Abubuwan da ke ciki
boye
Fara da Intel Trace Analyzer da Mai tarawa
Fara da Intel® Trace Analyzer da Mai tarawa
Yi amfani da wannan takaddar Farawa da alamar da aka riga aka tattara file don yin tafiya ta hanyar bincike na aikin MPI na asali tare da Intel® Trace Analyzer da Mai tarawa.
Intel Trace Analyzer da Mai tarawa yana taimakawa bincika ingantaccen amfani da hanyar wucewa ta saƙo (MPI) da gano wuraren sadarwa, ƙullawar aiki tare, da daidaita kaya. Don ƙarin bayani game da samfurin, duba Intel Trace Analyzer and Collector shafin samfurin.
Zazzage Intel Trace Analyzer da Mai Tara
- a matsayin wani ɓangare na Intel® oneAPI HPC Toolkit
- azaman kayan aiki na tsaye
Abubuwan da ake bukata
- Kafin gudanar da Intel Trace Analyzer da Collector, tabbatar cewa kun shigar da sabuwar Intel® MPI Library da Intel® oneAPI DPC++/C++ Compiler ko Intel® Fortran Compiler.
- Wannan yana saita sauye-sauyen yanayi da ake buƙata don masu tarawa, Laburaren Intel MPI, da Intel Trace Analyzer da Collector, kuma kuna shirye don gano aikace-aikacenku.
- Don ƙarin bayani, duba: Intel® oneAPI HPC Toolkit System Bukatun.
Fahimtar Tsarin Aiki
- Nemo Aikace-aikacenku
- Yi nazarin ayyukan MPI mafi aiki
- Gano ma'amala mai matsala
- Inganta aikin aikace-aikacen ku ta maye gurbin aikin da ke haifar da matsala
Nemo Aikace-aikacen MPI naku
Ƙirƙirar alama file don tattara rajistan ayyukan taron don nazarin halayen aikace-aikacen mai zuwa.
- Ƙirƙiri yanayi don ƙaddamar da Intel® Trace Analyzer da Mai tarawa ta hanyar gudanar da rubutun saiti daga daraktan shigarwa na oneAPI.
NOTE
Ta hanyar tsoho, Intel Trace Analyzer da Mai tarawa an shigar dashi zuwa / opt/intel/oneapi/itac don Linux* OS da zuwa Shirye-shirye. Files (x86)\Intel\oneAPI\itac\sabon don Windows* OS.
A Linux:
tushen $ /opt/intel/oneapi/setvars.sh
Na Windows:
"C:\Program Files (x86)\Intel oneAPIsetvars.bat" - Gudanar da aikace-aikacen MPI ɗin ku kuma samar da alama tare da zaɓin -trace.
A Linux:
$ mpirun -trace -n 4 ./poisson_sendecv.single
Na Windows:
Haɗa ƙa'idar kuma tattara alamar.
Don Intel oneAPI DPC++/C++ Compiler, gudu:
> mpicc -trace poisson_sendrecv.single.c
Don Intel Fortran Compiler, gudanar:
> mpiifort -trace poisson_sendrecv.single.f
Wannan example yana haifar da alama (stf*) don asample poisson_sendrcv. aikace-aikacen MPI guda ɗaya - Buɗe .stf file tare da Intel Trace Analyzer tare da Intel Trace Analyzer da Mai tarawa.
A Linux:
$ traceanalyzer ./ poisson_sendecv.single.stf
Na Windows:
traceanalyzer poisson_sendecv.single.stf
NOTE
Don dalilai na gwaji, zaku iya zazzage alamar da aka riga aka tattara file poisson_sendecv.single.stf don poisson da aka yi amfani da shi a cikin wannan takarda kuma buɗe shi tare da Intel Trace Analyzer da Mai tarawa.
Da .stf file yana buɗewa a cikin Shafin Takaitawa view, wanda ke wakiltar cikakken bayani game da aikin aikace-aikacenku:
NOTE Don ƙarin bayani game da Intel Trace Analyzer da ayyuka masu tarawa, duba Ƙara koyo.
Yi nazarin Ayyukan MPI Mafi Aiki
Yi nazarin halayen aikace-aikacen MPI, nemo ƙullun da gano serialization don nemo hanyoyin inganta aikin aikace-aikacen.
- Daga Shafi na Takaitawa bude Timeline view ta danna Ci gaba > Charts > Layin Lokaci don zurfafa nazarin manyan ayyukan MPI.
Taswirar tana nuna ayyukan aiwatar da mutum ɗaya akan lokaci.
Ayyukan aikace-aikacen na maimaitawa ne, inda kowane juzu'i ya ƙunshi ɓangaren lissafi da sadarwar MPI. - Gano juzu'i guda ɗaya don mayar da hankali a kai da zuƙowa cikinsa ta hanyar jan linzamin kwamfuta akan tazarar lokacin da ake buƙata:
Alamar view yana nuna sashe a cikin alamar da kuka zaɓa. Jadawalin Jadawalin Halittu yana nuna abubuwan da suka faru lokacin da aka zaɓa.
- Sandunan kwance suna wakiltar matakai tare da ayyukan da ake kira a cikin waɗannan matakan.
- Baƙar fata suna nuna saƙonnin da aka aika tsakanin matakai. Waɗannan layukan suna haɗa hanyoyin aikawa da karɓa.
- Layukan shuɗi suna wakiltar ayyukan gama kai, kamar watsa shirye-shirye ko rage ayyuka.
- Canja zuwa Flat Profile shafin (A) don samun kusancin duba ayyukan da ke aiwatarwa a cikin lokacin da kuka zaɓa (wanda aka zaɓa a cikin Tsarin Lokaci.
- Cire ayyukan MPI don tantance ayyukan tsarin MPI a cikin aikace-aikacen ku.
Don yin wannan, danna-dama Duk Tsari> Rukunin MPI (B) a cikin Flat Profile kuma zaɓi UngroupMPI. Wannan aikin yana fallasa kiran MPI guda ɗaya. - Yi nazarin hanyoyin sadarwa tare da maƙwabtansu kai tsaye ta amfani da MPI_Sendrecv a farkon ƙaddamarwa. Don misaliampda:
- a. A cikin sampHar ila yau, musayar bayanan MPI_Sendrecv yana da matsala: tsarin ba ya musayar bayanai tare da maƙwabcinsa na gaba har sai an kammala musayar tare da na baya. Jadawalin Lamarin view yana nuna wannan kuncin a matsayin matakala.
- b. MPI_Allreduce a ƙarshen ƙaddamarwa yana sake daidaita duk matakai; shi ya sa wannan shingen yana da kamannin bene na baya.
- Gano serialization, ta amfani da Aiki Profile da Message Profile views.
- a. Bude jadawalin a lokaci guda:
A cikin Aiki Profile ginshiƙi, buɗe Load Balancetab. - Jeka menu na Charts don buɗe Saƙon Profile.
- b. A cikin Load Balance tab, fadada MPI_Sendrecv da MPI_Allreduce. Daidaiton Load yana nuna cewa lokacin da aka kashe a MPI_Sendrecv yana ƙaruwa tare da lambar tsari, yayin da lokacin MPI_Allreduce ya ragu.
- c. Yi nazarin Saƙon Profile Jadawalin ƙasa zuwa kusurwar dama ta ƙasa.
Ƙididdiga launi na tubalan yana nuna cewa saƙonnin da ke tafiya daga matsayi mafi girma zuwa ƙananan matsayi suna buƙatar lokaci mai yawa yayin da saƙonnin da ke tafiya daga ƙananan matsayi zuwa matsayi mafi girma suna bayyana wani nau'i mai rauni ko da ban mamaki:
- a. Bude jadawalin a lokaci guda:
Sakamakon binciken kwatankwacin ya nuna cewa babu wani hadadden tsarin musayar a cikin aikace-aikacen, ana yin musayar ne kawai tare da matakai na makwabta. Bayanin zai zama mahimmanci don Inganta Ayyukan Aikace-aikacenku ta Canza matakin Sadarwa don haɓaka ƙirar sadarwar aikace-aikacen.
Gano Sadarwar Sadarwar da ba ta da Ma'ana
Kalli aikace-aikacen ku a ƙarƙashin ingantattun yanayi kuma kwatanta asalin asalin file tare da wanda aka keɓe don keɓe ma'amala mai matsala.
- Ƙirƙiri ingantaccen tsari file:
- a. Zaɓi Babba > Haɓakawa ko danna maɓallin
(Idealization) Toolbar button. - b. Duba ma'auni na manufa a cikin akwatin maganganu na Idealization (madaidaicin alama file suna da kewayon lokaci don canzawa).
- c. Danna Fara don daidaita alamar ku.
- a. Zaɓi Babba > Haɓakawa ko danna maɓallin
- Kwatanta alamar asali tare da ingantacciyar alama:
- a. Zaɓi Babba > Zane na rashin daidaituwa ko danna maɓallin
Maɓallin kayan aiki (Rashin daidaitawa). - b. A cikin akwatin maganganu na rashin daidaituwa, danna Buɗe Wani File maɓalli, kewaya zuwa ga alama mai kyau, kuma zaɓi shi.
- c. A cikin Tagar Rashin daidaituwa, danna maɓallin Yanayin Jumla kuma zaɓi Yanayin Breakdown.
- a. Zaɓi Babba > Zane na rashin daidaituwa ko danna maɓallin
Kuna iya ganin cewa MPI_Sendrecv shine aikin da ya fi cin lokaci. Ana nuna nauyin rashin daidaituwa a ciki
launin haske kuma ya ƙunshi kusan 10% don aikin MPI_Sendrecv. Wannan shine lokacin da hanyoyin ke ciyarwa suna jiran juna.
Inganta Ayyukan Aikace-aikacenku ta Canza Sadarwa
- Inganta aikin aikace-aikacen MPI ta hanyar canza toshewa zuwa hanyoyin sadarwa marasa toshewa.
A cikin lambar ku maye gurbin MPI_Sendrcv serial tare da sadarwar mara toshewa: MPI_Isend da MPI_Irecv. Don misaliample: Asalin code snippet:
// musanya iyaka
musayar mara amfani (para * p, grid * gr){
zan i,j;
MPI_Matsayin_100, status_200, status_300, status_400;
// saukar da layin farko
MPI_Send(gr->x_new[1], gr->col+2, MPI_DOUBLE, gr-> kasa, 100, MPI_COMM_WORLD); MPI_Recv(gr->x_new[gr->low+1], gr->col+2, MPI_DOUBLE, gr-> sama, 100, MPI_COMM_WORLD,
& matsayi_100);
// aika layi na ƙarshe
MPI_Send(gr->x_new[gr->low], gr->col+2, MPI_DOUBLE, gr-> sama, 200, MPI_COMM_WORLD);
MPI_Recv(gr->x_new[0], gr->col+2, MPI_DOUBLE, gr-> kasa, 200, MPI_COMM_WORLD, & status_200);
Yi amfani da Kwatancen Binciken Binciken Intel Trace view don kwatanta aikace-aikacen da aka jera tare da wanda aka bita
// kwafi shafi na hagu zuwa tmp arrays
idan (gr-> hagu! = MPI_PROC_NULL){
gr->x_new[i][gr->col+1] = dama_col[i]; right_col[i] = gr->x_new[i][gr->lcol];
// aika dama
MPI_Send (dama_col, gr->low+2, MPI_DOUBLE, gr-> dama, 400, MPI_COMM_WORLD); }
idan (gr-> hagu! = MPI_PROC_NULL)
{
MPI_Recv(left_col, gr->low+2, MPI_DOUBLE, gr->hagu, 400, MPI_COMM_WORLD,&status_400); don (i=0; i< gr->low+2; i++
{
gr->x_new[i][0] = hagu_col[i];
}
}
Snippet lambar da aka sabunta
MPI_Neman buqatar[7];
// saukar da layin farko
MPI_Isend (gr-> x_new [1], gr-> lcol +2, MPI_DOUBLE, gr-> ƙasa, 100, MPI_COMM_WORLD, & req[0]);
MPI_Irecv (gr-> x_new[gr-> lrow+1], gr->col+2, MPI_DOUBLE, gr-> sama, 100, MPI_COMM_WORLD, & req[1]);
…..
MPI_Waitall(7, req, MPI_STATUSES_IGNORE);
Da zarar an gyara, juzu'i guda na aikace-aikacen da aka bita zai yi kama da mai zuwaampda:
- Yi amfani da Kwatancen Binciken Binciken Intel Trace view don kwatanta aikace-aikacen da aka jera tare da wanda aka bita. Kwatanta alamu biyu tare da taimakon Kwatanta View, zuwa View > Kwatanta. Kwatanta View yayi kama da:
A cikin Kwatanta View, za ka iya ganin cewa yin amfani da ba tarewa sadarwa taimaka wajen cire serialization da kuma rage lokacin sadarwa na matakai.
NOTE Don ƙarin bayani game da aikin matakin kumburi na aikace-aikacenku, duba takaddun kayan aikin daban-daban: Intel® VTune™ Profiler MPI Code Analysis da Binciken aikace-aikacen Intel® MPI ta amfani da Intel® Advisor.
Ƙara Koyi
Bincika albarkatun masu zuwa don ƙarin bayani game da Intel Trace Analyzer da Mai tarawa.
Sanarwa da Rarrabawa
- Fasahar Intel na iya buƙatar kayan aikin da aka kunna, software ko kunnawa sabis.
- Babu samfur ko sashi wanda zai iya zama cikakkiyar amintacce.
- Kudin ku da sakamakon ku na iya bambanta.
- © Kamfanin Intel. Intel, tambarin Intel, da sauran alamun Intel alamun kasuwanci ne na Kamfanin Intel Corporation ko rassan sa. Ana iya da'awar wasu sunaye da alamun a matsayin mallakin wasu.
- Babu lasisi (bayyana ko fayyace, ta estoppel ko akasin haka) ga kowane haƙƙin mallakar fasaha da aka bayar ta wannan takaddar.
- Samfuran da aka siffanta na iya ƙunsar lahani na ƙira ko kurakurai da aka sani da errata wanda zai iya sa samfurin ya saba da ƙayyadaddun bayanai da aka buga. Ana samun siffa ta halin yanzu akan buƙata.
- Intel yana ƙin duk cikakkun bayanai da garanti mai ma'ana, gami da ba tare da iyakancewa ba, garantin ciniki, dacewa don wata manufa, da rashin cin zarafi, da kowane garanti da ya taso daga hanyar aiki, hanyar mu'amala, ko amfani a kasuwanci.
Takardu / Albarkatu
 |
intel Fara da Intel Trace Analyzer da Mai Tara [pdf] Jagorar mai amfani Fara da Intel Trace Analyzer da Collector, Fara da Intel, Trace Analyzer da Mai tarawa, Mai tarawa. |
