Kukholakala ukuthi amaseva abonakalayo ane-vGPU ayabiza. Ngokubuyekezwa okufushane ngizozama ukuphikisa le thesis.
Ukusesha ku-inthanethi kuveza ngokushesha ukuqashwa kwamakhompyutha amakhulu ane-NVIDIA Tesla V100 noma amaseva alula anama-GPU azinikele anamandla. Amasevisi afanayo ayatholakala, isibonelo, , noma . Izindleko zabo zanyanga zonke zikalwa ngamashumi ezinkulungwane zama-ruble, futhi bengifuna ukuthola izinketho ezishibhile ze-OpenCL kanye/noma izicelo ze-CUDA. Awekho ama-VPS amaningi ebhajethi enama-adaptha wevidiyo emakethe yaseRussia; esihlokweni esifushane ngizoqhathanisa amandla abo e-computing usebenzisa izivivinyo zokwenziwa.
abahlanganyeli
Amaseva abambayo afakiwe kuhlu lwamakhandidethi ukuze abambe iqhaza ekubuyekezweni. , , , ΠΈ . Azikho izinkinga ezithile ekutholeni ukufinyelela, njengoba cishe bonke abahlinzeki banesikhathi sesilingo samahhala. I-UltraVDS ngokusemthethweni ayinakho ukuhlolwa kwamahhala, kodwa akuzange kube nzima ukufika esivumelwaneni: ngemva kokufunda mayelana nokushicilelwa, izisebenzi zokusekela zanginika inani elidingekayo ukuze ngi-ode i-VPS ku-akhawunti yami yebhonasi. Kulesi sigaba, imishini ebonakalayo ye-VDS4YOU iphumile emjahweni, ngoba ukuze uhlole mahhala umsingathi udinga ukuthi unikeze iskena sekhadi lakho le-ID. Ngiyaqonda ukuthi udinga ukuzivikela ekuhlukunyezweni, kodwa ukuze uqinisekise, imininingwane yepasipoti noma, isibonelo, ukuxhuma i-akhawunti kunethiwekhi yokuxhumana nabantu - lokhu kudingwa yi-1Gb.ru.
Ukucushwa namanani
Ukuze sihlole, sithathe imishini esezingeni eliphakathi ebiza ngaphansi kwama-ruble ayizinkulungwane eziyi-10 ngenyanga: ama-computing cores angu-2, 4 GB we-RAM, 20 - 50 GB SSD, vGPU eno-256 MB VRAM kanye ne-Windows Server 2016. Ngaphambi kokuhlola ukusebenza kwe-VDS, ake sibheke izinhlelo zabo ezingaphansi zezithombe ngokubukeka okuhlomile. Idalwe yinkampani Umbuso ikuvumela ukuthi uthole ulwazi oluningiliziwe mayelana ne-hardware nezixazululo zesofthiwe ezisetshenziswa abasingathi. Ngosizo lwayo ungabona, isibonelo, inguqulo yomshayeli wevidiyo, inani lememori yevidiyo etholakalayo, kanye nedatha ekusekelweni kwe-OpenCL ne-CUDA.
1Gb.ru
I-GPUcloud
I-RuVDS
I-UltraVDS
I-Virtualization
I-Hyper-V
I-OpenStack
I-Hyper-V
I-Hyper-V
Computing cores
2*2,6 GHz
2*2,8 GHz
2*3,4 GHz
2*2,2 GHz
RAM, GB
4
4
4
4
Isitoreji, GB
30 (SSD)
50 (SSD)
20 (SSD)
30 (SSD)
vGPU
I-RemoteFX
I-NVIDIA GRID
I-RemoteFX
I-RemoteFX
I-adaptha yevidiyo
I-NVIDIA GeForce GTX 1080 Ti
I-NVIDIA Tesla T4
I-NVIDIA Quadro P4000
I-AMD FirePro W4300
vRAM, MB
256
4063
256
256
Usekelo lwe-OpenCL
+
+
+
+
Ukusekelwa kwe-CUDA
-
+
-
-
Inani ngenyanga (uma ikhokhwa njalo ngonyaka), hlikihla.
3494 (3015)
7923,60
1904 (1333)
1930 (1351)
Inkokhelo yezinsiza, hlikihla
akukho
CPU = 0,42 rub/ihora,
I-RAM = 0,24 rub/ihora,
I-SSD = 0,0087 rub/ihora,
I-OS Windows = 1,62 rub/ihora,
IPv4 = 0,15 rub/ihora,
vGPU (T4/4Gb) = 7 ruble/ihora.
kusuka ku-623,28 + 30 ngokufaka ngakunye
akukho
Isikhathi sokuhlola
Izinsuku ze-10
7 izinsuku noma ngaphezulu ngokuvumelana
Izinsuku ezi-3 ngenkokhelo yanyanga zonke
akukho
Kubahlinzeki ababuyekeziwe, yi-GPUcloud kuphela esebenzisa i-OpenStack virtualization kanye nobuchwepheshe be-NVIDIA GRID. Ngenxa yenani elikhulu lememori yevidiyo (amaphrofayili angu-4, 8 no-16 GB ayatholakala), isevisi ibiza kakhulu, kodwa iklayenti izosebenzisa izinhlelo zokusebenza ze-OpenCL ne-CUDA. Abanye abaqhudelana nabo banikeza ama-vGPU nge-VRAM encane, adalwe kusetshenziswa i-Microsoft RemoteFX. Abiza kancane kakhulu, kodwa asekela i-OpenCL kuphela.
Ukuhlolwa kokusebenza
IGeekBench 5
Ngalokhu okudumile Ungakwazi ukukala ukusebenza kwezithombe kuzinhlelo zokusebenza ze-OpenCL ne-CUDA. Ishadi elingezansi libonisa umphumela wesifinyezo, nedatha enemininingwane eminingi yamaseva abonakalayo , GPUcloud ( ΠΈ ), ΠΈ iyatholakala kuwebhusayithi yonjiniyela webhentshimakhi. Ukuwavula kuveza iqiniso elithokozisayo: I-GeekBench ikhombisa amanani e-VRAM aphezulu kakhulu kune-256 MB eyaliwe. Ijubane lewashi lamaphrosesa amaphakathi nalo lingase libe phezulu kunalokho okushiwo. Lokhu kuyinto evamile ezindaweni ezibonakalayo - okuningi kuncike emthwalweni ophethe i-VPS esebenza kuwo.
Ama-vGPU βesevaβ okwabelwana ngawo abuthakathaka kunama-adaptha evidiyo βedeskithophuβ asebenza kahle kakhulu uma esetshenziselwa izinhlelo zokusebenza zezithombe ezisindayo. Izixazululo ezinjalo zenzelwe ikakhulukazi imisebenzi yekhompyutha. Ezinye izivivinyo zokwenziwa zenziwa ukuze kuhlolwe ukusebenza kwazo.
I-FAHBench 2.3.1
Ukuze uthole ukuhlaziya okuphelele kwamakhono wekhompyutha we-vGPU ayifaneleki, kodwa ingasetshenziswa ukuqhathanisa ukusebenza kwama-adaptha evidiyo avela ku-VPS ehlukene ngezibalo eziyinkimbinkimbi kusetshenziswa i-OpenCL. Iphrojekthi Yekhompyutha Esabalalisiwe ixazulula inkinga ewumngcingo yokubunjwa kwekhompuyutha yokugoqa ama-molecule amaprotheni. Abacwaningi bazama ukuqonda izimbangela zama-pathologies ahambisana namaprotheni angenasici: Izifo ze-Alzheimer kanye ne-Parkinson, isifo sezinkomo ezihlanya, i-multiple sclerosis, njll. Kukalwa kusetshenziswa insiza abayidalile Ukusebenza okukodwa nokukabili kuboniswa eshadini. Ngeshwa, insiza ikhiqize iphutha emshinini obonakalayo we-UltraVDS.
Okulandelayo, ngizoqhathanisa imiphumela yokubala yendlela ye-dhfr-implicit modeling.
SiSoftware Sandra 20/20
Iphakheji Kuhle kakhulu ekuhloleni amandla okwenza ikhompuyutha ama-adaptha wevidiyo avela kubasingathi abahlukahlukene. Insiza iqukethe amasuite ebhentshimakhi ekhompiyutha enhloso ejwayelekile (GPGPU) futhi isekela i-OpenCL, i-DirectCompute ne-CUDA. Okokuqala, ukuhlolwa okuvamile kwama-vGPU ahlukene kwenziwa. Umdwebo ubonisa umphumela wesifinyezo, idatha enemininingwane eminingi yamaseva abonakalayo , GPUcloud () futhi iyatholakala kuwebhusayithi yonjiniyela webhentshimakhi.
Kuphinde kwaba nezinkinga ngokuhlolwa kukaSandra βokudeβ. Kumhlinzeki we-VPS i-GPUcloud, akukwazanga ukwenza ukuhlola okuvamile usebenzisa i-OpenCL. Lapho ukhetha inketho efanele, insiza isasebenza nge-CUDA. Umshini we-UltraVDS nawo uhlulekile kulokhu kuhlolwa: ibhentshimakhi yabanda ku-86% ngenkathi izama ukunquma ukubambezeleka kwenkumbulo.
Kuphakheji yokuhlola evamile, akunakwenzeka ukubona izinkomba ezinezinga elanele lemininingwane noma ukwenza izibalo ngokunemba okuphezulu. Kwadingeka senze izivivinyo ezimbalwa ezihlukene, siqale ngokuthola ukusebenza okuphezulu okuphezulu kwe-adaptha yevidiyo sisebenzisa isethi yezibalo ezilula zezibalo sisebenzisa i-OpenCL kanye (uma kungenzeka) ne-CUDA. Lokhu futhi kubonisa inkomba evamile kuphela, kanye nemiphumela enemininingwane ye-VPS evela , GPUcloud ( ΠΈ ), ΠΈ etholakala kuwebhusayithi.
Ukuze uqhathanise isivinini sombhalo wekhodi nokuqopha idatha, uSandra unesethi yokuhlolwa kwe-cryptographic. Imiphumela enemininingwane ye , GPUcloud ( ΠΈ ), ΠΈ .
Izibalo zezezimali ezifanayo zidinga ukubalwa kwe-adaptha enembile ephindwe kabili. Lena enye indawo ebalulekile yohlelo lokusebenza lwe-vGPUs. Imiphumela enemininingwane ye , GPUcloud ( ΠΈ ), ΠΈ .
I-Sandra 20/20 ikuvumela ukuthi uhlole amathuba okusebenzisa i-vGPU ekubaleni kwesayensi ngokunemba okuphezulu: ukuphindaphinda kwe-matrix, ukuguqulwa okusheshayo kwe-Fourier, njll. Imiphumela enemininingwane ye , GPUcloud ( ΠΈ ), ΠΈ .
Ekugcineni, ukuhlolwa kwamakhono okucubungula izithombe kwe-vGPU kwenziwa. Imiphumela enemininingwane ye , GPUcloud ( ΠΈ ), ΠΈ .
okutholakele
Iseva ebonakalayo ye-GPUcloud ibonise imiphumela emihle kakhulu ekuhlolweni kwe-GeekBench 5 ne-FAHBench, kodwa ayizange ikhuphuke ngaphezu kwezinga elivamile ekuhlolweni kwebhentshimakhi kaSandra. Ibiza kakhulu kunezinsizakalo zezimbangi, kodwa inenani elikhulu kakhulu lememori yevidiyo futhi isekela i-CUDA. Ezivivinyweni zikaSandra, i-VPS evela ku-1Gb.ru yayingumholi onokunemba okuphezulu kwezibalo, kodwa futhi ayishibhile futhi yenziwa ngokwesilinganiso kwezinye izivivinyo. I-UltraVDS ivele yaba ngumuntu wangaphandle osobala: Angazi ukuthi kukhona yini ukuxhumana lapha, kodwa lo msingathi kuphela onikeza amaklayenti amakhadi evidiyo e-AMD. Ngokwesilinganiso senani/sokusebenza, iseva ye-RuVDS ibonakala kimi njengengcono kakhulu. Kubiza ngaphansi kwama-ruble ayi-2000 XNUMX ngenyanga, futhi izivivinyo ziphase kahle impela. Ukuma kokugcina kubukeka kanje:
Indawo
Umsingathi
Usekelo lwe-OpenCL
Ukusekelwa kwe-CUDA
Ukusebenza okuphezulu ngokuya nge-GeekBench 5
Ukusebenza okuphezulu ngokusho kwe-FAHBench
Ukusebenza okuphezulu ngokusho kukaSandra 20/20
Intengo ephansi
I
I-RuVDS
+
-
+
+
+
+
II
1Gb.ru
+
-
+
+
+
+
III
I-GPUcloud
+
+
+
+
+
-
IV
I-UltraVDS
+
-
-
-
-
+
Nganginokungabaza mayelana nomnqobi, kodwa ukubuyekezwa kunikezelwe kubhajethi ye-VPS nge-vGPU, futhi umshini we-RuVDS virtual ubiza cishe uhhafu kunombangi wawo oseduze futhi izikhathi ezingaphezu kwezine kunesipho esibiza kakhulu esibuyekeziwe. Indawo yesibili neyesithathu nayo bekungelula ukuyihlukanisa, kodwa nalapha intengo ibidlula ezinye izici.
Njengomphumela wokuhlola, kuvele ukuthi ama-vGPU ezinga lokungena awabizi kangako futhi angasetshenziswa kakade ukuxazulula izinkinga zekhompyutha. Yiqiniso, ukusebenzisa izivivinyo zokwenziwa kunzima ukubikezela ukuthi umshini uzoziphatha kanjani ngaphansi komthwalo wangempela, futhi ngaphandle kwalokho, ikhono lokwaba izinsiza ngokuqondile lincike kumakhelwane walo kumphathi womzimba - yenza izibonelelo zalokhu. Uma uthola enye i-VPS yesabelomali nge-vGPU ku-Inthanethi yaseRussia, ungangabazi ukubhala ngabo kumazwana.
Source: www.habr.com