{"id":157425,"date":"2026-02-07T17:12:29","date_gmt":"2026-02-07T15:12:29","guid":{"rendered":"https:\/\/prohoster.info\/blog\/novosti-interneta\/inicziativa-po-vstraivaniyu-modelej-mashinnogo-obucheniya-v-yadro-linux"},"modified":"2026-02-07T17:12:29","modified_gmt":"2026-02-07T15:12:29","slug":"inicziativa-po-vstraivaniyu-modelej-mashinnogo-obucheniya-v-yadro-linux","status":"publish","type":"post","link":"https:\/\/prohoster.info\/sk\/blog\/news\/inicziativa-po-vstraivaniyu-modelej-mashinnogo-obucheniya-v-yadro-linux","title":{"rendered":"Iniciat\u00edva na integr\u00e1ciu modelov strojov\u00e9ho u\u010denia do jadra Linuxu","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"
Vja\u010deslav Dubejko z firmy IBM spustil na zozname distrib\u00facie pre v\u00fdvoj\u00e1rov jadra Linux diskusiu o vyu\u017eit\u00ed modelov strojov\u00e9ho u\u010denia v jadre a navrhol na testovanie s\u00fabor z\u00e1plat s kni\u017enicou na integr\u00e1ciu ML modelov do jadra a pr\u00edkladom symbolick\u00e9ho ovl\u00e1da\u010da, ktor\u00fd t\u00fato kni\u017enicu pou\u017e\u00edva. <\/p>\n
Integr\u00e1cia ML modelov do jadra m\u00f4\u017ee by\u0165 u\u017eito\u010dn\u00e1 pre zmenu logiky fungovania subsyst\u00e9mov na z\u00e1klade spracovan\u00fdch d\u00e1t, optimaliz\u00e1ciu pr\u00e1ce a prisp\u00f4sobenie konfigur\u00e1cie v z\u00e1vislosti od vn\u00fatorn\u00e9ho stavu syst\u00e9mov. Pou\u017eitie strojov\u00e9ho u\u010denia, ktor\u00e9 dok\u00e1\u017ee odha\u013eova\u0165 vzory a robi\u0165 predpovede bez manu\u00e1lnych implement\u00e1ci\u00ed algoritmov, zjednodu\u0161\u00ed v\u00fdber najefekt\u00edvnej\u0161ej konfigur\u00e1cie jadra s oh\u013eadom na zlo\u017eitosti a variabilitu modern\u00fdch pracovn\u00fdch z\u00e1\u0165a\u017e\u00ed a umo\u017en\u00ed rie\u0161i\u0165 tak\u00e9 probl\u00e9my, ako je predpovedanie por\u00fach syst\u00e9mov skladovania. <\/p>\n
Okrem enginu na vykon\u00e1vanie modelov sa zva\u017euje v\u00fdvoj n\u00e1strojov na zber d\u00e1t pre tr\u00e9ning ML modelov, samotn\u00fd tr\u00e9ning modelu a testovanie v\u00fdsledkov. Ke\u010f\u017ee na vykon\u00e1vanie ML modelu s\u00fa potrebn\u00e9 oper\u00e1cie s pl\u00e1vaj\u00facou desatinnou \u010diarou a priamo pou\u017eitie FPU v jadre nie je povolen\u00e9, navrhnut\u00fd prototyp predstavuje vrstvenie pre pr\u00edstup z r\u00f4znych subsyst\u00e9mov jadra k ML modelom, ktor\u00e9 sa vykon\u00e1vaj\u00fa v pou\u017e\u00edvate\u013eskom priestore, podobne ako vy\u010dlenenie spracovate\u013eov SPDK, DPDK a ublk do pou\u017e\u00edvate\u013esk\u00e9ho priestoru. <\/p>\n
Vy\u010dlenenie vykon\u00e1vania a tr\u00e9ningu modelu do pou\u017e\u00edvate\u013esk\u00e9ho priestoru zjednodu\u0161uje \u00fadr\u017ebu a izoluje jadro od probl\u00e9mov v k\u00f3de vykon\u00e1vania modelu. V \u0161t\u00e1diu tr\u00e9ningu m\u00f4\u017eu by\u0165 d\u00e1ta o stave a parametroch jadra bu\u010f po\u017eadovan\u00e9 spracovate\u013eom z pou\u017e\u00edvate\u013esk\u00e9ho priestoru, alebo pren\u00e1\u0161an\u00e9 vrstvou, ktor\u00e1 be\u017e\u00ed na \u00farovni jadra. Na riadenie interakcie komponentov v jadre a pou\u017e\u00edvate\u013eskom priestore sa pou\u017e\u00edva sysfs. Je mo\u017en\u00e9 adapt\u00edvne u\u010denie ML modelu, pri ktorom subsyst\u00e9m jadra dost\u00e1va odpor\u00fa\u010dania od ML modelu, aplikuje odpor\u00fa\u010dan\u00e9 zmeny a hodnot\u00ed \u00fa\u010dinnos\u0165 odpor\u00fa\u010dan\u00ed na zmenu stavu. \n Zdroj: opennet.ru<\/a> <\/p>","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"excerpt":{"rendered":"