{"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\/sq\/blog\/administrirovanie\/situatsiya-virtualnye-gpu-ne-ustupayut-po-proizvoditelnosti-zheleznym-resheniyam","title":{"rendered":"Situata: GPU virtuale nuk i n\u00ebnshtrohen performanc\u00ebs s\u00eb zgjidhjeve harduerike","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"

N\u00eb shkurt, n\u00eb Stanford u mbajt nj\u00eb konferenc\u00eb, e cila trajtoi llogarit\u00eb me performanc\u00eb t\u00eb lart\u00eb (HPC). P\u00ebrfaq\u00ebsuesit e VMware njoftuan se, kur punojn\u00eb me GPU, sistemi i bazuar n\u00eb t\u00eb modifikuar hipervizorin ESXi nuk i n\u00ebnshtrohet ndryshimeve t\u00eb shpejt\u00ebsis\u00eb p\u00ebrball\u00eb zgjidhjeve bare metal.<\/p>\n

Ne flasim p\u00ebr teknologjit\u00eb q\u00eb e b\u00ebn\u00eb k\u00ebt\u00eb t\u00eb mundur.<\/p>\n

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

Problemi i performanc\u00ebs<\/h2>\n

\nSipas analist\u00ebve, rreth 70% e ngarkesave t\u00eb pun\u00ebs n\u00eb qendrat e t\u00eb dh\u00ebnave jan\u00eb virtualizuar<\/a><\/noindex>. Megjithat\u00eb, 30% e mbetur ende punon n\u00eb bare metal pa hipervizor\u00eb. K\u00ebto 30% kryesisht p\u00ebrb\u00ebhen nga aplikacione me ngarkes\u00eb t\u00eb lart\u00eb q\u00eb lidhen, p\u00ebr shembull, me trajnimin e rrjeteve neurale dhe p\u00ebrdorin karta grafike.<\/p>\n

Ekspert\u00ebt e shpjegojn\u00eb k\u00ebt\u00eb trend duke th\u00ebn\u00eb se hipervizori si nj\u00eb shtres\u00eb nd\u00ebrmjet\u00ebsimi mund t\u00eb ndikoj\u00eb n\u00eb performanc\u00ebn e t\u00ebr\u00eb sistemit. N\u00eb studimet pes\u00ebvje\u00e7are q\u00eb kan\u00eb kaluar mund t\u00eb gjenden t\u00eb dh\u00ebna<\/a><\/noindex> p\u00ebr nj\u00eb ulje prej 10% n\u00eb shpejt\u00ebsin\u00eb e pun\u00ebs. Prandaj, kompanit\u00eb dhe operator\u00ebt e qendrave t\u00eb t\u00eb dh\u00ebnave nuk jan\u00eb t\u00eb nxituar t\u00eb transferojn\u00eb ngarkesat HPC n\u00eb mjedisin virtual.<\/p>\n

Por teknologjit\u00eb e virtualizimit po zhvillohen dhe p\u00ebrmir\u00ebsohen. N\u00eb konferenc\u00ebn e nj\u00eb muaji m\u00eb par\u00eb, VMware informoi se hipervizori ESXi nuk ka ndikim negativ n\u00eb performanc\u00ebn e GPU-ve. Shpejt\u00ebsia e llogaritjeve mund t\u00eb ulet me tre p\u00ebr qind, e cila \u00ebsht\u00eb e krahasueshme me parametrat e bare metal.<\/p>\n

Si funksionon<\/h2>\n

\nP\u00ebr t\u00eb rritur performanc\u00ebn e sistemeve HPC me karta grafike, VMware kan\u00eb b\u00ebr\u00eb disa ndryshime n\u00eb funksionimin e hipervizorit. Ve\u00e7an\u00ebrisht, e kan\u00eb eliminuar funksionin e vMotion. Ky funksion nevojitet p\u00ebr balancimin e ngarkes\u00ebs dhe zakonisht transferon makinat virtuale (VM) midis server\u00ebve ose GPU-ve. Shk\u00ebputja e vMotion ka b\u00ebr\u00eb q\u00eb \u00e7do VM tani t\u00eb lidhet me nj\u00eb GPU specifike. Kjo ka ndihmuar n\u00eb uljen e kostove gjat\u00eb shk\u00ebmbimit t\u00eb t\u00eb dh\u00ebnave.<\/p>\n

Nj\u00eb komponent tjet\u00ebr ky\u00e7 i sistemit \u00ebsht\u00eb teknologjia<\/a><\/noindex> DirectPath I\/O. Ajo lejon q\u00eb drejtuesi CUDA p\u00ebr llogaritjet paralel t\u00eb nd\u00ebrveproj\u00eb me makinat virtuale n\u00eb m\u00ebnyr\u00eb direkte, duke anashkaluar hipervizorin. Kur n\u00eb nj\u00eb GPU nevojitet t\u00eb ekzekutohen disa VM nj\u00ebkoh\u00ebsisht, zbatohet zgjidhja GRID vGPU. Ajo ndan memorien e kart\u00ebs n\u00eb disa segmente (por ciklet llogarit\u00ebse nuk ndahen).<\/p>\n

Skema e funksionimit t\u00eb dy makinave virtuale n\u00eb k\u00ebt\u00eb rast do t\u00eb duket si m\u00eb posht\u00eb:<\/p>\n

\"Situata:<\/p>\n

Rezultatet dhe parashikimet<\/h2>\n

\nKompania u kryen teste<\/a><\/noindex> t\u00eb hipervizorit, duke trajnuar nj\u00eb model gjuh\u00ebsor t\u00eb bazuar TensorFlow<\/a><\/noindex>. \"D\u00ebmi\" i performanc\u00ebs ishte vet\u00ebm 3\u20134%, krahasuar me bare metal. Megjithat\u00eb, sistemi fitoi mund\u00ebsin\u00eb p\u00ebr t\u00eb shp\u00ebrndar\u00eb burimet sipas k\u00ebrkes\u00ebs n\u00eb var\u00ebsi t\u00eb ngarkesave aktuale.<\/p>\n

Giganti IT gjithashtu kryente teste<\/a><\/noindex> me kontejner\u00ebt. Inxhinier\u00ebt e kompanis\u00eb trajnuan rrjetet neurale p\u00ebr t\u00eb njohur imazhet. Gjat\u00eb k\u00ebsaj, burimet e nj\u00eb GPU-je u shp\u00ebrndan\u00eb midis kat\u00ebr VM-ve kontejner. Si rezultat, performanca e makinave individuale u ul me 17% (krahasuar me nj\u00eb VM q\u00eb kishte qasje t\u00eb plot\u00eb n\u00eb burimet e GPU). Megjithat\u00eb, numri i imazheve t\u00eb p\u00ebrpunuara p\u00ebr sekond\u00eb u rrit<\/a><\/noindex> u rrit tre her\u00eb. Pritet q\u00eb sisteme t\u00eb tilla t\u00eb gjejn\u00eb<\/a><\/noindex> p\u00ebrdorim n\u00eb fush\u00ebn e analiz\u00ebs s\u00eb t\u00eb dh\u00ebnave dhe modelimit kompjuterik.<\/p>\n

Midis problemeve t\u00eb mundshme me t\u00eb cilat mund t\u00eb p\u00ebrballet VMware, ekspert\u00ebt p\u00ebrmendin<\/a><\/noindex> kan\u00eb nj\u00eb audienc\u00eb shum\u00eb t\u00eb ngusht\u00eb. Me sistemet e llogaritjes s\u00eb performanc\u00ebs t\u00eb lart\u00eb, aktualisht punojn\u00eb nj\u00eb num\u00ebr i vog\u00ebl kompanish. Megjith\u00ebse n\u00eb Statista theksojn\u00eb<\/a><\/noindex>, q\u00eb deri n\u00eb vitin 2021 do t\u00eb ken\u00eb virtualizuar 94% t\u00eb ngarkesave t\u00eb pun\u00ebs globale. Sipas<\/a><\/noindex> parashikimeve<\/p>\n

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

Zgjidhje t\u00eb ngjashme<\/h2>\n

\nanalist\u00ebve, vlera e tregut HPC do t\u00eb rritet nga 32 n\u00eb 45 miliard dollar\u00eb n\u00eb period\u00ebn 2017 deri n\u00eb 2022.<\/p>\n

N\u00eb treg ka disa analoge q\u00eb po zhvillohen nga kompani t\u00eb m\u00ebdha IT: AMD dhe Intel. ofron<\/noindex> Kompania e par\u00eb p\u00ebr virtualizim t\u00eb GPU-ve<\/p>\n

ndjek nj\u00eb qasje bazuar n\u00eb SR-IOV (virtualizimi i hyrjes\/daljes me rr\u00ebnj\u00eb t\u00eb vetme). Kjo teknologji ofron VM-ve qasje n\u00eb nj\u00eb pjes\u00eb t\u00eb kapaciteteve harduerike t\u00eb sistemit. Zgjidhja lejon ndarjen e GPU-s\u00eb midis 16 p\u00ebrdoruesve me performanc\u00eb t\u00eb barabart\u00eb t\u00eb sistemeve t\u00eb virtualizuara. Sa i p\u00ebrket gjigantit t\u00eb dyt\u00eb IT, teknologjia e tyre<\/a><\/noindex> bazohet n\u00eb hipervizorin Citrix XenServer 7. Ajo bashkon pun\u00ebn e drejtuesit standard t\u00eb GPU-ve dhe makin\u00ebs virtuale, duke lejuar k\u00ebshtu q\u00eb ajo t\u00eb shfaq\u00eb aplikacione 3D dhe desktop-e n\u00eb pajisjet e qindra p\u00ebrdoruesve.<\/p>\n

E ardhmja e teknologjis\u00eb<\/h2>\n

\nZhvilluesit e GPU-ve virtuale po vendosin<\/a><\/noindex> n\u00eb sistemet e AI q\u00eb po b\u00ebhen gjithnj\u00eb e m\u00eb t\u00eb njohura n\u00eb tregun e teknologjive t\u00eb biznesit. Ata shpresojn\u00eb se nevoja p\u00ebr p\u00ebrpunimin e volumit t\u00eb madh t\u00eb t\u00eb dh\u00ebnave do t\u00eb rris\u00eb k\u00ebrkes\u00ebn p\u00ebr vGPU.<\/p>\n

Aktualisht prodhuesit po k\u00ebrkojn\u00eb nj\u00eb m\u00ebnyr\u00eb<\/a><\/noindex> t\u00eb bashkohen funksionalitetet e CPU dhe GPU n\u00eb nj\u00eb b\u00ebrtham\u00eb, p\u00ebr t\u00eb p\u00ebrshpejtuar zgjidhjen e problemeve q\u00eb lidhen me grafik\u00ebn, kryerjen e llogaritjeve matematikore, operacione logjike, p\u00ebrpunimin e t\u00eb dh\u00ebnave. Shfaqja n\u00eb treg e till\u00eb e b\u00ebrthamave n\u00eb t\u00eb ardhmen do t\u00eb ndryshoj\u00eb qasjen ndaj virtualizimit t\u00eb burimeve dhe shp\u00ebrndarjes s\u00eb tyre mes ngarkesave t\u00eb pun\u00ebs n\u00eb mjedisin virtual dhe n\u00eb re.<\/p>\n

\n\u00c7far\u00eb t\u00eb lexoni mbi k\u00ebt\u00eb tem\u00eb n\u00eb blogun ton\u00eb korporativ:<\/p>\n