{"id":36902,"date":"2019-10-31T22:14:38","date_gmt":"2019-10-31T19:14:38","guid":{"rendered":"https:\/\/prohoster.info\/blog\/vgpu-ispolzovat-nelzya-ignorirovat\/"},"modified":"2019-10-31T22:14:38","modified_gmt":"2019-10-31T19:14:38","slug":"vgpu-ispolzovat-nelzya-ignorirovat","status":"publish","type":"post","link":"https:\/\/prohoster.info\/cs\/blog\/administrirovanie\/vgpu-ispolzovat-nelzya-ignorirovat","title":{"rendered":"vGPU \u2013 nelze ignorovat","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"

\"vGPU
\n
\nV \u010dervnu a \u010dervenci se na n\u00e1s obr\u00e1tilo t\u00e9m\u011b\u0159 dvacet spole\u010dnost\u00ed, kter\u00e9 m\u011bly z\u00e1jem o mo\u017enosti virtu\u00e1ln\u00edch GPU. S technologi\u00ed \u201eGrafika\u201c od Cloud4Y ji\u017e vyu\u017e\u00edv\u00e1 jedna z velk\u00fdch dcer Sberbank, ale celkov\u011b tato slu\u017eba nen\u00ed p\u0159\u00edli\u0161 popul\u00e1rn\u00ed. Tato aktivita n\u00e1s tedy pot\u011b\u0161ila. Vzhledem k rostouc\u00edmu z\u00e1jmu o technologii jsme se rozhodli podrobn\u011bji p\u0159edstavit vGPU.<\/i> <\/a><\/noindex><\/p>\n

Data lakes, z\u00edskan\u00e9 v d\u016fsledku v\u011bdeck\u00fdch experiment\u016f a v\u00fdzkum\u016f, Deep Learning a dal\u0161\u00ed sm\u011bry pr\u00e1ce s AI, modelov\u00e1n\u00ed velk\u00fdch a slo\u017eit\u00fdch objekt\u016f \u2014 to v\u0161e vy\u017eaduje vysoce v\u00fdkonn\u00fd hardware. Je dobr\u00e9, pokud je k dispozici a umo\u017e\u0148uje rychle \u0159e\u0161it aktu\u00e1ln\u00ed \u00fakoly. Jen\u017ee kv\u016fli rostouc\u00ed v\u00fdpo\u010detn\u00ed slo\u017eitosti \u00faloh (zejm\u00e9na se to t\u00fdk\u00e1 obchodn\u00ed analytiky, renderov\u00e1n\u00ed, DL-algoritm\u016f a r\u00e1mc\u016f) se hardwarov\u00e9 zdroje stoln\u00edch a dokonce serverov\u00fdch CPU st\u00e1le \u010dast\u011bji st\u00e1vaj\u00ed neefektivn\u00edmi.<\/p>\n

\u0158e\u0161en\u00ed bylo nalezeno v pou\u017eit\u00ed v\u00fdpo\u010dt\u016f na GPU. Tato technologie urychlen\u00ed grafiky umo\u017e\u0148uje rozd\u011blen\u00ed zdroj\u016f jednoho grafick\u00e9ho procesoru mezi v\u00edce virtu\u00e1ln\u00edmi po\u010d\u00edta\u010di. GPU byl p\u016fvodn\u011b navr\u017een pro pr\u00e1ci s grafikou, a proto se skl\u00e1d\u00e1 z tis\u00edc\u016f drobn\u00fdch jader, kter\u00e1 jsou pou\u017e\u00edv\u00e1na pro efektivn\u00ed zpracov\u00e1n\u00ed paraleln\u00edch \u00faloh. Na GPU se prov\u00e1d\u00ed \u010d\u00e1st nejn\u00e1ro\u010dn\u011bj\u0161\u00edch v\u00fdpo\u010dt\u016f, zbytek zaji\u0161\u0165uje CPU.<\/p>\n

\"vGPU
\n
\nV\u00fdpo\u010dty pomoc\u00ed GPU vymyslela spole\u010dnost Nvidia <\/a><\/noindex>ji\u017e v roce 2007. Dnes tato technologie dosp\u011bla na novou \u00farove\u0148 a vyu\u017e\u00edv\u00e1 se v datov\u00fdch centrech nejv\u011bt\u0161\u00edch podnik\u016f a v\u011bdeck\u00fdch laborato\u0159\u00ed. Tradi\u010dn\u00ed p\u0159\u00edstup m\u00e1 v\u0161ak jeden v\u00fdznamn\u00fd nedostatek: n\u00e1kup fyzick\u00e9ho vybaven\u00ed vych\u00e1z\u00ed velmi draho. A pokud si vzpomeneme na rychlost zastar\u00e1v\u00e1n\u00ed \u201ehardware\u201c, situace se st\u00e1v\u00e1 je\u0161t\u011b nelichotiv\u011bj\u0161\u00ed.<\/p>\n

Tento probl\u00e9m m\u00e1 vy\u0159e\u0161it technologie virtu\u00e1ln\u00edch grafick\u00fdch procesor\u016f: vGPU. S jej\u00ed pomoc\u00ed mohou u\u017eivatel\u00e9 vzd\u00e1len\u011b spou\u0161t\u011bt n\u00e1ro\u010dn\u00e9 aplikace jako AutoCAD, 3DS Max, Maya, Sony Vegas Pro. Virtualizace rychle z\u00edskala sv\u016fj pod\u00edl na trhu. Kter\u00fd rusk\u00fd datov\u00fd v\u011bdec by nemiloval rychl\u00e9 v\u00fdpo\u010dty na grafick\u00fdch kart\u00e1ch NVidia Tesla? <\/p>\n

Je t\u0159eba poznamenat, \u017ee p\u0159ed p\u0159\u00edchodem vGPU se pou\u017e\u00edvaly jin\u00e9 metody zrychlen\u00ed zpracov\u00e1n\u00ed grafiky: Virtual Shared Graphics Acceleration (vSGA) a Virtual Dedicated Graphics Acceleration (vDGA). \u0158e\u0161en\u00ed vGPU spojilo to nejlep\u0161\u00ed z obou technologi\u00ed. Stejn\u011b jako v p\u0159\u00edpad\u011b vSGA se v prost\u0159ed\u00ed vGPU p\u0159edpokl\u00e1d\u00e1 sd\u00edlen\u00ed GPU a RAM mezi n\u011bkolika virtu\u00e1ln\u00edmi pracovn\u00edmi stanicemi, ale ka\u017ed\u00e1 VM pos\u00edl\u00e1 p\u0159\u00edkazy p\u0159\u00edmo na GPU, jako v p\u0159\u00edpad\u011b vDGA.<\/p>\n

Pro\u010d jsou vGPU v\u016fbec pot\u0159eba? <\/h2>\n

\nCloudov\u00e9 v\u00fdpo\u010dty s vyu\u017eit\u00edm vGPU umo\u017e\u0148uj\u00ed spole\u010dnostem zvl\u00e1dat \u00fakoly, kter\u00e9 d\u0159\u00edve nebylo mo\u017en\u00e9 \u0159e\u0161it. Nebo bylo mo\u017en\u00e9, ale vy\u017eadovalo to neskute\u010dn\u00e9 mno\u017estv\u00ed zdroj\u016f. Jeden modern\u00ed GPU server dok\u00e1\u017ee nahradit a\u017e 100 b\u011b\u017en\u00fdch CPU. Existuj\u00ed i dal\u0161\u00ed, je\u0161t\u011b impozantn\u011bj\u0161\u00ed \u010d\u00edsla<\/a><\/noindex>. To nen\u00ed vtip: \u0159e\u0161en\u00ed Nvidia zpracov\u00e1vaj\u00ed petabajty dat n\u011bkolikr\u00e1t rychleji ne\u017e klasick\u00e9 CPU servery. A nap\u0159\u00edklad Google Cloud nab\u00edz\u00ed virtu\u00e1ln\u00ed stroje s GPU, kter\u00e9 dosahuj\u00ed a\u017e 960 teraflop\u016f.<\/p>\n

Mnoz\u00ed specialist\u00e9 pot\u0159ebuj\u00ed v\u00fdkonn\u00e1 za\u0159\u00edzen\u00ed, kter\u00e1 dok\u00e1\u017eou prov\u00e1d\u011bt paraleln\u00ed v\u00fdpo\u010dty. Architekti a in\u017een\u00fd\u0159i vyu\u017e\u00edvaj\u00ed technologii vGPU v n\u00e1vrh\u00e1\u0159sk\u00fdch syst\u00e9mech (t\u0159eba Autodesk). Design\u00e9\u0159i pracuj\u00ed s digit\u00e1ln\u00edm fotografick\u00fdm a video obsahem (Photoshop, CorelDraw). Virtu\u00e1ln\u00ed stroje<\/a><\/noindex> Grafick\u00e9 procesory jsou pot\u0159eba i zdravotnick\u00fdm za\u0159\u00edzen\u00edm, kter\u00e1 shroma\u017e\u010fuj\u00ed a analyzuj\u00ed data o pacientech a onemocn\u011bn\u00edch. Pracuje s GPU i \"Yandex<\/a><\/noindex>\u00bb.<\/p>\n

Mysl\u00edte, \u017ee je to v\u0161e? Kdepak. Technologii vyu\u017e\u00edvaj\u00ed i pro automatickou kontrolu p\u0159esnosti satelitn\u00edch sn\u00edmk\u016f<\/a><\/noindex>, a pro predikci epidemi\u00ed<\/a><\/noindex>, meteorologick\u00e9 v\u00fdzkumy<\/a><\/noindex>, modelov\u00e1n\u00ed slune\u010dn\u00edch cykl\u016f<\/a><\/noindex> a byznysovou analytiku<\/a><\/noindex>. A je\u0161t\u011b existuje skv\u011bl\u00fd \u010dl\u00e1nek o modelov\u00e1n\u00ed<\/a><\/noindex> v prost\u0159ed\u00ed Unity3D od ThisIsZolden<\/a><\/noindex>.<\/p>\n

I p\u0159es to v\u0161e zat\u00edm \u0159e\u0161en\u00ed na b\u00e1zi vGPU nez\u00edskala \u0161irok\u00e9 roz\u0161\u00ed\u0159en\u00ed na sv\u011bt\u011b. V roce 2018 provedla spole\u010dnost NetApp anketu <\/a><\/noindex>mezi firmami, kter\u00e9 ve sv\u00e9 pr\u00e1ci vyu\u017e\u00edvaj\u00ed grafick\u00e9 procesory. V\u00fdsledky uk\u00e1zaly, \u017ee 60 % organizac\u00ed st\u00e1le pracuje na vlastn\u00ed IT infrastruktu\u0159e. \"Cloud\" pou\u017e\u00edv\u00e1 pouze 23 %. V Rusku m\u00e1 pr\u016fnik technologie cloudov\u00fdch v\u00fdpo\u010dt\u016f men\u0161\u00ed v\u00fdznam. Ale d\u00edky nov\u00fdm hardwarov\u00fdm a softwarov\u00fdm \u0159e\u0161en\u00edm po\u010det spole\u010dnost\u00ed vyu\u017e\u00edvaj\u00edc\u00edch virtu\u00e1ln\u00ed stroje s GPU st\u00e1le roste.<\/p>\n

\u0158e\u0161en\u00ed pro vGPU<\/h2>\n

\n\"vGPU
\n
\nV\u00fdvojem technologi\u00ed virtualizace grafick\u00fdch akceler\u00e1tor\u016f se zab\u00fdv\u00e1 mnoho spole\u010dnost\u00ed, ale mezi nimi jsou bezesporu i l\u00edd\u0159i.<\/p>\n

Jedn\u00edm z nejuzn\u00e1van\u011bj\u0161\u00edch v\u00fdvoj\u00e1\u0159\u016f \u0159e\u0161en\u00ed v oblasti cloudov\u00fdch technologi\u00ed je spole\u010dnost VMware <\/b>nab\u00edzej\u00edc\u00ed hypervizor ESXi<\/a><\/noindex>, pod n\u00edm\u017e je rychlost virtualizovan\u00fdch grafick\u00fdch procesor\u016f srovnateln\u00e1 s implementacemi na hol\u00e9m hardwaru. V ned\u00e1vn\u00e9 aktualizaci v\u00fdvoj\u00e1\u0159 deaktivoval vyva\u017eova\u010d z\u00e1t\u011b\u017ee vMotion a p\u0159idal podporu technologie DirectPath I\/O, kter\u00e1 spojuje ovlada\u010d CUDA s VM mimo hypervizor a urychluje p\u0159enos dat.<\/p>\n

Nvidia <\/b>se tak\u00e9 sna\u017e\u00ed splnit o\u010dek\u00e1v\u00e1n\u00ed trhu a k tomu vydala opensource platformu Rapids<\/a><\/noindex>. \u0158e\u0161en\u00ed spojuje n\u011bkolik knihoven pro pr\u00e1ci s architekturou CUDA, co\u017e usnad\u0148uje pr\u00e1ci s daty b\u011bhem tr\u00e9nov\u00e1n\u00ed neuronov\u00fdch s\u00edt\u00ed a umo\u017e\u0148uje automatizaci pr\u00e1ce s Python k\u00f3dem. Pou\u017eit\u00ed Rapids s algoritmem strojov\u00e9ho u\u010den\u00ed XGBoost p\u0159in\u00e1\u0161\u00ed 50n\u00e1sobn\u00e9 zv\u00fd\u0161en\u00ed v\u00fdkonu ve srovn\u00e1n\u00ed se syst\u00e9my na b\u00e1zi CPU.<\/p>\n

Vlastn\u00ed technologii m\u00e1 i AMD<\/b>. Platforma se naz\u00fdv\u00e1 ROCm<\/a><\/noindex>. Pou\u017e\u00edv\u00e1 technologii SR-IOV, kter\u00e1 rozd\u011bluje hardwarov\u00e9 mo\u017enosti fyzick\u00e9ho za\u0159\u00edzen\u00ed mezi n\u011bkolik virtu\u00e1ln\u00edch stroj\u016f. Zdroje jednoho akceler\u00e1toru lze rozd\u011blit mezi \u0161estn\u00e1ct u\u017eivatel\u016f a udr\u017eovat stejn\u00fd v\u00fdkon pro ka\u017ed\u00e9ho z nich. T\u00edm se zrychluje p\u0159enos dat mezi cloudov\u00fdmi CPU a GPU. Tak\u00e9 se pou\u017e\u00edv\u00e1 speci\u00e1ln\u00ed dialekt C++ nazvan\u00fd HIP, kter\u00fd usnad\u0148uje prov\u00e1d\u011bn\u00ed matematick\u00fdch operac\u00ed na GPU.<\/p>\n

Intel <\/b>stav\u00ed svou technologii na b\u00e1zi multiplatformn\u00edho hypervizoru Citrix <\/a><\/noindex>XenServer 7, kter\u00fd v roce 2017 z\u00edskal certifik\u00e1t shody od FST\u00c9K. \u0158e\u0161en\u00ed spojuje pr\u00e1ci standardn\u00edho GPU ovlada\u010de a virtu\u00e1ln\u00edho stroje. To znamen\u00e1, \u017ee \"virtu\u00e1lka\" m\u016f\u017ee udr\u017eovat <\/a><\/noindex>prov\u00e1d\u011bt n\u00e1ro\u010dn\u00e9 aplikace na za\u0159\u00edzen\u00edch s velk\u00fdm (n\u011bkolika stovkov\u00fdm) mno\u017estv\u00edm u\u017eivatel\u016f.<\/p>\n

Tr\u017en\u00ed perspektivy<\/h2>\n

\n\"vGPU
\n
\nNez\u00e1visl\u00ed analytici se domn\u00edvaj\u00ed, \u017ee objem prodeje \u0159e\u0161en\u00ed pro HPC syst\u00e9my dos\u00e1hne 45 miliard dolar\u016f do roku 2022. V\u00fdvoj\u00e1\u0159i platforem tak\u00e9 o\u010dek\u00e1vaj\u00ed zv\u00fd\u0161en\u00ed popt\u00e1vky po vysoce v\u00fdkonn\u00fdch syst\u00e9mech. To o\u010dek\u00e1v\u00e1n\u00ed je pos\u00edleno popularitou Big Data a \u010dastou pot\u0159ebou zpracov\u00e1vat velk\u00e9 objemy dat.<\/p>\n

R\u016fst popt\u00e1vky po vGPU m\u016f\u017ee rovn\u011b\u017e podn\u00edtit rozvoj hybridn\u00edch technologi\u00ed, kter\u00e9 spojuj\u00ed GPU a CPU v jednom za\u0159\u00edzen\u00ed. V takov\u00fdch integrovan\u00fdch \u0159e\u0161en\u00edch dva typy jader sd\u00edlej\u00ed spole\u010dnou cache, co\u017e urychluje p\u0159enos dat mezi grafick\u00fdmi a tradi\u010dn\u00edmi procesory. <\/p>\n

Hybridn\u00ed technologie mohou z\u00e1sadn\u011b zm\u011bnit p\u0159\u00edstup k virtualizaci a rozd\u011blen\u00ed virtu\u00e1ln\u00edch zdroj\u016f v r\u00e1mci datov\u00fdch center. \u0158e\u0161en\u00ed s otev\u0159en\u00fdm zdrojov\u00fdm k\u00f3dem, jako jsou ROCm a Rapids, umo\u017e\u0148uj\u00ed oper\u00e1tor\u016fm datov\u00fdch center efektivn\u011bji vyu\u017e\u00edvat v\u00fdpo\u010detn\u00ed zdroje a zvy\u0161ovat v\u00fdkon za\u0159\u00edzen\u00ed.<\/p>\n

Existuje i jin\u00fd n\u00e1zor. Nap\u0159\u00edklad, \u017ee virtu\u00e1ln\u00ed GPU budou vytla\u010dov\u00e1ny optick\u00fdmi \u010dipy s fotonov\u00fdm k\u00f3dov\u00e1n\u00edm dat. Takov\u00e1 \u0159e\u0161en\u00ed ji\u017e existuj\u00ed a pou\u017e\u00edvaj\u00ed se pro strojov\u00e9 u\u010den\u00ed. Nav\u00edc se zdaj\u00ed b\u00fdt v\u00fdkonn\u011bj\u0161\u00ed <\/a><\/noindex>ne\u017e b\u011b\u017en\u00e9 GPU. Ale technologie je st\u00e1le v ran\u00e9 f\u00e1zi.<\/p>\n

Jak\u00fd z\u00e1v\u011br z toho vypl\u00fdv\u00e1? I p\u0159es mo\u017en\u00fd vznik alternativ je vGPU zcela perspektivn\u00edm sm\u011brem, kter\u00fd m\u016f\u017ee \u0159e\u0161it \u0159adu \u00faloh. Ale ne ka\u017ed\u00fd m\u00e1 na to, aby je vyu\u017e\u00edval. Tak\u017ee \u010d\u00e1rku v nadpisu si m\u016f\u017eete d\u00e1t sami.<\/p>\n

P.S.<\/b>
\nSledujte n\u00e1\u0161 Telegram<\/a><\/noindex>-kan\u00e1l, abyste nep\u0159i\u0161li o dal\u0161\u00ed \u010dl\u00e1nek! P\u00ed\u0161eme maxim\u00e1ln\u011b dvakr\u00e1t t\u00fddn\u011b a pouze relevantn\u011b.<\/i><\/p>\n

Pouze registrovan\u00ed u\u017eivatel\u00e9 se mohou \u00fa\u010dastnit ankety. P\u0159ihlaste se<\/a><\/noindex>, pros\u00edm.<\/p>\n

Pou\u017e\u00edv\u00e1te vGPU?<\/h2>\n