NgoFebhuwari, uStanford ubambe ingqungquthela ye-high-performance computing (HPC). Abamele i-VMware bathi lapho besebenza ne-GPU, uhlelo olusekelwe ku-ESXi hypervisor eguquliwe alukho ngaphansi ngesivinini kunezixazululo zensimbi ezingenalutho.
Sikhuluma ngobuchwepheshe obenze kwaba nokwenzeka ukufeza lokhu.
/isithombe
Inkinga yokusebenza
Ngokusho kwabahlaziyi, cishe u-70% wemithwalo yemisebenzi ezikhungweni zedatha . Kodwa-ke, ama-30% asele asasebenza ngensimbi engenalutho ngaphandle kwama-hypervisors. Lokhu 30% ikakhulukazi kuqukethe izinhlelo zokusebenza ezinomthwalo omkhulu, njengalezo ezihlobene nokuqeqeshwa kwamanethiwekhi e-neural, nokusebenzisa ama-GPU.
Ochwepheshe bachaza lo mkhuba ngokuthi i-hypervisor, njenge-intermediate abstraction layer, ingathinta ukusebenza kwalo lonke uhlelo. Ezifundweni eminyakeni emihlanu edlule mayelana nokunciphisa ijubane lomsebenzi ngo-10%. Ngakho-ke, izinkampani nabaqhubi besikhungo sedatha abajahi ukudlulisa imithwalo yemisebenzi ye-HPC endaweni ebonakalayo.
Kepha ubuchwepheshe be-virtualization buyathuthuka futhi buyathuthuka. Engqungqutheleni enyangeni edlule, i-VMware yathi i-ESXi hypervisor ayinawo umthelela omubi ekusebenzeni kwe-GPU. Ijubane lekhompyutha lingancishiswa ngamaphesenti amathathu, eliqhathaniseka nensimbi engenalutho.
Kanjani lo msebenzi
Ukuthuthukisa ukusebenza kwezinhlelo ze-HPC ngama-GPU, i-VMware yenze izinguquko eziningi ku-hypervisor. Ikakhulukazi, kukhishwe umsebenzi we-vMotion. Iyadingeka ekulinganiseni umthwalo futhi ngokuvamile idlulisela imishini ebonakalayo (ama-VM) phakathi kwamaseva noma ama-GPU. Ukukhubaza i-vMotion kuholele ekutheni i-VM ngayinye manje inikezwe i-GPU ethile. Lokhu kusize ukunciphisa izindleko uma ushintshanisa idatha.
Enye ingxenye eyinhloko yesistimu I-DirectPath I/O. Ivumela umshayeli we-CUDA parallel computing ukuthi axhumane nemishini ebonakalayo ngokuqondile, edlula i-hypervisor. Uma udinga ukusebenzisa ama-VM ambalwa ku-GPU eyodwa ngesikhathi esisodwa, kusetshenziswa isisombululo se-GRID vGPU. Ihlukanisa inkumbulo yekhadi ibe izingxenye eziningana (kodwa imijikelezo yokubala ayihlukaniswanga).
Umdwebo wokusebenza wemishini emibili ebonakalayo kuleli cala uzobukeka kanje:
Imiphumela kanye nezibikezelo
Firm hypervisor ngokuqeqesha imodeli yolimi esekelwe . Ukusebenza "ukulimala" bekungama-3-4% kuphela uma kuqhathaniswa nensimbi engenalutho. Ngokubuyisela, uhlelo lukwazile ukusabalalisa izinsiza ngokufunwa kuye ngomthwalo wamanje.
Umdondoshiya we-IT futhi ngezitsha. Onjiniyela benkampani baqeqeshe amanethiwekhi e-neural ukuze abone izithombe. Ngaso leso sikhathi, izinsiza ze-GPU eyodwa zasatshalaliswa phakathi kwama-VM ama-container amane. Ngenxa yalokho, ukusebenza kwemishini ngayinye kwehle ngo-17% (uma kuqhathaniswa ne-VM eyodwa enokufinyelela okugcwele kuzinsiza ze-GPU). Nokho, inani lezithombe ezicutshungulwayo ngomzuzwana kathathu. Kulindeleke ukuthi izinhlelo ezinjalo izinhlelo zokusebenza ekuhlaziyweni kwedatha nasekufaniseni ikhompyutha.
Phakathi kwezinkinga ezingase zibhekane ne-VMware, ochwepheshe izethameli eziqondiwe eziwumngcingo. Izinkampani ezimbalwa zisasebenza ngezinhlelo ezisebenza kahle kakhulu. Nakuba ku-Statista ukuthi ngo-2021, u-94% wemisebenzi yesikhungo sedatha yomhlaba izobe seyenziwe nge-virtual. Ngu Abahlaziyi, inani lemakethe ye-HPC lizokhula lisuka ku-32 liye kuma-dollar ayizigidi eziyizinkulungwane ezingama-45 esikhathini esisuka ku-2017 kuya ku-2022.
/isithombe PD
Izixazululo ezifanayo
Kukhona ama-analogue amaningana emakethe athuthukiswa yizinkampani ezinkulu ze-IT: i-AMD ne-Intel.
Inkampani yokuqala ye-GPU virtualization indlela esekelwe ku-SR-IOV (i-single-root input/outputization). Lobu buchwepheshe bunikeza i-VM ukufinyelela engxenyeni yamakhono wehadiwe yesistimu. Isixazululo sikuvumela ukuthi wabelane nge-GPU phakathi kwabasebenzisi abayi-16 ngokusebenza okulinganayo kwezinhlelo ezibonwayo.
Ngokuqondene ne-IT yesibili giant, bona ku-hypervisor ye-Citrix XenServer 7. Ihlanganisa umsebenzi womshayeli we-GPU ojwayelekile kanye nomshini obonakalayo, ovumela lona wakamuva ukuthi ubonise izinhlelo zokusebenza ze-3D namadeskithophu kumadivayisi angamakhulu abasebenzisi.
Ikusasa lobuchwepheshe
Virtual GPU Developers ekusetshenzisweni kwezinhlelo ze-AI kanye nokuthandwa okukhulayo kwezixazululo ezisebenza kahle emakethe yobuchwepheshe bebhizinisi. Bathemba ukuthi isidingo sokucubungula amanani amakhulu wedatha sizokwandisa isidingo sama-vGPU.
Manje abakhiqizi hlanganisa ukusebenza kwe-CPU ne-GPU kungqikithi eyodwa ukusheshisa ukuxazulula izinkinga ezihlobene nemifanekiso, ukwenza izibalo zezibalo, ukusebenza okunengqondo, nokucubungula idatha. Ukubonakala kwama-cores anjalo emakethe esikhathini esizayo kuzoshintsha indlela yokwenza izinto ezibonakalayo kanye nokusatshalaliswa kwazo phakathi kwemithwalo yemisebenzi ezindaweni ezibonakalayo nezamafu.
Ongakufunda ngesihloko kubhulogi yethu yebhizinisi:
Okuthunyelwe okumbalwa okuvela esiteshini sethu seTelegram:
Source: www.habr.com