,Clojure,…)•  sourcecode•  compiledintoVMbytecode•  hybridrun-Omeenvironment(plaQormdependentVMimplementaOon)

–  interpretedbytecode–  compliedassembly-code(naOveCPUcode)–  automatedplaQormcapabilityopOmizaOons(e.g.useofSIMD)

»  comparisonofbytecodetoassembly-code•  (+)plaQormindependence(portable)–architecture(RISC/CISC,bits),OS•  (+)reflecOon–observe,modifyownstructureatrun-Ome•  (+)smallsize•  (-)slowerexecuOon–interpretedmode,compilaOonlatencies•  (-)lesscontrolonassemblycode–lessopOonsforcustomopOmizaOon


JAVAVirtualMachine–MemoryLayout

Thread specific Shared by many threads


JAVAVirtualMachine–MemoryLayout


JAVAVirtualMachine–Stack-orientedMachine

»  stack-oriented-stackmachinemodelforpassingparametersandoutputforinstrucOons

(2+3)×11+1»  JVMbytecode–sequenceofinstruc(onscomposedof

•  opcode–operaOoncode,whatshouldbedone•  opcodespecificparameters–somehasnoparams,somemulOple


JAVAVirtualMachine-Frame

»  frame»  eachthreadhasstackwithframes(outsideofheap,fixedlength)

StackOverflowErrorvs.OutOfMemoryError»  frameiscreatedeachOmemethodisinvoked(destroyedaderreturn)

-  interpretedframeperexactlyonemethod-  compliedframeincludesallin-linedmethods

»  framesizedeterminedatcompile-Ome(inclassfileforinterpreted)»  variables(anytype)

»  {this}–instancecallonly!»  {methodparameters}»  {localvariables}

»  operandstack(anytype)»  LIFO

»  referencetorun-%meconstantpool(classdef)»  method+classisassociated


JAVAVirtualMachine–Opcodes

»  JVMopcode(1Byteonlyalways):»  loadandstore(aload_0,istore,aconst_null,…)»  arithmeOcandlogic(ladd,fcmpl,…)»  typeconversion(i2b,d2i,…)»  objectmanipulaOon(new,puQield,geQield,…)»  stackmanagement(swap,dup2,…)»  controltransfer(ifeq,goto,…)»  methodinvocaOon(invokespecial,areturn,…)–framemanipulaOon»  excepOonsandmonitorconcurrency(athrow,monitorenter,…)

»  prefix/suffix–i,l,s,b,c,f,danda(reference)»  variablesasregisters–e.g.istore_1(variable0isthisforinstancemethod)

VS.

CPU assembly-code JVM bytecode


JAVAVirtualMachine

»  JVMisusedtoimplementalsootherlanguagesthanJAVA»  Erlang->Erjang»  JavaScript->Rhino»  Python->Jython»  Ruby->Jruby»  Scala,Clojure–funcOonalprogramming»  others

»  bytecodeisverifiedbeforeexecuted:»  branches(jumps)arealwaystovalidlocaOons–onlywithinmethod»  anyinstrucOonoperatesonafixedstacklocaOon(helpsJITfor

registersmapping)»  dataisalwaysiniOalizedandreferencesarealwaystype-safe»  accesstoprivate,packageiscontrolled


JAVAVirtualMachine–ObjectOrientedLanguage

»  Classfile–productofsourcecodecompilaOon•  onepereachclass•  methodbytecodeisincluded


JAVAVirtualMachine–Example1–SourceCode


JAVAVirtualMachine–Example1–ClassFileContent


JAVAVirtualMachine–Example1–DisassembledConstants

»  javap–JAVAdisassemblerincludedinJDK


JAVAVirtualMachine–Example1–DisassembledFields


JAVAVirtualMachine–Example1–DisassembledMethod

»  geQield•  takes1reffromstack•  buildanindexintorunOmepoolofclassinstancebyreferencethis

»  areturn•  takes1reffromstack•  pushontothestackofcallingmethod

opcode offset in bytecode for the method employeeData


JAVAVirtualMachine–Example1–DisassembledConstructor


JAVAVirtualMachine–Example1–Decompiler

»  procyon–open-sourceJAVAdecompiler

Original source code De-compiled source code


JAVAVirtualMachine–Example2–SourceCode


JAVAVirtualMachine–Example2–daysInMonthBytecode


JAVAVirtualMachine–Example2–computeBytecode

No optimization during source code compilation !


JAVAVirtualMachine–SourceCodeCompila%on

»  sourcecodecompila%on(sourcecode=>bytecode)»  bytecodeisnotbeperthanyoursourcecode

»  invariantsinlooparenotremoved»  noopOmizaOonslike

»  loopunrolling»  algebraicsimplificaOon»  strengthreducOon

»  opOonallybytecodecouldbemodifiedbeforeexecuOonbyJVM•  e.g.ProGuard–obfuscatorincludingbytecodeopOmizaOons

–  shrinker–compactcode,removedeadcode–  opOmizer

•  modifyaccesspapern(private,staOc,final)•  inlinebytecode

–  obfuscator–renaming,layoutchanges–  preverifier–ensureclassloading

Test yourself - compute method is simplified - faster interpretation - better JIT output

obfuscation = make code difficult to be understood by humans but with the same functionality


JAVAVirtualMachine–BytecodeCompila%oninrun-%me

»  Just-in-%me(JIT)»  convertsbytecodeintoassemblycodeinrun-Ome»  checkOpenJDKsourcesforverydetailedinformaOon

hpp://openjdk.java.net»  JITincludesadap%veop%miza%on(adapOveOeredcompilaOonsinceversion7)

»  balancetrade-offbetweenJITandinterpreOnginstrucOons»  monitorsfrequentlyexecutedparts“hotspots”includingdataoncaller-callee

relaOonshipforvirtualmethodinvocaOon»  triggersdynamicre-compilaOonbasedoncurrentexecuOonprofile»  inlineexpansiontoremovecontextswitching»  opOmizebranches»  canmakeriskyassumpOon(e.g.skipcode)->

»  unwindtovalidstate»  deopOmizepreviouslyJITedcodeevenifcodeisalreadyexecuted

»  Ahead-of-TimeCompilaOon(AOT)–removewarm-upphase•  compileintoassemblycodepriortolaunchingthevirtualmachine


JAVAVirtualMachine–JITCompila%on

»  Just-in-6me(JIT)compilers–asynchronous(3C1,7C2threadsfor32cores)»  C1compiler–muchfasterthanC2

»  simplifiedinlining,usingCPUregistry»  window-basedopOmizaOonoversmallsetofinstrucOons»  intrinsicfuncOonswithvectoroperaOons(Math,arraycopy,…)

»  C2compiler–high-endfullyopOmizingcompiler»  deadcodeeliminaOon,loopunrolling,loopinvarianthoisOng,commonsub-expressioneliminaOon,constantpropagaOon

»  fullinlining,fulldeopOmizaOon(backtolevel0)»  escapeanalysis,nullcheckeliminaOon,»  papern-basedloopvectorizaOonandsuperwordpacking(SIMD)

»  JITcompila%on%ers

»  on-stackreplacement(OSR)–opOmizaOonduringexecuOonofamethod»  startatbytecodejumptargets(goto,if_)


AssemblyCode

»  reasonstostudyassemblycode(bothJavaandC/C++)•  educaOonalreasons

–  predictefficientcodingtechniques•  debuggingandverificaOon

–  howwellthecodelookslike•  opOmizecode

–  forspeed•  avoidpoorlycompiledpaperns•  datafitsintocache•  predictablebranchesornobranches•  usevectorprogramingifpossible(SIMD)

»  256bitregisterswithAVX2sinceIntelSandyBridge»  512bitAVX-512sinceIntelKnightLanding(XeonPhi)

–  forsize•  primarilycodecacheefficiency


JAVAVirtualMachine–Example2–TieredCompila%on

»  -XX:+PrintCompilaOon(-XX:+PrintInlining)

{millis from start} {compilation_task_id} {flags} {tier} {class:method} (bytecode size)@OSR {removing not rentrant/zombie}

Notice standard compilation path 0 -> 3 -> 4


JVM–Example2–daysInMonthAssemblyCode–Tier3

»  -XX:+UnlockDiagnosOcVMOpOons-XX:+PrintAssembly»  allexamplesareinJVM864-bit,IntelHaswellCPU,AT&Tsyntax%er3-C1withinvoca%on&backedgecounters+MethodDataOopcounter

because:count="256"iicount="256”hot_count="256”stackiniOalizaOon,invoca%oncounterinMDO(0xDC)+triggerC2(Oer4)

0x1ff8 >> 3 = 1024 invocations trigger tier 4 (C2)

month, year stacking banging technique, StackOverflowException stack allocation, saving registers



ESI is month input

default jump



targetformonth=4,backedgecountertrackinginMDO(0x290):jumptarget,inlinedTLABalloca%onofIntegerobject:

no space in TLAB -> new TLAB + external allocation with header init returns after the inlined allocation

EBX=30 is retVal

RAX Integer instance address Object structure (64-bit JVM): -  header 12 or 16 Bytes -  object data super class first, type grouped

8B - mark word

4B / 8B – Klass ref.

… object data

Array object structure (64-bit JVM): -  header 16 or 20 Bytes -  sequence of array values

8B - mark word

4B / 8B – Klass ref.

sequence of values

4B – array length

0x10 Integer instance size

object initialization, header filed with prototype mark



inlinedIntegerconstructorwithsupers,invocaOoncountsinMDOs(0xDC) Integer::<init>,Number::<init>,Object::<init>

-currentlyinOer3(C1countersinMDO) invocation cnt of Integer::<init> in daysInMonth for inline

invocation cnt in Integer::<init> + trigger its C2 (tier 4)

invocation cnt of Number::<init> in Int::<init> for inline

invocation cnt in Number::<init> + trigger its C2 (tier 4)

invocation cnt of Object::<init> in Numb::<init> for inline

invocation cnt in Object::<init> + trigger its C2 (tier 4)

RAX.value = EBX (retVal)



finalcleanupandreturn,RAXcontainsreturnvalue(pointertoIntegerinstance)»  OrdinaryObjectPointer(Oop)–flexiblereferencetoanobject»  safepoint–OopsinperfectlydescribedstatebyOopMap(GCmaps)

•  Oopcanbesafelymanipulatedexternallywhilethreadissuspended•  ininterpretedmode–betweenany2bytecodes•  inC1/C2compiled–endofallmethods(notin-lined),non-countedloopbackedge,

duringJVMrun-Omecall•  parked,blockedonIO,monitororlock•  whilerunningJNI(donotneedthreadsuspension)•  globalsafepoint(allthreads)–stoptheworld

–  GC,printthreads,threaddumps,heapdump,getallstacktrace–  enableBiasedLocking,RevokeBias–  classredefiniOon(e.g.instrumentaOon),debug

•  localsafepoint(justexecu%ngthread)–  de-opOmizaOon,enable/revokebiaslocking,OSR

stack dealocation, reload register safepoint poll check


JVM–TimeToSafePoint

»  TimeToSafePoint(TTSP)–howlongittakestoentersafepoint-XX:+PrintSafepointStaOsOcs-XX:+PrintGCApplicaOonStoppedTime-XX:PrintSafepointStaOsOcsCount=1

TTSPoverheadinprofilerwhilecallingGetStackTraceexamplewith5threads:

TTS

P



%er4–C2compiler–noprofilecountersbecause:count="5376"iicount="5376”hot_count="5376”

stackiniOalizaOon,uselookuptablejumpfortableswitch

default (>=12)

month, year



targetformonth=4Integer.<init>,Number.<init>,Object.<init>-iicount=“5376”->Inline(hot)op%mizedbranching,inlinedTLABalloca%on,inlinedconstructors,nonulling,cachingop%miza%on

EBP=30 is retVal

TLAB Integer object allocation, ref in RAX

MarkWord fetch from class and then store compressed OOP to Integer class

RAX.value = EBX (retVal)

final cleanup

RAX contains return value (pointer to Integer instance)

cache optimization 3 cache lines ahead



targetfordefaultclassIllegalArgumentExcepOonnoprofile->uncommon->reinterpretremapinputs,returnbacktoreinterpreterthendiscardOer3version


JVM–Example2–computeAssemblyCode–Tier4OSR

OSR@10–OnStackReplacementatbytecode10%er4–C2(beforetherewasOer3OSR@10because60416loopsandOer3)

because:backedge_count=”101376"hot_count=”101376”copy4localsonstackfromOer3OSR@10toregs

RSI compiled stack of tier 3 OSR @10


JVM–Example2–computeAssemblyCode–Tier4OSR

loopcriteriathenthereisinlinedOer4daysOfMonth(lookupjump)becausethecallishotendingwithaddiOonintoaccumulatororeinterpretonendofcyclejump(unstableif_bytecode),save3localstostack

EBX is local I; 0xF4240 = 1_000_000


JVM–Example2–computeAssemblyCode–Tier4

%er4–C2because:count=”2”backedge_count=”150528”

usecombinaOonoffullinline,deadcodeelimina%on,objectescape,loopinvarianthois%ng,strengthreduc%on

30_000_000

RAX contains return value (primitive int)


JavaVirtualMachine–Performance

»  requireswarm-uptou%lizebenefitsofC2(orC1)»  compilerscannotdoallmagic->writebegeralgorithms

»  32-bitvs64bitsJVMs•  32-bit(max~3GBheap)

–  smallermemoryfootprint–  slowerlong&doubleoperaOons

•  64-bitmax32GBvirtualmemory(withdefaultObjectAlignmentInBytes)-  fasterperformanceforlong&double–  slightincreaseofmemoryfootprint–  compressedOOPsareslightlyslowerforreferencesuponusage–  compressedOOPslessmemory->lessfrequentGC->fasterprogram

•  64-bit>32GBvirtualmemory(largeheap)–  fastreferenceusage–  wasOngalotofmemory(48GB~32GBwithcompressedOOPs)


JavaVirtualMachine–CPUandMemoryProfiling

»  profiling•  CPU–Omespentinmethods•  memory–usage,allocaOons

»  modes•  sampling

–  periodicsamplingofstacksofrunningthreadstoesOmateslowest–  noinvocaOoncounts,no100%accuracy(varioussamplingerrors)–  nobytecode(&assemblycode)modificaOons–  1-2%impacttostandardperformance(TTSP,threaddumps,analysis)

•  tracing(instrumetaOon)-methodentry,exit,traceObjAllocaOons–  instrumentedbytecode->affectedperformance->affectedcompilerop%miza%ons

»  jvisualvm•  JVMmonitoring,troubleshooOngandprofilingtool•  includedinallJDKs•  profiledthreadlimit32


JVM–Example2–CPUTracingofdaysOfMonth

assemblycodeof%er4–C2(beforetherewasverycomplexOer3)inlineddaysInMonthrootMethodEntrytracking

749 Bytes of assembly code for each rootMethodEntry


JVM–Example2–CPUTracingofdaysOfMonth

addiOonalrootMethodEntryandrootMethodExittrackingsforInteger::<init>andNumber::<init>

inlinedrootMethodExitaderIntegerinstance.value=retVal

313 Bytes of assembly code for each rootMethodEntry


JVM–Example2–CPUTracingOutcome


JVM–Example2–ProfilingPerformance

»  CPUtracingofcomputeresultsintomuchslowercode•  noobjectescapefromdaysInMonthcall•  noinvarianthoisOng•  nostrengthreducOon(fullloopremainsthere)

»  objectallocaOonissimilarwithtraceObjAllocinjectedcalls

»  recommendedapproach•  dosamplingfirst•  idenOfyperformanceboplenecks(wheremostOmeisspent)

–  itcouldbeoutsideofJVM(e.g.latencyofexternalDB,filesystem)•  focuswithtracingjusttoidenOfiedparts


JVM–JavaMissionControl

jmc–JRockitJVM,includedincommercialJDKs,samplinginFlightrecorder


ApproachtoPerformanceTes%ng

»  testrealapplica%on–ideallythewayitisused•  microbenchmarks–measureverysmallunits

–  warm-up–tomeasurerealcode,notcompilersitself,biasedlocks•  keepinmindcaching

–  bewareofcompilers–useresults,reorderingofoperaOons–  synchronizaOon–mulO-threadedbenchmarks–  varypre-calculatedrightparametersaffecOngcomplexity–differentopOmizaOoninreality

•  macrobenchmarks–measureapplicaOoninput/output–  leastperformingcomponentaffectsthewholeapplicaOon

»  understandthroughput,elapsedandresponse%me•  outlierscanoccur–e.g.GC•  useexisOnggeneratorsthanwriOngown


ApproachtoPerformanceTes%ng

»  understandvariability–changesoverOme•  internalstate•  backgroundeffects–load,network•  probabilisOcanalysis–workswithuncertainty

»  testearly,testo*en–ideallypartofdevelopmentcycle•  ideallysomeproperlyrepeatedmesobenchmarking•  automatetests–scripted•  propertestcoverageoffuncOonalityandinputs•  testontargetsystem–differentcodeondifferentsystems

Eﬀec%ve Soware - cvut.cz · 2018. 2. 25. · » JVM is used to implement also other languages...

Documents

Transcript of Eﬀec%ve Soware - cvut.cz · 2018. 2. 25. · » JVM is used to implement also other languages...