NgoFebruwari, iStanford yabamba inkomfa kwi-high-performance computing (HPC). VMware abameli bathi xa usebenza GPU, inkqubo esekelwe ukuguqulwa ESXi hypervisor ayikho ngaphantsi ngesantya izisombululo zentsimbi.
Sithetha ngobuchwephesha obenze ukuba kwenzeke oku.
/ ifoto
Umba wokusebenza
Ngokutsho kwabahlalutyi, malunga ne-70% yomthwalo wemisebenzi kumaziko edatha . Nangona kunjalo, i-30% eseleyo isaqhuba kwintsimbi engenanto ngaphandle kwe-hypervisors. Le 30% ubukhulu becala iqulathe izicelo ezinomthwalo ophezulu, ezifana nezo zinxulumene noqeqesho lothungelwano lwe-neural, kunye nokusetyenziswa kweGPUs.
Iingcali zichaza lo mkhwa ngento yokuba i-hypervisor, njenge-intermediate abstraction layer, inokuchaphazela ukusebenza kwenkqubo yonke. Kwizifundo kwiminyaka emihlanu eyadlulayo malunga nokunciphisa isantya somsebenzi nge-10%. Ke ngoko, iinkampani kunye nabaqhubi beziko ledatha abangxamanga ukuhambisa imithwalo yemisebenzi ye-HPC kwindawo ebonakalayo.
Kodwa itekhnoloji ye-virtualization iyakhula kwaye iyaphucuka. Kwinkomfa kwinyanga edlulileyo, i-VMware yathi i-ESXi hypervisor ayinayo impembelelo embi ekusebenzeni kwe-GPU. Isantya sekhompyutha sinokuncitshiswa ngeepesenti ezintathu, esifaniswa nentsimbi engenanto.
ntoni lo msebenzi
Ukuphucula ukusebenza kweenkqubo ze-HPC kunye ne-GPUs, i-VMware yenze utshintsho oluninzi kwi-hypervisor. Ngokukodwa, yakhutshwa umsebenzi we-vMotion. Iyafuneka ekulinganiseni umthwalo kwaye idla ngokudlulisela oomatshini bokwenene (VMs) phakathi kweeseva okanye ii-GPU. Ukukhubaza i-vMotion kubangele ukuba i-VM nganye inikwe i-GPU ethile. Oku kuncede ukunciphisa iindleko xa utshintshisa idatha.
Elinye icandelo eliphambili lenkqubo I-DirectPath I/O. Ivumela umqhubi we-CUDA we-computing parallel ukuba adibane noomatshini benyani ngokuthe ngqo, ngokudlula i-hypervisor. Xa ufuna ukuqhuba ii-VM ezininzi kwi-GPU enye ngaxeshanye, kusetyenziswa isisombululo se-GRID vGPU. Yahlula imemori yekhadi ibe ngamacandelo amaninzi (kodwa imijikelo yokubala ayohlulwanga).
Umzobo wokusebenza koomatshini ababini abanenyani kule meko iya kujongeka ngolu hlobo:
Iziphumo kunye noqikelelo
Inkampani hypervisor ngokuqeqesha imodeli yolwimi esekelwe . Ukusebenza "umonakalo" kwakuyi-3-4% kuphela xa kuthelekiswa nentsimbi engenanto. Ngembuyekezo, inkqubo ikwazile ukusasaza izibonelelo kwimfuno ngokuxhomekeke kumthwalo wangoku.
Isigebenga se-IT kwakhona ngezikhongozeli. Iinjineli zenkampani ziqeqeshe iinethiwekhi ze-neural ukubona imifanekiso. Kwangelo xesha, izixhobo ze-GPU enye zasasazwa phakathi kwee-VM ezine zekhonteyina. Ngenxa yoko, ukusebenza koomatshini ngabanye kwehle nge-17% (xa kuthelekiswa ne-VM enye kunye nokufikelela ngokupheleleyo kwizibonelelo ze-GPU). Nangona kunjalo, inani lemifanekiso eqhutywe ngesekhondi amaxa amathathu. Kulindeleke ukuba iinkqubo ezinjalo izicelo kuhlalutyo lwedatha kunye nemodeli yekhompyuter.
Phakathi kweengxaki ezinokujamelana nazo i-VMware, iingcali abaphulaphuli ekujoliswe kubo kunomxinwa. Inani elincinci leenkampani zisasebenza kunye neenkqubo eziphezulu zokusebenza. Nangona kwi-Statista ukuba ngo-2021, i-94% yomthwalo wedatha yeziko ledatha yehlabathi iya kuba yinto ebonakalayo. Ngu abahlalutyi, ixabiso lemarike ye-HPC liya kukhula ukusuka kwi-32 ukuya kwi-45 yeebhiliyoni zeedola kwixesha elivela kwi-2017 ukuya kwi-2022.
/ ifoto PD
Izisombululo ezifanayo
Kukho ii-analogue ezininzi kwimarike eziphuhliswa ziinkampani ezinkulu ze-IT: i-AMD kunye ne-Intel.
Inkampani yokuqala ye-GPU virtualization indlela esekelwe kwi-SR-IOV (ingcambu enye / imveliso ebonakalayo). Le teknoloji inika i-VM ukufikelela kwinxalenye yezakhono ze-hardware yenkqubo. Isisombululo sikuvumela ukuba wabelane ngeGPU phakathi kwabasebenzisi be-16 kunye nokusebenza okulinganayo kweenkqubo ezibonakalayo.
Ngokubhekiselele kwisigebenga se-IT sesibini, bona kwi-hypervisor ye-Citrix XenServer 7. Idibanisa umsebenzi womqhubi we-GPU oqhelekileyo kunye nomatshini obonakalayo, ovumela ukuba umva ubonise izicelo ze-3D kunye needesktops kwizixhobo zamakhulu abasebenzisi.
Ikamva lobuchwephesha
Abaphuhlisi be-GPU ebonakalayo ekuphunyezweni kweenkqubo ze-AI kunye nokuthandwa okukhulayo kwezisombululo eziphezulu zokusebenza kwimarike yeteknoloji yoshishino. Banethemba lokuba imfuno yokucubungula inani elikhulu ledatha iya kwandisa imfuno ye-vGPUs.
Ngoku abavelisi ukudibanisa ukusebenza kwe-CPU kunye ne-GPU kumbindi omnye ukukhawulezisa ukusombulula iingxaki ezinxulumene nemizobo, ukwenza izibalo zemathematika, ukusebenza okunengqiqo, kunye nokucubungula idatha. Ukubonakala kwee-cores ezinjalo kwimarike kwixesha elizayo kuya kutshintsha indlela yokwenza izinto ezibonakalayo kunye nokusabalalisa kwabo phakathi komthwalo wemisebenzi kwiindawo ezibonakalayo kunye nefu.
Yintoni ekufuneka uyifunde ngesihloko kwibhlog yethu yeshishini:
Izithuba ezimbalwa kwijelo lethu leTelegram:
umthombo: www.habr.com