Musalole Kusungirako Kukhale Bottleneck Yofunika Kwambiri pa Maphunziro Achitsanzo

Zanenedwa kuti makampani aukadaulo akuthamangira ma GPU kapena akupita kukawapeza.Mu Epulo, wamkulu wa Tesla Elon Musk adagula ma GPU 10,000 ndipo adati kampaniyo ipitiliza kugula ma GPU ambiri kuchokera ku NVIDIA.Kumbali yamabizinesi, ogwira ntchito pa IT akukankhiranso mwamphamvu kuti awonetsetse kuti ma GPU akugwiritsidwa ntchito nthawi zonse kuti apititse patsogolo kubweza ndalama.Komabe, makampani ena atha kupeza kuti ngakhale kuchuluka kwa ma GPU akuchulukirachulukira, kusagwira ntchito kwa GPU kumakhala kovutirapo.

Ngati mbiriyakale yatiphunzitsa chilichonse chokhudza makompyuta apamwamba kwambiri (HPC), ndikuti kusungirako ndi ma intaneti siziyenera kuperekedwa nsembe chifukwa choyang'ana kwambiri pa computing.Ngati kusungirako sikungathe kusamutsa deta kumayunitsi apakompyuta, ngakhale mutakhala ndi ma GPU ambiri padziko lapansi, simungathe kuchita bwino.

Malinga ndi Mike Matchett, wowunikira pa Small World Big Data, mitundu yaying'ono imatha kuchitidwa kukumbukira (RAM), kulola kuyang'ana kwambiri pakuwerengera.Komabe, mitundu yokulirapo ngati ChatGPT yokhala ndi mabiliyoni a ma node sangathe kusungidwa pamtima chifukwa cha kukwera mtengo.

"Simungathe kukwanira mabiliyoni ambiri a kukumbukira, kotero kusungirako kumakhala kofunika kwambiri," akutero Matchett.Tsoka ilo, kusungirako deta nthawi zambiri sikumanyalanyazidwa panthawi yokonzekera.

Mwambiri, mosasamala kanthu za momwe mungagwiritsire ntchito, pali mfundo zinayi zodziwika bwino pakuphunzitsira kwachitsanzo:

1. Maphunziro a Chitsanzo
2. Inference Application
3. Kusungirako Deta
4. Makompyuta Ofulumira

Popanga ndi kutumiza zitsanzo, zofunikira zambiri zimayika patsogolo umboni wamalingaliro ofulumira (POC) kapena malo oyesera kuti ayambitse maphunziro achitsanzo, osafunikira kusungitsa deta.

Komabe, vuto liri chifukwa chakuti maphunziro kapena kupititsa patsogolo kutha kukhala miyezi kapena zaka.Makampani ambiri amakulitsa kukula kwamitundu yawo panthawiyi, ndipo zomangamanga ziyenera kukulirakulira kuti zigwirizane ndi mitundu yomwe ikukula komanso ma dataset.

Kafukufuku wochokera ku Google pa mamiliyoni ambiri a ntchito zophunzitsira za ML akuwonetsa kuti pafupifupi 30% ya nthawi yophunzitsira imagwiritsidwa ntchito popanga payipi ya data.Ngakhale kafukufuku wam'mbuyomu adayang'ana kwambiri kukhathamiritsa ma GPU kuti afulumizitse maphunziro, zovuta zambiri zikadali pakukwaniritsa magawo osiyanasiyana a mapaipi a data.Mukakhala ndi mphamvu zowerengera, vuto lenileni limakhala momwe mungayankhire mwachangu powerengera kuti mupeze zotsatira.

Makamaka, zovuta zosungirako deta ndi kasamalidwe zimafunika kukonzekera kukula kwa deta, kukulolani kuti mutulutse mtengo wa deta nthawi zonse pamene mukupita patsogolo, makamaka pamene mukugwiritsa ntchito njira zapamwamba kwambiri monga kuphunzira mozama ndi neural network, zomwe zimafuna kwambiri. kusungirako malinga ndi mphamvu, ntchito, ndi scalability.

Makamaka:

Scalability
Kuphunzira pamakina kumafuna kugwiritsa ntchito deta yochulukirapo, ndipo kuchuluka kwa deta kumachulukira, kulondola kwamitundu kumakweranso.Izi zikutanthauza kuti mabizinesi amayenera kusonkhanitsa ndikusunga zambiri tsiku lililonse.Pamene kusungirako sikungathe kukulirakulira, kuchulukitsitsa kwa data kumapangitsa kuti pakhale zopinga, kuchepetsa magwiridwe antchito ndikupangitsa kuti pakhale nthawi yodula ya GPU.

Kusinthasintha
Thandizo losinthika la ma protocol angapo (kuphatikizapo NFS, SMB, HTTP, FTP, HDFS, ndi S3) ndizofunikira kukwaniritsa zosowa za machitidwe osiyanasiyana, m'malo mongokhala ndi mtundu umodzi wa chilengedwe.

Kuchedwa
I/O latency ndiyofunikira pakumanga ndi kugwiritsa ntchito zitsanzo popeza deta imawerengedwa ndikuwerengedwanso kangapo.Kuchepetsa I/O latency kungafupikitse nthawi yophunzitsira ya zitsanzo ndi masiku kapena miyezi.Kukula kwachitsanzo kwachangu kumatanthauzira mwachindunji phindu lalikulu labizinesi.

Kupititsa patsogolo
Mayendedwe a kachitidwe kakusungirako ndikofunikira kuti tiphunzitse bwino zachitsanzo.Njira zophunzitsira zimaphatikizapo kuchuluka kwa data, makamaka ma terabytes pa ola limodzi.

Parallel Access
Kuti akwaniritse ntchito zambiri, zitsanzo zophunzitsira zimagawaniza ntchito kukhala ntchito zingapo zofanana.Izi nthawi zambiri zimatanthawuza kuti makina ophunzirira makina amapeza mafayilo omwewo kuchokera kumachitidwe angapo (mwina pa maseva angapo akuthupi) nthawi imodzi.Dongosolo losungirako liyenera kuthana ndi zofuna nthawi imodzi popanda kusokoneza magwiridwe antchito.

Ndi luso lake lapadera mu latency yotsika, kutulutsa kwakukulu, ndi kufanana kwakukulu kwa I/O, Dell PowerScale ndi malo abwino osungiramo makompyuta a GPU-accelerated computing.PowerScale imachepetsa bwino nthawi yofunikira pakuwunikira zitsanzo zomwe zimaphunzitsa ndi kuyesa ma dataset a multi-terabyte.Mu PowerScale zosungira zonse, bandwidth imawonjezeka ndi nthawi za 18, kuchotsa mabotolo a I / O, ndipo akhoza kuwonjezeredwa kumagulu omwe alipo a Isilon kuti afulumizitse ndi kutsegula mtengo wa deta yochuluka yosakonzedwa.

Kuphatikiza apo, kuthekera kofikira kwa PowerScale kwamitundu yambiri kumapereka kusinthasintha kopanda malire pakuyendetsa ntchito, kulola kuti deta isungidwe pogwiritsa ntchito protocol imodzi ndikufikiridwa ndi ina.Makamaka, mawonekedwe amphamvu, kusinthasintha, kuchulukira, komanso magwiridwe antchito amagawo a PowerScale amathandizira kuthana ndi zovuta zotsatirazi:

- Limbikitsani luso mpaka nthawi 2.7, kuchepetsa kuzungulira kwa maphunziro.

- Chotsani zovuta za I/O ndikupereka maphunziro ofulumira achitsanzo ndi kutsimikizira, kulondola kwachitsanzo, kupititsa patsogolo zokolola za sayansi ya data, komanso kubweza ndalama zambiri pamakompyuta pogwiritsa ntchito zida zamabizinesi, magwiridwe antchito apamwamba, kugwirizana, komanso kusakhazikika.Limbikitsani kulondola kwachitsanzo ndi ma dataseti akuya, otsimikiza kwambiri mwa kugwiritsa ntchito mpaka 119 PB yosungirako bwino mugulu limodzi.

- Fikirani kutumizidwa pamlingo poyambira pang'onopang'ono komanso modziyimira pawokha makulitsidwe ndi kusunga, kupereka chitetezo champhamvu cha data ndi njira zotetezedwa.

- Limbikitsani zokolola za sayansi ya data ndi ma analytics am'malo ndi mayankho omwe adatsimikizidwa kale kuti atumizidwe mwachangu, opanda chiopsezo chochepa.

- Kugwiritsa ntchito mapangidwe otsimikiziridwa kutengera matekinoloje apamwamba kwambiri, kuphatikiza mathamangitsidwe a NVIDIA GPU ndi kamangidwe kake ndi machitidwe a NVIDIA DGX.Kuchita kwapamwamba kwa PowerScale ndi kugwirizanitsa kumakwaniritsa zofunikira pakusungirako pagawo lililonse la kuphunzira kwamakina, kuyambira pakupeza deta ndi kukonzekera mpaka kuphunzitsira kwachitsanzo ndi kulingalira.Pamodzi ndi makina ogwiritsira ntchito OneFS, ma node onse amatha kugwira ntchito mosasunthika mkati mwa gulu lomwelo loyendetsedwa ndi OneFS, lomwe lili ndi magawo abizinesi monga kasamalidwe ka magwiridwe antchito, kasamalidwe ka data, chitetezo, ndi chitetezo cha data, zomwe zimathandiza kumaliza mwachangu maphunziro achitsanzo ndi kutsimikizira kwa mabizinesi.


Nthawi yotumiza: Jul-03-2023