Hello Habr! Saitin bayanai don Babban Bayanai da koyon injin suna girma sosai kuma muna buƙatar ci gaba da su. Matsayinmu game da wata sabuwar fasaha a fagen ƙididdiga masu girma (HPC, Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwaƙwalwa ) ) wanda aka nuna a rumfar Kingston a . Wannan shine amfani da tsarin adana bayanai na Hi-End (SDS) a cikin sabobin tare da raka'o'in sarrafa hoto (GPU) da fasahar bas ɗin Adanawa na GPUDirect. Godiya ga musayar bayanai kai tsaye tsakanin tsarin ajiya da GPU, ketare CPU, ɗora bayanai a cikin masu haɓaka GPU ana haɓaka ta hanyar tsari mai girma, don haka Big Data aikace-aikacen yana gudana a matsakaicin aikin da GPUs ke bayarwa. Hakanan, masu haɓaka tsarin HPC suna sha'awar ci gaba a cikin tsarin ajiya tare da mafi girman saurin I/O, kamar waɗanda Kingston ke samarwa.
Ayyukan GPU ya zarce lodin bayanai
Tun lokacin da CUDA, kayan aikin GPU na tushen GPU da keɓaɓɓen kayan aikin kwamfuta na software don haɓaka aikace-aikacen maƙasudin gabaɗaya, an ƙirƙira su a cikin 2007, ƙarfin kayan aikin GPUs da kansu sun girma sosai. A yau, ana ƙara amfani da GPUs a aikace-aikacen HPC kamar Big Data, koyon injin (ML), da zurfin koyo (DL).
Lura cewa duk da kamanceceniya na sharuddan, na ƙarshe biyu ayyuka ne daban-daban na algorithmically. ML yana horar da kwamfutar bisa ga bayanan da aka tsara, yayin da DL ke horar da kwamfutar bisa ga martani daga hanyar sadarwa na jijiyoyi. Misali don taimakawa fahimtar bambance-bambancen abu ne mai sauƙi. Bari mu ɗauka cewa dole ne kwamfutar ta bambanta tsakanin hotunan kuliyoyi da karnuka waɗanda aka loda daga tsarin ajiya. Don ML, ya kamata ku ƙaddamar da saitin hotuna tare da alamomi masu yawa, kowannensu yana bayyana takamaiman fasalin dabbar. Don DL, ya isa a loda adadi mafi girma na hotuna, amma tare da tag ɗaya kawai "wannan cat" ko "wannan kare ne". DL yayi kama da yadda ake koyar da yara ƙanana - ana nuna su kawai hotuna na karnuka da kuliyoyi a cikin littattafai da kuma rayuwa (mafi sau da yawa, ba tare da bayyana cikakken bambanci ba), kuma kwakwalwar yaron kanta ta fara ƙayyade nau'in dabba bayan. wani adadi mai mahimmanci na hotuna don kwatanta (Bisa ga ƙididdiga, muna magana ne game da nunin ɗari ko biyu kawai a duk lokacin ƙuruciya). Algorithms na DL har yanzu ba su zama cikakke ba: don hanyar sadarwar jijiyoyi suma suyi nasarar aiki akan gano hotuna, ya zama dole don ciyarwa da sarrafa miliyoyin hotuna cikin GPU.
Takaitacciyar gabatarwar: dangane da GPUs, zaku iya gina aikace-aikacen HPC a fagen Big Data, ML da DL, amma akwai matsala - saitin bayanan suna da girma har lokacin da ake ɗaukar bayanai daga tsarin ajiya zuwa GPU. ya fara rage yawan aikin aikace-aikacen. A takaice dai, GPUs masu sauri suna kasancewa marasa amfani saboda jinkirin bayanan I/O da ke fitowa daga wasu tsarin ƙasa. Bambanci a cikin saurin I/O na GPU da bas zuwa tsarin CPU/ajiya na iya zama tsari na girma.
Ta yaya fasahar Ma'ajiya ta GPUDirect ke aiki?
Tsarin I/O na CPU ne ke sarrafa shi, kamar yadda ake loda bayanai daga ajiya zuwa GPUs don ƙarin aiki. Wannan ya haifar da buƙatar fasahar da za ta ba da damar kai tsaye tsakanin GPUs da kuma NVMe don sadarwa da juna cikin sauri. NVIDIA ita ce ta farko da ta ba da irin wannan fasaha kuma ta kira shi GPUDirect Storage. A haƙiƙa, wannan bambancin fasaha ce ta GPUDirect RDMA (Adreshin Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwaƙwalwa ) da suka ƙirƙira a baya.
Jensen Huang, Shugaba na NVIDIA, zai gabatar da GPUDirect Storage a matsayin bambancin GPUDirect RDMA a SC-19. Source: NVIDIA
Bambanci tsakanin GPUDirect RDMA da GPUDirect Storage yana cikin na'urorin da ake aiwatar da jawabi. GPUDirect RDMA an sake amfani da fasahar don matsar da bayanai kai tsaye tsakanin katin dubawar cibiyar sadarwa na gaba (NIC) da ƙwaƙwalwar GPU, kuma GPUDirect Storage yana ba da hanyar bayanai kai tsaye tsakanin ma'ajiyar gida ko na nesa kamar NVMe ko NVMe akan Fabric (NVMe-oF) da kuma GPU ƙwaƙwalwar ajiya.
Dukansu GPUDirect RDMA da GPUDirect Storage suna guje wa motsin bayanan da ba dole ba ta hanyar buffer a cikin ƙwaƙwalwar ajiyar CPU kuma suna ba da damar hanyar ƙwaƙwalwar kai tsaye (DMA) don matsar da bayanai daga katin cibiyar sadarwa ko ajiya kai tsaye zuwa ko daga ƙwaƙwalwar GPU - duk ba tare da lodi akan CPU ta tsakiya ba. Don Ma'ajiyar GPUDirect, wurin ajiyar ba shi da mahimmanci: yana iya zama faifan NVME a cikin naúrar GPU, a cikin rack, ko haɗa ta hanyar sadarwa azaman NVMe-oF.
Tsarin aiki na GPUDirect Storage. Source: NVIDIA
Tsarin ajiya na Hi-End akan NVMe ana buƙata a cikin kasuwar aikace-aikacen HPC
Sanin cewa tare da zuwan GPUDirect Storage, za a jawo sha'awar manyan abokan ciniki don ba da tsarin ajiya tare da saurin I / O daidai da kayan aikin GPU, a nunin SC-19 Kingston ya nuna demo na tsarin da ke kunshe da tsarin ajiya wanda ya danganci fayafai NVMe da naúrar mai GPU, wanda ya bincika dubban hotunan tauraron dan adam a sakan daya. Mun riga mun rubuta game da irin wannan tsarin ajiya dangane da 10 DC1000M U.2 NVMe drives .
Tsarin ajiya wanda ya dogara da 10 DC1000M U.2 NVMe yana tafiyar da daidaitaccen sabar tare da masu haɓaka hoto. Source: Kingston
An tsara wannan tsarin ajiya azaman 1U ko naúrar rack mafi girma kuma ana iya ƙididdige shi dangane da adadin abubuwan tafiyar DC1000M U.2 NVMe, kowanne yana da ƙarfin 3.84-7.68 TB. DC1000M shine samfurin NVMe SSD na farko a cikin nau'in nau'in U.2 a layin Kingston na abubuwan tafiyar da bayanai. Yana da ƙimar juriya (DWPD, Drive ya rubuta kowace rana), yana ba shi damar sake rubuta bayanai zuwa cikakkiyar ƙarfinsa sau ɗaya a rana don tabbacin rayuwar tuƙi.
A cikin gwajin fio v3.13 akan tsarin aiki na Ubuntu 18.04.3 LTS, Linux kernel 5.0.0-31-generic, samfurin ajiya na nunin ya nuna saurin karantawa (An ci gaba da karantawa) na IOPS miliyan 5.8 tare da ci gaba mai dorewa (Bandiwidth mai dorewa). 23.8 Gbit/s.
Ariel Perez, manajan kasuwanci na SSD a Kingston, ya ce game da sabbin tsarin ajiya: “Muna shirye don samar da sabar sabar na gaba tare da mafita na U.2 NVMe SSD don kawar da yawancin matsalolin canja wurin bayanai waɗanda aka saba da alaƙa da ajiya. Haɗin kayan aikin NVMe SSD da babban sabar Premier DRAM ɗin mu ya sa Kingston ya zama mafi cikakkiyar masana'antar samar da mafita na ƙarshen-zuwa-ƙarshen.
Gwajin gfio v3.13 ya nuna kayan aiki na 23.8 Gbps don tsarin ajiya na demo akan abubuwan tafiyar DC1000M U.2 NVMe. Source: Kingston
Menene tsarin al'ada na aikace-aikacen HPC zai yi kama da amfani da GPUDirect Storage ko fasaha makamancin haka? Wannan gine-gine ne tare da rabuwa ta jiki na raka'a masu aiki a cikin rak: raka'a ɗaya ko biyu don RAM, da yawa don GPU da nodes ɗin kwamfuta na CPU, da ɗaya ko fiye na tsarin ajiya.
Tare da sanarwar Ma'ajiyar GPUDirect da yuwuwar fitowar irin wannan fasahar daga sauran dillalai na GPU, buƙatun Kingston na tsarin ajiya da aka ƙera don amfani da ƙididdiga masu girma na haɓaka. Alamar alama za ta zama saurin karanta bayanai daga tsarin ajiya, kwatankwacin yadda ake fitar da katunan sadarwar 40- ko 100-Gbit a ƙofar sashin kwamfuta tare da GPU. Don haka, tsarin ajiya mai saurin-sauri, gami da NVMe na waje ta hanyar Fabric, za su tafi daga kasancewa mai ban mamaki zuwa ga al'ada don aikace-aikacen HPC. Baya ga lissafin kimiyya da na kuɗi, za su sami aikace-aikace a wasu fagage masu amfani da yawa, kamar tsarin tsaro a matakin babban birni na Safe City ko cibiyoyin sa ido kan sufuri, inda ake buƙatar fitarwa da gano saurin miliyoyin hotuna HD a sakan daya, ”in ji shi. niche kasuwa na babban tsarin Adanawa
Ana iya samun ƙarin bayani game da samfuran Kingston a kamfani.
source: www.habr.com