多核辦理器Linux的進度綁定辦理器核運轉名詞CPUaffinity：中文稱作“CPU親和力”，是指在CMP架構下，能夠將一個或多個進度綁定到一個或多個辦理器上運轉。一、Linux代碼中綁定多核運轉1、假如自己寫代碼，要把進度綁定到CPU，該怎么做？能夠用sched_setaffinity函數(shù)。在Linux上，這會觸發(fā)一次系統(tǒng)調用。intsched_setaffinity(pid_tpid,unsignedintlen,unsignedlong*mask);sched_setaffinity的第一個參數(shù)是pid（進度ID），設置進度為pid的這個進度，讓它運轉在mask所設定的CPU上。假如pid的值為0，則表示指定的是目行進度，使目行進度運轉在mask所設定的那些CPU上；第二個參數(shù)cpusetsize是mask所指定的數(shù)的長度。往常設定為sizeof(cpu_set_t)；假如目前pid所指定的CPU此時沒有運轉在mask所指定的隨意一個CPU上,則該指定的進度會從其余CPU上遷徙到mask的指定的一個CPU上運轉。intsched_getaffinity(pid_tpid,unsignedintlen,unsignedlong*mask);該函數(shù)獲取pid所指示的進度的CPU位掩碼，并將該掩碼返回到mask所指向的構造中，即獲取指定pid目前能夠運轉在哪些CPU上。相同，假如pid的值為0.也表示的是目行進度。voidCPU_ZERO(cpu_set_t*set)這個宏對CPU集set進行初始化，將其設置為空集。voidCPU_SET(intcpu,cpu_set_t*set)這個宏將cpu加入CPU集set中。voidCPU_CLR(intcpu,cpu_set_t*set)這個宏將cpu從CPU集set中刪除。intCPU_ISSET(intcpu,constcpu_set_t*set)假如cpu是CPU集set的一員，這個宏就返回一個非零值（true），不然就返回零false）。Example：我是在一個虛構機上運轉的程序，機器CPU是雙核的，我設置虛構機模擬四核。在linux上履行top指令看結果，點擊“1”查察每個CPU核的運轉狀況。/*Shorttestprogramtotestsched_setaffinity(whichsetstheaffinityofprocessestoprocessors).Compile:gccsched_setaffinity_test.c*-osched_setaffinity_test.o-lmUsage:./sched_setaffinity_test.oOpena"top"-windowatthesametimeandseeallthework*beingdoneonCPU0firstandafterashortwaitonCPU1，2，3.Repeatwithdifferentnumberstomakesure,itisnotacoincidence.*/#include<stdio.h>#include<math.h>#include<sched.h>doublewaste_time(longn){doubleres=0;longi=0;while(i<n*200000){i++;res+=sqrt(i);}returnres;}intmain(intargc,char**argv){unsignedlongmask=1;/*二進制1,processor0*//*bindprocesstoprocessor0*/if(sched_setaffinity(0,sizeof(mask),&mask)<0){perror("sched_setaffinity");}/*wastesometimesotheworkisvisiblewith"top"*/printf("result:%f\n",waste_time(2000));mask=2;/*二進制10，processswitchestoprocessor1now*/if(sched_setaffinity(0,sizeof(mask),&mask)<0){perror("sched_setaffinity");}/*wastesomemoretimetoseetheprocessorswitch*/printf("result:%f\n",waste_time(2000));mask=4;/*二進制100，processor2*//*bindprocesstoprocessor2*/if(sched_setaffinity(0,sizeof(mask),&mask)<0){perror("sched_setaffinity");}/*wastesometimesotheworkisvisiblewith"top"*/printf("result:%f\n",waste_time(2000));mask=8;/*二進制1000，processswitchestoprocessor3now*/if(sched_setaffinity(0,sizeof(mask),&mask)<0){perror("sched_setaffinity");}/*wastesomemoretimetoseetheprocessorswitch*/printf("result:%f\n",waste_time(2000));}2、運轉結果圖2-1：圖中顯示編譯及履行一個使用“CPU親和力”的程序。圖2-2能夠看見CPU0已在履行程序（CPU達到100.0%？有必需嗎），第一段程序的結果還未出來。圖2-3能夠看見CPU1已在履行程序，第一段程序的結果已經顯示（圖片中顯示了上一次履行的結果，由于截圖比較慢，履行了3次^-^），CPU1履行的結果還未顯示出來。圖2-4：能夠看見CPU2已在履行程序，第二段程序（CPU1履行）的結果已經顯示，CPU2履行的程序結果還未出來。圖2-5：:CPU3再履行程序。圖2-6：結果已經顯示，CPU恢復到履行以前的狀態(tài)二、多進度（在兩個核上各自運轉一個進度）1、代碼以下/*Shorttestprogramtotestparallel_setaffinity_test.c(whichsetstheaffinityofprocessestoprocessors).Compile:gccparallel_setaffinity_test.c-oparallel_setaffinity_test.o-lmUsage:./parallel_setaffinity_test.o*Opena"top"-windowatthesametimeandseealltheworkbeingdoneonCPU0firstandafterashortwaitonCPU1.Repeatwithdifferentnumberstomakesure,itisnotacoincidence.*/#include<stdlib.h>#include<stdio.h>#include<sys/types.h>#include<sys/sysinfo.h>#include<unistd.h>#include<string.h>#include<math.h>#include<sched.h>doublewaste_time(longn){doubleres=0;longi=0;while(i<n*200000){i++;res+=sqrt(i);}returnres;}intmain(){charbuf[100];pid_tcld_pid;intfd;intstatus;unsignedlongmask=0;strcpy(buf,"Thisisparentprocesswrite!\n");if((cld_pid=fork())==0){strcpy(buf,"Thisischildprocesswrite!\n");printf("Thisischildprocess!\n");printf("MyPID(child)is%d\n",getpid());printf("MyPID(child)is%d\n",getppid());mask=1;/*二進制1,processor0*//*bindprocesstoprocessor0*/if(sched_setaffinity(0,sizeof(mask),&mask)<0){perror("sched_setaffinity");}/*wastesometimesotheworkisvisiblewith"top"*/printf("result:%f.\n",waste_time(2000));write(fd,buf,strlen(buf));close(fd);exit(0);}else{printf("Thisisparentprocess!\n");printf("MyPID(parent)is%d.\n",getpid());printf("MychildPIDis%d.\n",cld_pid);mask=2;/*二進制10，processswitchestoprocessor1now*/if(sched_setaffinity(0,sizeof(mask),&mask)<0){perror("sched_setaffinity");}/*wastesomemoretimetoseetheprocessorswitch*/printf("result:%f.\n",waste_time(2000));write(fd,buf,strlen(buf));close(fd);}wait(&status);return0;}2、運轉結果圖3-1：圖中顯示編譯及履行一個使用“CPU親和力”的多進度程序，雙核（虛構四核太慢）。圖3-2：，父子進度并發(fā)履行，父進度在CPU1上運轉，子進度運轉在CPU0，能夠看出兩個辦理器目前都在運轉程序。圖3-3：結果已經顯示，CPU0和CPU1都恢復到履行以前的狀態(tài)附錄1、Processoraffinityisamodificationofthenativecentralqueueschedulingalgorithminasymmetricmultiprocessingoperatingsystem.Eachtask(beitprocessorthread)inthequeuehasatagindicatingitspreferred/kinprocessor.Atallocationtime,eachtaskisallocatedtoitskinprocessorinpreferencetoothers.Processoraffinitytakesadvantageofthefactthatsomeremnantsofaprocessmayremaininoneprocessor'sstate(inparticular,initscache)fromthelasttimetheprocessran.Schedulingittorunonthesameprocessorthenexttimecouldresultintheprocess'srunningmoreefficientlybyreducingperformance-degradingsituationssuchascachemisses.Apracticalexamplemightberunningmultipleinstancesofanapplicationwhichdoesnotusemultiplethreads,suchassomegraphics-renderingsoftware.Overallsystemefficiencyincreases.Schedulingalgorithmimplementationsvaryinadherencetoprocessoraffinity.Undercertaincircumstancessomeimplementationswillallowatasktochangetoanotherprocessorifthisisdeemedtobemostefficient.Thismaybethecaseiftwoprocessor-intensivetasks(A&B)haveaffinitytooneprocessorwhileanotherprocessorliesunused.ManyalgorithmswouldshifttaskBtothesecondprocessorinordertomaximizeprocessoruse.TaskBwouldthenacquireaffinitywiththesecondprocessorwhiletaskAwouldcontinuetohaveaffinitywiththeoriginalprocessor.Processoraffinitycaneffectivelyreducecacheproblemsbutitdoesnotcurbthepersistentload-balancingproblem.[1]Further,processoraffinitybecomesmorecomplicatedinsystemswithnon-uniformarchitectures.Forexample,asystemwithtwodual-corehyper-threadedCPUspresentsachallengetoaschedulingalgorithm.ThereiscompleteaffinitybetweentwovirtualCPUsimplementedonthesamecoreviahyper-threading,partialaffinitybetweentwocoresonthesamephysicalchip(asthecoressharesome,butnotall,cache),andnoaffinitybetweenseparatephysicalchips.Asotherresourcesarealsoshared,processoraffinityalonecannotbeusedasthebasisforCPUdispatching.Ifaprocesshasrecentlyrunononevirtualhyper-threadedCPUinagivencore,andthatvirtualCPUiscurrentlybusybutitspartnerisnot,cacheaffinitywouldsuggestthattheprocessshouldbedispatchedtotheidlepartner.However,thetwovirtualCPUscompeteforessentiallyallcomputing,cache,andmemoryresources.ItwouldtypicallybemoreefficientinthiscasetodispatchtheprocesstoadifferentcoreorCPUifoneisavailable.Thiswouldlikelyincurapenaltywhenprocessrepopulatesthecache,butoverallperformancewouldlikelybehigherastheprocesswouldnothavetocompeteforresourceswithintheCPU.OnLinuxtheCPUaffinityofaprocessmightbealteredwiththetaskset(1)program.[2]OnSGIsystems,dplacebindsaprocesstoasetofCPUs.[3]NetBSD5.0,FreeBSD7.2andlaterversionscanusepthread_setaffinity_npandpthread_getaffinity_np.[4]InNetBSD,thepsrsetutility[5]tosetathread'saffinitytoacertainCPUset.InFreeBSD,cpuset[6]utilityisusedtocreateCPUsetsandtoassignprocessestothesesets.OnWindowsNT,threadandprocessCPUaffinitiescanbesetseparatelybyusingSetThreadAffinityMask[7]andSetProcessAffinityMask[8]APIcallsorviatheTaskManagerinterface(forprocessaffinityonly).MacOSXexposesanaffinityAPIthatprovideshintstothekernelhowtoschedulethreadsaccordingtoaffinitysets.2、sched_setaffinity,sched_getaffinity-setandgetaprocess'sCPUaffinitymaskSYNOPSIS#include<sched.h>intsched_setaffinity(pid_tpid,unsignedintlen,unsignedlong*mask);intsched_getaffinity(pid_tpid,unsignedintlen,unsignedlong*mask);DESCRIPTIONsched_setaffinitysetstheCPUaffinitymaskoftheprocessdenotedbypid.Ifpidiszero,thenthecurrentprocessisused.Theaffinitymaskisrepresentedbythebitmaskstoredinmask.Theleastsignificantbitcorrespondstothefirstlogicalprocessornumberonthesystem,whilethemostsignificantbitcorrespondstothelastlogicalprocessornumberonthesystem.AsetbitcorrespondstoalegallyschedulableCPUwhileanunsetbitcorrespondstoanillegallyschedulableCPU.Inotherwords,aprocessisboundtoandwillonlyrunonprocessorswhosecorrespondingbitisset.Usually,allbitsinthemaskareset.Theargumentlenisthelengthofthedatapointedtobymask.Normallythisisthesizeofawordonthesystem.ForcompatibilitywithfutureversionsoftheLinuxkernel,sincethissizecanchange,thebitmasksuppliedmustbeatleastaslargeastheaffinitymaskstoredinthekernel.Thefunctionsched_getaffinitywritesintothepointersuppliedbymaskthatissizelentheaffinitymaskofprocesspid.Ifpidiszero,thenthemaskofthecurrentprocessisreturned.RETURNVALUEOnsuccess,sched_setaffinityandsched_getaffinitybothreturn0.Onerror,-1isreturned,anderrnoissetappropriately.ERRORSEFAULTAsuppliedmemoryaddresswasinvalid.ESRCHTheprocesswhoseIDispidcouldnotbefound.EPERMThecallingprocessdoesnothaveappropriateprivileges.Theprocesscallingsched_setaffinityneedsaneffectiveuidequaltotheeuidoruidoftheprocessidentifiedbypid,oritmustpossesstheCAP_SYS_NICEcapability.EINVALTheaffinitybitmaskmaskcontainsnoprocessorsthatarephysicallyonthesystemorthelengthlenissmallerthanthesizeoftheaffinitymaskusedbythekernel.HISTORYTheaffinitysyscallswereintroducedinLinuxkernel2.5.8.Thelibrarycallswereintroducedinglibc2.3,andarestillinglibc2.3.2.Laterglibc2.3.2developmentversionschangedthisinterfacetoonewithoutthelenfield,andstilllaterversionsrevertedagain.Theglibcprototypeisnow/*SettheCPUaffinityforatask*/externintsched_setaffinity(pid_tpid,size_tcpusetsize,constcpu_set_t*cpuset);/*GettheCPUaffinityforatask*/externintsched_getaffinity(pid_tpid,3、TASKSET(1)

size_tcpusetsize,cpu_set_t*cpuset);LinuxUser’sManualTASKSET(1)NAMEtaskset-retrieveorsetaprocesses’sCPUaffinitySYNOPSIStaskset[options][mask|list][pid|command[arg]...]DESCRIPTIONtasksetisusedtosetorretrievetheCPUaffinityofarunningpro-cessgivenitsPIDortolaunchanewCOMMANDwithagivenCPUaffin-ity.CPUaffinityisaschedulerpropertythat"bonds"aprocesstoagivensetofCPUsonthesystem.TheLinuxschedulerwillhonorthegivenCPUaffinityandtheprocesswillnotrunonanyotherCPUs.NotethattheLinuxscheduleralsosupportsnaturalCPUaffinity:theschedulerattemptstokeepprocessesonthesameCPUaslongaspracti-calforperformancereasons.Therefore,forcingaspecificCPUaffin-ityisusefulonlyincertainapplications.TheCPUaffinityisrepresentedasabitmask,withthelowestorderbitcorrespondingtothefirstlogicalCPUandthehighestorderbitcorre-spondingtothelastlogicalCPU.NotallCPUsmayexistonagivensystembutamaskmayspecifymoreCPUsthanarepresent.AretrievedmaskwillreflectonlythebitsthatcorrespondtoCPUsphysicallyonthesystem.Ifaninvalidmaskisgiven(i.e.,onethatcorrespondstonovalidCPUsonthecurrentsystem)anerrorisreturned.Themasksaretypicallygiveninhexadec