{"id":42117,"date":"2019-03-18T00:00:00","date_gmt":"2019-03-17T21:00:00","guid":{"rendered":"https:\/\/prohoster.info\/blog\/blog_prohoster\/situatsiya-virtualnye-gpu-ne-ustupayut-po-proizvoditelnosti-zheleznym-resheniyam"},"modified":"2020-02-18T13:44:55","modified_gmt":"2020-02-18T10:44:55","slug":"situatsiya-virtualnye-gpu-ne-ustupayut-po-proizvoditelnosti-zheleznym-resheniyam","status":"publish","type":"post","link":"https:\/\/prohoster.info\/cs\/blog\/administrirovanie\/situatsiya-virtualnye-gpu-ne-ustupayut-po-proizvoditelnosti-zheleznym-resheniyam","title":{"rendered":"Situace: virtu\u00e1ln\u00ed GPU se v\u00fdkonem rovnaj\u00ed hardwarov\u00fdm \u0159e\u0161en\u00edm","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"

V \u00fanoru se na Stanfordu konala konference v\u011bnovan\u00e1 vysoce v\u00fdkonn\u00e9mu po\u010d\u00edt\u00e1n\u00ed (HPC). P\u0159edstavitel\u00e9 VMware uvedli, \u017ee p\u0159i pr\u00e1ci s GPU syst\u00e9m zalo\u017een\u00fd na modifikovan\u00e9m hypervizoru ESXi v\u00fdkonem nijak nezaost\u00e1v\u00e1 za \u0159e\u0161en\u00edmi typu bare metal.<\/p>\n

Popisujeme technologie, kter\u00e9 k tomu pomohly.<\/p>\n

\"Situace:<\/a><\/noindex><\/a><\/noindex>
\n\/ \u0444\u043e\u0442\u043e Victorgrigas<\/a><\/noindex> CC BY-SA<\/a><\/noindex><\/i><\/p>\n

Probl\u00e9m v\u00fdkonu<\/h2>\n

\nPodle anal\u00fdz se p\u0159edpokl\u00e1d\u00e1, \u017ee p\u0159ibli\u017en\u011b 70 % pracovn\u00edch z\u00e1t\u011b\u017e\u00ed v datov\u00fdch centrech je virtualizov\u00e1no<\/a><\/noindex>. Nicm\u00e9n\u011b zbyl\u00fdch 30 % st\u00e1le b\u011b\u017e\u00ed na bare metal bez hypervizor\u016f. Tento pod\u00edl v\u011bt\u0161inou tvo\u0159\u00ed vysoce zat\u00ed\u017een\u00e9 aplikace, spojen\u00e9 nap\u0159\u00edklad s tr\u00e9nov\u00e1n\u00edm neuronov\u00fdch s\u00edt\u00ed, a vyu\u017e\u00edvaj\u00edc\u00ed grafick\u00e9 procesory.<\/p>\n

Odborn\u00edci tento trend vysv\u011btluj\u00ed t\u00edm, \u017ee hypervizor jako mezivrstva abstrakce m\u016f\u017ee ovlivnit v\u00fdkon cel\u00e9ho syst\u00e9mu. V p\u011bt let star\u00fdch studi\u00edch nalezneme \u00fadaje<\/a><\/noindex> o poklesu rychlosti pr\u00e1ce o 10 %. Proto se spole\u010dnosti a provozovatel\u00e9 datov\u00fdch center nechystaj\u00ed p\u0159ech\u00e1zet HPC z\u00e1t\u011b\u017e do virtu\u00e1ln\u00edho prost\u0159ed\u00ed.<\/p>\n

Technologie virtualizace se v\u0161ak vyv\u00edjej\u00ed a zlep\u0161uj\u00ed. Na konferenci p\u0159ed m\u011bs\u00edcem bylo ve VMware sd\u011bleno, \u017ee hypervizor ESXi nem\u00e1 negativn\u00ed vliv na v\u00fdkon GPU. Rychlost v\u00fdpo\u010dt\u016f m\u016f\u017ee klesnout o t\u0159i procenta, co\u017e je srovnateln\u00e9 s ukazateli bare metal.<\/p>\n

Jak to funguje<\/h2>\n

\nAby se zv\u00fd\u0161il v\u00fdkon HPC syst\u00e9m\u016f s grafick\u00fdmi procesory, VMware provedl v pr\u00e1ci hypervizoru \u0159adu zm\u011bn. Zvl\u00e1\u0161t\u011b se zbavili funkce vMotion. Ta je pot\u0159ebn\u00e1 pro vyva\u017eov\u00e1n\u00ed z\u00e1t\u011b\u017ee a obvykle p\u0159esouv\u00e1 virtu\u00e1ln\u00ed stroje (VM) mezi servery nebo GPU. Vypnut\u00ed vMotion vedlo k tomu, \u017ee k ka\u017ed\u00e9mu VM je nyn\u00ed p\u0159id\u011blen konkr\u00e9tn\u00ed grafick\u00fd procesor. To pomohlo sn\u00ed\u017eit n\u00e1klady p\u0159i v\u00fdm\u011bn\u011b dat.<\/p>\n

Dal\u0161\u00ed kl\u00ed\u010dovou sou\u010d\u00e1st\u00ed syst\u00e9mu je technologie<\/a><\/noindex> DirectPath I\/O. Ta umo\u017e\u0148uje ovlada\u010di CUDA pro paraleln\u00ed v\u00fdpo\u010dty komunikovat s virtu\u00e1ln\u00edmi stroji p\u0159\u00edmo, obch\u00e1zej\u00edc hypervizor. Kdy\u017e je na jednom GPU pot\u0159eba spustit v\u00edce VM, je pou\u017eito \u0159e\u0161en\u00ed GRID vGPU. To d\u011bl\u00ed pam\u011b\u0165 karty na n\u011bkolik segment\u016f (ale v\u00fdpo\u010detn\u00ed cykly nejsou rozd\u011bleny).<\/p>\n

Sch\u00e9ma pr\u00e1ce dvou virtu\u00e1ln\u00edch stroj\u016f v tomto p\u0159\u00edpad\u011b bude vypadat takto:<\/p>\n

\"Situace:<\/p>\n

V\u00fdsledky a p\u0159edpov\u011bdi<\/h2>\n

\nSpole\u010dnost provedla testy<\/a><\/noindex> hypervisoru, kter\u00fd vy\u0161kolil jazykov\u00fd model na z\u00e1klad\u011b TensorFlow<\/a><\/noindex>. 'Ztr\u00e1ta' v\u00fdkonu \u010dinila pouze 3\u20134% v porovn\u00e1n\u00ed s bare metal. Na opl\u00e1tku syst\u00e9m z\u00edskal mo\u017enost distribuovat zdroje podle pot\u0159eby v z\u00e1vislosti na aktu\u00e1ln\u00edch z\u00e1t\u011b\u017e\u00edch.<\/p>\n

IT gigant tak\u00e9 prov\u00e1d\u011bl testy<\/a><\/noindex> s kontejnery. In\u017een\u00fd\u0159i spole\u010dnosti vyu\u010dovali neuronov\u00e9 s\u00edt\u011b rozpozn\u00e1vat obr\u00e1zky. P\u0159itom byly zdroje jednoho grafick\u00e9ho procesoru rozd\u011bleny mezi \u010dty\u0159i kontejnerov\u00e9 VM. V d\u016fsledku toho v\u00fdkon jednotliv\u00fdch stroj\u016f klesl o 17% (ve srovn\u00e1n\u00ed s jednou VM, kter\u00e1 m\u011bla pln\u00fd p\u0159\u00edstup k prost\u0159edk\u016fm GPU). Nicm\u00e9n\u011b, mno\u017estv\u00ed zpracov\u00e1van\u00fdch obr\u00e1zk\u016f za sekundu vzrostlo<\/a><\/noindex> trojn\u00e1sobn\u011b. O\u010dek\u00e1v\u00e1 se, \u017ee podobn\u00e9 syst\u00e9my naleznou<\/a><\/noindex> uplatn\u011bn\u00ed v oblasti anal\u00fdzy dat a po\u010d\u00edta\u010dov\u00e9ho modelov\u00e1n\u00ed.<\/p>\n

Mezi potenci\u00e1ln\u00edmi probl\u00e9my, se kter\u00fdmi se m\u016f\u017ee VMware setkat, odborn\u00edci vyzdvihuj\u00ed<\/a><\/noindex> pom\u011brn\u011b \u00fazkou c\u00edlovou skupinu. S vysoce v\u00fdkonn\u00fdmi syst\u00e9my moment\u00e1ln\u011b pracuje pouze mal\u00e9 mno\u017estv\u00ed firem. P\u0159esto\u017ee v Statista uv\u00e1d\u011bj\u00ed<\/a><\/noindex>, \u017ee do roku 2021 bude virtualizov\u00e1no ji\u017e 94% pracovn\u00edch z\u00e1t\u011b\u017e\u00ed glob\u00e1ln\u00edch datov\u00fdch center. Podle p\u0159edpov\u011bd\u00ed<\/a><\/noindex> analytik\u016f, cena trhu HPC vzroste z 32 na 45 miliard dolar\u016f v obdob\u00ed od 2017 do 2022.<\/p>\n

\"Situace:
\n\/ \u0444\u043e\u0442\u043e Global Access Point<\/a><\/noindex> PD<\/i><\/p>\n

Podobn\u00e1 \u0159e\u0161en\u00ed<\/h2>\n

\nNa trhu existuje n\u011bkolik analog\u016f, kter\u00e9 vyv\u00edjej\u00ed velk\u00e9 IT spole\u010dnosti: AMD a Intel.<\/p>\n

Prvn\u00ed spole\u010dnost pro virtualizaci grafick\u00fdch procesor\u016f nab\u00edz\u00ed<\/noindex> p\u0159\u00edstup zalo\u017een\u00fd na SR-IOV (single-root input\/output virtualization). Tato technologie poskytuje VM p\u0159\u00edstup k \u010d\u00e1sti hardwarov\u00fdch mo\u017enost\u00ed syst\u00e9mu. \u0158e\u0161en\u00ed umo\u017e\u0148uje rozd\u011blit grafick\u00fd procesor mezi 16 u\u017eivateli p\u0159i stejn\u00e9 v\u00fdkonu virtualizovan\u00fdch syst\u00e9m\u016f.<\/p>\n

Co se t\u00fd\u010de druh\u00e9ho IT giganta, jejich technologie se zakl\u00e1d\u00e1<\/a><\/noindex> na hypervisoru Citrix XenServer 7. Ten spojuje pr\u00e1ci standardn\u00edho GPU ovlada\u010de a virtu\u00e1ln\u00edho stroje, co\u017e umo\u017e\u0148uje tomuto druh\u00e9mu zobrazovat 3D aplikace a plochy na za\u0159\u00edzen\u00edch stovek u\u017eivatel\u016f.<\/p>\n

Budoucnost technologie<\/h2>\n

\nV\u00fdvoj\u00e1\u0159i virtu\u00e1ln\u00edch grafick\u00fdch procesor\u016f s\u00e1z\u00ed na<\/a><\/noindex> na zaveden\u00ed syst\u00e9m\u016f AI a r\u016fst popularity vysoce v\u00fdkonn\u00fdch \u0159e\u0161en\u00ed na trhu byznysov\u00fdch technologi\u00ed. Doufaj\u00ed, \u017ee popt\u00e1vka po zpracov\u00e1n\u00ed velk\u00fdch objem\u016f dat zv\u00fd\u0161\u00ed popt\u00e1vku po vGPU.<\/p>\n

Nyn\u00ed v\u00fdrobci hledaj\u00ed zp\u016fsob<\/a><\/noindex> jak spojit funk\u010dnost CPU a GPU v jednom j\u00e1dru, aby urychlili \u0159e\u0161en\u00ed \u00faloh souvisej\u00edc\u00edch s grafikou, prov\u00e1d\u011bn\u00ed matematick\u00fdch v\u00fdpo\u010dt\u016f, logick\u00fdch operac\u00ed a zpracov\u00e1n\u00ed dat. Objevov\u00e1n\u00ed takov\u00fdch jader na trhu v budoucnu zm\u011bn\u00ed p\u0159\u00edstup k virtualizaci zdroj\u016f a jejich rozd\u011blen\u00ed mezi pracovn\u00edmi z\u00e1t\u011b\u017eemi ve virtu\u00e1ln\u00edm a cloudov\u00e9m prost\u0159ed\u00ed.<\/p>\n

\nCo si p\u0159e\u010d\u00edst na toto t\u00e9ma v na\u0161em firemn\u00edm blogu:<\/p>\n