Nested OpenMP - National Energy Research Scientific ... · • Hopper: NERSC Cray XE6 ... Courtesy of Robert Hager ... • Nested OpenMP can be used to enable mulple threads for

Helen He!NUG Meeting, 10/08/2015

Nested OpenMP

OpenMP Execution Model

•  ForkandJoinModel– Masterthreadforksnewthreadsatthebeginningofparallelregions.

– Mul6plethreadsshareworkinparallel.–  Threadsjoinattheendoftheparallelregions.

-2-

Hopper/Edison Compute Nodes

-3-

•  Hopper:NERSCCrayXE6,6,384nodes,153,126cores.•  4NUMAdomainspernode,6coresperNUMAdomain.

•  Edison:NERSCCrayXC30,5,576nodes,133,824cores.•  2NUMAdomainspernode,12coresperNUMAdomain.

2hardwarethreadspercore.•  Memorybandwidthisnon-homogeneousamongNUMAdomains.

MPI Process Affinity: aprun “-S” Option •  Processaffinity:orCPUpinning,bindsMPIprocesstoaCPUorarangesof

CPUsonthenode.•  ImportanttospreadMPIranksevenlyontodifferentNUMAnodes.•  Usethe“-S”opWonforHopper/Edison.

-4-

0200400600800

100012001400

24576*1 12288*2 8192*3 4096*6 2048*12

RunTime(sec)

MPITasks*OpenMPThreads

GTCHybridMPI/OpenMPonHopper,24,576cores

with-S-ss

no-S-ss

aprun–n4–S1–d6

aprun–n4–d6

Lower is B

etter

-S2–d3

Thread Affinity: aprun “-cc” Option

•  Threadaffinity:forceseachprocessorthreadtorunonaspecificsubsetofprocessors,totakeadvantageoflocalprocessstate.

•  Threadlocalityisimportantsinceitimpactsbothmemoryandintra-nodeperformance.

•  OnHopper/Edison:–  Thedefaultop6onis“-cccpu”(useitfornon-Intelcompilers),

bindseachPEtoaCPUwithintheassignedNUMAnode.–  PayaWen6ontoIntelcompiler,whichusesanextrathread.

•  Use“-ccnone”if1MPIprocesspernode•  Use“-ccnuma_node”(Hopper)or“-ccdepth”(Edison)ifmul6pleMPIprocessespernode

-5-

NERSC KNC Testbed: Babbage

-6-

•  NERSCIntelXeonPhiKnightsCorner(KNC)testbed

•  45computenodes,eachhas:–  Hostnode:2IntelXeon

Sandybridgeprocessors,8coreseach.

–  2MICcardseachhas60na6vecoresand4hardwarethreadspercore.

–  MICcardsaWachedtohostnodesviaPCI-express.

–  8GBmemoryoneachMICcard

•  Recommendtouseatleast2threadspercoretohidelatencyofin-orderexecuWon.

TobestpreparecodesonBabbageforCori:•  Use“na6ve”modeonKNCtomimicKNL,

whichmeansignorethehost,justruncompletelyonKNCcards.

•  Encouragetoexploresinglenodeop6miza6onforthreadingscalingandvectoriza6ononKNCcardswithproblemsizesthatcanfit.

•  “Symmetric”,“Offload”modesonKNCand“OpenMP4.0target”work,butarenotourpromotedusagemodelsforBabbage.

Babbage MIC Card

-7-

Babbage:NERSCIntelXeonPhitestbed,45nodes.2MICcardspernode.•  1NUMAdomainperMICcard:60physicalcores,240logicalcores.OpenMP

threadingpotenWalto240-way.•  KMP_AFFINITY,KMP_PLACE_THREADS,OMP_PLACES,OMP_PROC_BINDfor

threadaffinitycontrol•  I_MPI_PIN_DOMAINforMPI/OpenMPprocessandthreadaffinitycontrol.

Full OpenMP 4.0 Support in Compilers

•  GNUcompiler–  From4.9.0forC/C++–  Fromgcc/4.9.1forFortran

•  Intelcompiler–  Fromintel/15.0:supportsmostfeaturesinOpenMP4.0;

FromIntel/16.0:fullsupport

•  Craycompiler–  Fromcce/8.4.0

-8-

Thread Affinity Control in OpenMP 4.0 •  OMP_PLACES:alistofplacesthatthreadscanbepinnedon

–  threads:Eachplacecorrespondstoasinglehardwarethreadonthetargetmachine.

–  cores:Eachplacecorrespondstoasinglecore(havingoneormorehardwarethreads)onthetargetmachine.

–  sockets:Eachplacecorrespondstoasinglesocket(consis6ngofoneormorecores)onthetargetmachine.

–  Alistwithexplicitplacevalues:suchas:•  "{0,1,2,3},{4,5,6,7},{8,9,10,11},{12,13,14,15}”•  “{0:4},{4:4},{8:4},{12:4}”

•  OMP_PROC_BIND–  spread:Bindthreadsasevenlydistributed(spread)aspossible–  close:Bindthreadsclosetothemasterthreadwhiles6lldistribu6ng

threadsforloadbalancing,wraparoundonceeachplacereceivesonethread

–  master:Bindthreadsthesameplaceasthemasterthread

-9-

Nested OpenMP Thread Affinity Illustration

-10-

setenvOMP_PLACESthreadssetenvOMP_NUM_THREADS4,4setenvOMP_PROC_BINDspread,close

spread

close

Sample Nested OpenMP Code #include<omp.h>#include<stdio.h>voidreport_num_threads(intlevel){#pragmaompsingle{prinl("Level%d:numberofthreadsintheteam:%d\n",level,omp_get_num_threads());}}intmain(){omp_set_dynamic(0);#pragmaompparallelnum_threads(2){report_num_threads(1);#pragmaompparallelnum_threads(2){report_num_threads(2);#pragmaompparallelnum_threads(2){report_num_threads(3);}}}return(0);}

-11-

%a.outLevel1:numberofthreadsintheteam:2Level2:numberofthreadsintheteam:1Level3:numberofthreadsintheteam:1Level2:numberofthreadsintheteam:1Level3:numberofthreadsintheteam:1

%setenvOMP_NESTEDTRUE%a.outLevel1:numberofthreadsintheteam:2Level2:numberofthreadsintheteam:2Level2:numberofthreadsintheteam:2Level3:numberofthreadsintheteam:2Level3:numberofthreadsintheteam:2Level3:numberofthreadsintheteam:2Level3:numberofthreadsintheteam:2

Level0:P0Level1:P0P1Level2:P0P2;P1P3Level3:P0P4;P2P5;P1P6;P3P7

When to Use Nested OpenMP

•  SomeapplicaWonteamsareexploringwithnestedOpenMPtoallowmorefine-grainedthreadparallelism.–  MPI/Hybridnotusingnodefullypacked–  ToplevelOpenMPloopdoesnotuseallavailablethreads–  Mul6plelevelsofOpenMPloopsarenoteasilycollapsed–  Certaincomputa6onalintensivekernelscouldusemorethreads–  MKLcanuseextracoreswithnestedOpenMP

-12-

Process and Thread Affinity in Nested OpenMP

•  AchievingbestprocessandthreadaffinityiscrucialingernggoodperformancewithnestedOpenMP,yetitisnotstraighlorwardtodoso.

•  AcombinaWonofOpenMPenvironmentvariablesandrunWmeflagsareneededfordifferentcompilersanddifferentbatchschedulersondifferentsystems.

-13-

Example:UseIntelcompilerwithTorque/MoabonEdison:setenvOMP_NESTEDtruesetenvOMP_NUM_THREADS4,3setenvOMP_PROC_BINDspread,closeaprun-n2-S1-d12–ccnuma_node./nested.intel.edison

Edison: Run Time Environment Variables •  setenvOMP_NESTEDtrue

–  Defaultisfalseformostcompilers•  setenvOMP_MAX_ACTIVE_LEVELS2

–  Thedefaultwas1forCCEpriortocce/8.4.0•  setenvOMP_NUM_THREADS4,3•  setenvOMP_PROC_BINDspread,close•  setenvKMP_HOT_TEAMS1

–  Intelonlyenv.Defaultisfalse•  setenvKMP_HOT_TEAMS_MAX_LEVELS2

–  Intelonlyenv.AllownestedlevelOpenMPthreadstostayaliveinsteadofbeingdestroyedandcreatedagaintoreducethreadcrea6onoverhead.

•  aprun-n2-S1-d12–ccnuma_node./nested.intel.edison–  Use-dfortotalnumberofthreads(productsofnum_threadsfromall

levels).–ccnuma_nodetoallowthreadsmigratewithinNUMAnodetonotaffectedbyIntel’sextramanagerthread.

-14-

Babbage: Run Time Environment Variables •  SetI_MPI_PIN_DOMAIN=autotogetbasicMPIprocessaffinity•  DonotsetKMP_AFFINITY,otherwiseOMP_PROC_BINDwill

beignored.•  UseOMP_PLACES=threads(default)insteadofsockets•  setenvOMP_NESTEDtrue•  setenvOMP_NUM_THREADS4,3•  setenvOMP_PROC_BINDspread,close•  setenvKMP_HOT_TEAMS1•  setenvKMP_HOT_TEAMS_MAX_LEVELS2•  mpirun.mic-n2-hostbc1109-mic0./xthi-nested.mic|sort

-15-

XGC1: Nested OpenMP •  Alwaysmakesuretousebestthreadaffinity.Avoidusingthreadsacross

NUMAdomains.•  Currently:

•  Isabitslowerthan(workongoing):

•  Willtry:

•  Usenum_threadsclauseinsourcecodestosetthreadsfornestedregions.

Formostothernon-nestedregions,useOMP_NUM_THREADSenvforsimplicityandflexibility.

-16-

exportOMP_NUM_THREADS=6,4exportOMP_PROC_BIND=spread,closeexportOMP_NESTED=TRUEexportOMP_STACKSIZE=8000000aprun-n200-N2-S1-j2-ccnuma_node./xgca

exportOMP_NUM_THREADS=24exportOMP_NESTED=TRUEexportOMP_STACKSIZE=8000000aprun-n200-d24-N2-S1-j2-ccnuma_node./xgca

exportKMP_HOT_TEAMS=1exportKMP_HOT_TEAMS_MAX_LEVELS=2

CourtesyofRobertHager,PPPLandNESAPXGC1team.

Use Multiple Threads in MKL

•  ByDefault,inOpenMPparallelregions,only1threadwillbeusedforMKLcalls.–  MKL_DYNAMICSistruebydefault

•  NestedOpenMPcanbeusedtoenablemulWplethreadsforMKLcalls.TreatMKLasanestedinnerOpenMPregion.

•  Sampleserngs

-17-

exportOMP_NESTED=trueexportOMP_PLACES=coresexportOMP_PROC_BIND=closeexportOMP_NUM_THREADS=6,4exportMKL_DYNAMICS=falseexportKMP_HOT_TEAMS=1exportKMP_HOT_TEAMS_MAX_LEVELS=2

NWChem: OpenMP “Reduce” Algorithm

-18-

•  PlanewaveLagrangemulWplier–  Manymatrixmul6plica6onsofcomplexnumbers,C=AxB–  Smallermatrixproducts:FFM,typicalsize100x10,000x100–  OriginalthreadingscalingwithMKLnotsa6sfactory

•  OpenMP“Reduce”or“Block”algorithm-  DistributeworkonAandBalongthekdimension-  Athreadputsitscontribu6oninabufferofsizemxn-  BuffersreducedtoproduceC-  OMPteamsofthreads

FFM

CourtesyofMathiasJacquelin,LBNL

NWChem: OpenMP “Reduce” Algorithm •  Bezerforsmallerinnerdimensions,i.e.forFFMs•  MulWpleFFMscanbedoneconcurrentlyindifferentthreadpools•  Threadingenablesustouseall240hardwarethreads•  Best“Reduce”:10MPI,6teamsof4threads(nestedOpenMPwithMKL)

-19-

MKL1MPI,240threads

Best“Reduce”10MPI,6teamsof4threads

CourtesyofMathiasJacquelin,LBNL

FFT3D on KNC, Ng=643

-20-

CourtesyofJeongnimKim,Intel

Nested OpenMP on NERSC Systems

•  Pleaseseedetailedexampleserngsinthe“NestedOpenMP”webpage:–  RunonEdisonandBabbage– WithIntelandCraycompilers–  UseTorque/MoabandSLURMbatchschedulers–  hWps://www.nersc.gov/users/computa6onal-systems/edison/running-jobs/using-openmp-with-mpi/nested-openmp/

-21-

Thank you.

-22-