A watan Fabrairu, Stanford ya shirya wani taro a kan manyan ayyuka (HPC). Wakilan VMware sun ce lokacin aiki tare da GPU, tsarin da ya dogara da ingantaccen hypervisor na ESXi bai yi ƙasa da sauri zuwa mafita na ƙarfe ba.
Muna magana game da fasahohin da suka ba da damar cimma wannan.
/ hoto
Batun aiki
A cewar manazarta, kusan kashi 70% na yawan aiki a cibiyoyin bayanai . Koyaya, sauran kashi 30% har yanzu suna gudana akan ƙaramin ƙarfe ba tare da hawan jini ba. Wannan 30% galibi ya ƙunshi aikace-aikace masu ɗaukar nauyi, kamar waɗanda ke da alaƙa da horar da hanyoyin sadarwa na jijiyoyi, da amfani da GPUs.
Masana sun bayyana wannan yanayin ta hanyar gaskiyar cewa hypervisor, a matsayin tsaka-tsakin tsaka-tsakin tsaka-tsakin, zai iya rinjayar aikin dukan tsarin. A cikin karatu shekaru biyar da suka wuce game da rage saurin aiki da kashi 10%. Don haka, kamfanoni da ma'aikatan cibiyar bayanai ba sa gaggawar canja wurin aikin HPC zuwa yanayin kama-da-wane.
Amma fasahohin haɓakawa suna haɓaka kuma suna haɓaka. A wani taro wata daya da suka gabata, VMware ya ce ESXi hypervisor ba shi da wani mummunan tasiri akan aikin GPU. Ana iya rage saurin kwamfuta da kashi uku cikin ɗari, wanda ya yi daidai da ƙarfe maras tushe.
Ta yaya wannan aikin
Don inganta aikin tsarin HPC tare da GPUs, VMware ya yi sauye-sauye da dama ga hypervisor. Musamman, an kawar da aikin vMotion. Ana buƙatar don daidaita nauyi kuma yawanci yana canja wurin injunan kama-da-wane (VMs) tsakanin sabar ko GPUs. Kashe vMotion ya haifar da kowane VM yanzu ana sanya masa takamaiman GPU. Wannan ya taimaka rage farashi lokacin musayar bayanai.
Wani maɓalli mai mahimmanci na tsarin Hanyar kai tsaye I/O. Yana ba da damar CUDA daidaitaccen direban kwamfuta don yin hulɗa tare da injina kai tsaye, ketare hypervisor. Lokacin da kuke buƙatar gudanar da VM da yawa akan GPU ɗaya lokaci ɗaya, ana amfani da maganin GRID vGPU. Yana raba žwažwalwar ajiyar katin zuwa sassa da yawa (amma ba a raba zagayowar lissafin ba).
Jadawalin aiki na inji guda biyu a cikin wannan yanayin zai yi kama da haka:
Sakamako da hasashen
M hypervisor ta horar da samfurin harshe bisa . Ayyukan "lalacewar" shine kawai 3-4% idan aka kwatanta da karfe maras tushe. A sakamakon haka, tsarin ya iya rarraba albarkatu akan buƙata dangane da nauyin da ake ciki yanzu.
Giant din IT kuma tare da kwantena. Injiniyoyin kamfanin sun horar da hanyoyin sadarwa na jijiyoyi don gane hotuna. A lokaci guda, an rarraba albarkatun GPU ɗaya tsakanin VMs kwantena guda huɗu. Sakamakon haka, aikin injuna ɗaya ya ragu da 17% (idan aka kwatanta da VM guda ɗaya tare da cikakken damar samun albarkatun GPU). Koyaya, adadin hotunan da aka sarrafa kowace daƙiƙa sau uku. Ana sa ran cewa irin wannan tsarin aikace-aikace a cikin nazarin bayanai da ƙirar kwamfuta.
Daga cikin yuwuwar matsalolin da VMware ka iya fuskanta, masana maimakon kunkuntar manufa masu sauraro. Ƙananan kamfanoni har yanzu suna aiki tare da tsarin aiki mai girma. Ko da yake a Statista cewa nan da shekarar 2021, kashi 94% na yawan ayyukan cibiyar bayanai na duniya za su kasance da inganci. By Manazarta, darajar kasuwar HPC za ta karu daga dala biliyan 32 zuwa dala biliyan 45 a tsakanin shekarar 2017 zuwa 2022.
/ hoto PD
Makamantan mafita
Akwai analogues da yawa a kasuwa waɗanda manyan kamfanonin IT suka haɓaka: AMD da Intel.
Kamfani na farko don haɓakawa na GPU tsarin da ya danganci SR-IOV (shigarwar tushen-guda ɗaya/fitarwa na zahiri). Wannan fasaha tana ba wa VM damar zuwa ɓangaren ƙarfin kayan aikin tsarin. Maganin yana ba ku damar raba GPU tsakanin masu amfani da 16 tare da daidaitaccen aiki na tsarin ƙima.
Amma ga na biyu IT giant, su a kan Citrix XenServer hypervisor 7. Yana haɗuwa da aikin daidaitaccen direba na GPU da na'ura mai mahimmanci, wanda ya ba da damar na ƙarshe don nuna aikace-aikacen 3D da tebur akan na'urorin daruruwan masu amfani.
Makomar fasaha
Virtual GPU Developers akan aiwatar da tsarin AI da kuma karuwar shaharar hanyoyin samar da ayyukan yi a kasuwar fasahar kasuwanci. Suna fatan cewa buƙatar aiwatar da adadi mai yawa na bayanai zai ƙara buƙatar vGPUs.
Yanzu masana'antun hada ayyuka na CPU da GPU a cikin cibiya ɗaya don hanzarta magance matsalolin da suka shafi zane-zane, yin lissafin lissafi, ayyukan ma'ana, da sarrafa bayanai. Bayyanar irin waɗannan nau'ikan a kasuwa a nan gaba zai canza hanyar da za a iya amfani da kayan aiki na kayan aiki da kuma rarraba su tsakanin ayyukan aiki a cikin yanayin kama-da-wane da girgije.
Abin da za mu karanta a kan batun a cikin rukunin yanar gizon mu:
Ga wasu abubuwa guda biyu daga tasharmu ta Telegram:
source: www.habr.com