5.8 lab IOPS: vim li cas thiaj ntau?

Nyob zoo Habr! Cov ntaub ntawv teev tseg rau Cov Ntaub Ntawv Loj thiab kev kawm tshuab tau loj hlob zuj zus thiab peb yuav tsum ua raws li lawv. Peb cov ntawv tshaj tawm txog lwm cov thev naus laus zis tshiab hauv kev ua haujlwm siab (HPC, High Performance Computing), qhia ntawm Kingston lub rooj muag khoom ntawm Supercomputing - 2019. Qhov no yog kev siv Hi-End cov ntaub ntawv khaws cia (SDS) hauv cov servers nrog graphic processing units (GPU) thiab GPUDirect Storage bus technology. Ua tsaug rau kev sib pauv cov ntaub ntawv ncaj qha ntawm lub kaw lus cia thiab GPU, dhau ntawm CPU, thauj cov ntaub ntawv mus rau GPU accelerators yog nrawm los ntawm qhov kev txiav txim ntawm qhov loj, yog li Cov Ntaub Ntawv Loj tau khiav ntawm qhov ua tau zoo tshaj plaws uas GPUs muab. Nyob rau hauv tas li ntawd, HPC system developers txaus siab rau kev nce qib nyob rau hauv cia systems nrog lub siab tshaj plaws I / O speeds, xws li cov tsim los ntawm Kingston.

5.8 lab IOPS: vim li cas thiaj ntau?

GPU kev ua tau zoo tshaj cov ntaub ntawv thauj khoom

Txij li thaum CUDA, GPU-based hardware thiab software parallel computing architecture rau kev tsim cov phiaj xwm dav dav, tau tsim nyob rau xyoo 2007, lub peev xwm kho vajtse ntawm GPUs lawv tus kheej tau loj hlob zoo kawg. Niaj hnub no, GPUs tau siv ntau dua hauv HPC daim ntawv thov xws li Cov Ntaub Ntawv Loj, kev kawm tshuab (ML), thiab kev kawm tob (DL).

Nco ntsoov tias txawm tias qhov zoo sib xws ntawm cov ntsiab lus, ob qhov kawg yog cov haujlwm sib txawv algorithmically. ML cob qhia lub khoos phis tawj raws li cov ntaub ntawv tsim, thaum DL cob qhia lub khoos phis tawj raws li kev tawm tswv yim los ntawm neural network. Ib qho piv txwv los pab kom nkag siab qhov sib txawv yog qhov yooj yim heev. Cia peb xav tias lub khoos phis tawj yuav tsum paub qhov txawv ntawm cov duab ntawm miv thiab dev uas tau thauj khoom los ntawm lub kaw lus cia. Rau ML, koj yuav tsum xa cov duab nrog ntau cov cim npe, txhua tus uas txhais tau tias ib qho tshwj xeeb ntawm tus tsiaj. Rau DL, nws yog qhov txaus los xa cov duab ntau dua, tab sis tsuas yog ib daim ntawv "qhov no yog miv" lossis "qhov no yog dev". DL zoo ib yam li cov menyuam yaus raug qhia - lawv tsuas yog pom cov duab ntawm dev thiab miv hauv cov phau ntawv thiab hauv lub neej (feem ntau, tsis tau piav qhia txog qhov sib txawv), thiab tus menyuam lub hlwb nws tus kheej pib txiav txim siab hom tsiaj tom qab. ib qho tseem ceeb ntawm cov duab rau kev sib piv (raws li kev kwv yees, peb tab tom tham txog tsuas yog ib puas lossis ob qho yeeb yam thoob plaws thaum yau). DL algorithms tseem tsis tau zoo tag nrho: rau lub network neural tseem ua tiav kev txheeb xyuas cov duab, nws yog qhov yuav tsum tau noj thiab ua cov duab ntau lab rau hauv GPU.

Cov ntsiab lus ntawm cov lus ua ntej: raws li GPUs, koj tuaj yeem tsim HPC daim ntawv thov hauv thaj tsam ntawm Cov Ntaub Ntawv Loj, ML thiab DL, tab sis muaj ib qho teeb meem - cov ntaub ntawv teev muaj ntau heev uas lub sij hawm siv cov ntaub ntawv los ntawm lub kaw lus cia rau GPU pib txo qhov kev ua tau zoo tag nrho ntawm daim ntawv thov. Hauv lwm lo lus, ceev GPUs tseem tsis tau siv vim qeeb I / O cov ntaub ntawv los ntawm lwm cov subsystems. Qhov sib txawv ntawm I / O ceev ntawm GPU thiab lub tsheb npav mus rau CPU / lub kaw lus cia tuaj yeem yog qhov kev txiav txim loj.

GPUDirect Storage technology ua haujlwm li cas?

Cov txheej txheem I/O yog tswj los ntawm CPU, ib yam li cov txheej txheem ntawm kev thauj cov ntaub ntawv los ntawm kev cia rau GPUs rau kev ua haujlwm ntxiv. Qhov no ua rau muaj kev thov rau thev naus laus zis uas yuav muab kev nkag ncaj qha ntawm GPUs thiab NVMe tsav kom sib txuas lus sai sai. NVIDIA yog thawj zaug muab cov thev naus laus zis no thiab hu ua GPUDirect Storage. Qhov tseeb, qhov no yog qhov hloov pauv ntawm GPUDirect RDMA (Remote Direct Memory Address) thev naus laus zis uas lawv tau tsim yav dhau los.

5.8 lab IOPS: vim li cas thiaj ntau?
Jensen Huang, CEO ntawm NVIDIA, yuav nthuav qhia GPUDirect Storage raws li qhov sib txawv ntawm GPUDirect RDMA ntawm SC-19. Source: NVIDIA

Qhov sib txawv ntawm GPUDirect RDMA thiab GPUDirect Storage yog nyob rau hauv cov khoom siv ntawm qhov chaw nyob yog nqa tawm. GPUDirect RDMA thev naus laus zis yog repurposed kom txav cov ntaub ntawv ncaj qha ntawm lub hauv ntej-kawg network interface card (NIC) thiab GPU nco, thiab GPUDirect Storage muab cov ntaub ntawv ncaj qha ntawm lub zos lossis chaw cia khoom xws li NVMe lossis NVMe hla Fabric (NVMe-oF) thiab GPU nco.

Ob leeg GPUDirect RDMA thiab GPUDirect Storage zam tsis tas yuav tsum muaj cov ntaub ntawv txav los ntawm qhov tsis nyob hauv CPU nco thiab tso cai rau lub cim xeeb ncaj qha nkag (DMA) mechanism txav cov ntaub ntawv los ntawm daim npav network lossis cia ncaj qha mus rau lossis los ntawm GPU nco - txhua yam tsis muaj kev thauj khoom ntawm lub hauv paus CPU. Rau GPUDirect Cia, qhov chaw ntawm qhov chaw cia tsis muaj teeb meem: nws tuaj yeem yog NVME disk hauv GPU chav tsev, hauv khib, lossis txuas nrog lub network li NVMe-oF.

5.8 lab IOPS: vim li cas thiaj ntau?
Lub tswv yim ntawm kev ua haujlwm ntawm GPUDirect Storage. Source: NVIDIA

Hi-End cia systems ntawm NVMe yog qhov xav tau hauv HPC daim ntawv thov kev lag luam

Paub txog tias nrog kev tshwm sim ntawm GPUDirect Storage, qhov kev txaus siab ntawm cov neeg siv khoom loj yuav raug kos los muab cov txheej txheem khaws cia nrog I / O nrawm sib xws rau kev dhau los ntawm GPU, ntawm SC-19 exhibition Kingston tau qhia ib qho kev nthuav qhia ntawm lub kaw lus suav nrog qhov system cia raws li NVMe disks thiab ib chav tsev nrog GPU, uas txheeb xyuas ntau txhiab tus duab satellite ib ob. Peb twb tau sau txog xws li ib tug cia system raws li 10 DC1000M U.2 NVMe drives nyob rau hauv ib daim ntawv qhia los ntawm lub supercomputer exhibition.

5.8 lab IOPS: vim li cas thiaj ntau?
Lub kaw lus cia raws li 10 DC1000M U.2 NVMe tsav kom ua tiav cov neeg rau zaub mov nrog cov duab nrawm. Source: Kingston

Lub kaw lus cia no yog tsim los ua 1U lossis loj rack chav tsev thiab tuaj yeem ntsuas raws li tus naj npawb ntawm DC1000M U.2 NVMe drives, txhua qhov muaj peev xwm ntawm 3.84-7.68 TB. DC1000M yog thawj tus qauv NVMe SSD hauv U.2 daim ntawv foos hauv Kingston txoj kab ntawm cov ntaub ntawv chaw tsav tsheb. Nws muaj qhov ntsuas kev ua siab ntev (DWPD, Tsav sau ib hnub), tso cai rau nws rov sau cov ntaub ntawv rau nws lub peev xwm tag nrho ib hnub rau lub neej lav ntawm lub tsav.

Hauv kev sim fio v3.13 ntawm Ubuntu 18.04.3 LTS kev khiav hauj lwm qhov system, Linux ntsiav 5.0.0-31-generic, cov qauv cia khoom nthuav qhia pom qhov kev nyeem nrawm (Sustained Read) ntawm 5.8 lab IOPS nrog kev txhawb nqa mus ntev (Sustained Bandwidth ) ntawm 23.8 Gbit / s.

Ariel Perez, SSD tus thawj tswj hwm kev lag luam ntawm Kingston, tau hais txog cov txheej txheem khaws cia tshiab: "Peb tau npaj los npaj cov neeg siv khoom txuas ntxiv nrog U.2 NVMe SSD cov kev daws teeb meem los tshem tawm ntau cov ntaub ntawv hloov pauv cov fwj uas ib txwm muaj cuam tshuam nrog kev khaws cia. Kev sib xyaw ua ke ntawm NVMe SSD drives thiab peb cov hwm Server Premier DRAM ua rau Kingston yog ib qho ntawm kev lag luam cov ntaub ntawv tshaj tawm txog kev daws teeb meem. "

5.8 lab IOPS: vim li cas thiaj ntau?
Qhov kev sim gfio v3.13 tau pom qhov kev sib tw ntawm 23.8 Gbps rau qhov demo cia qhov system ntawm DC1000M U.2 NVMe drives. Source: Kingston

Yuav ua li cas yuav ib tug raug system rau HPC daim ntaub ntawv zoo li siv GPUDirect Storage los yog zoo sib xws technology? Qhov no yog ib qho qauv tsim nrog kev sib cais ntawm lub cev ntawm cov haujlwm ua haujlwm nyob rau hauv ib lub khib: ib lossis ob chav rau RAM, ob peb ntxiv rau GPU thiab CPU suav cov nodes, thiab ib lossis ntau chav rau cov tshuab cia.

Nrog rau kev tshaj tawm ntawm GPUDirect Storage thiab qhov muaj peev xwm tshwm sim ntawm cov thev naus laus zis zoo sib xws los ntawm lwm tus neeg muag khoom GPU, Kingston qhov kev thov rau cov tshuab khaws cia tsim los siv rau hauv kev ua haujlwm siab tau nthuav dav. Tus cim yuav yog qhov ceev ntawm kev nyeem cov ntaub ntawv los ntawm lub kaw lus cia, piv rau qhov kev xa tawm ntawm 40- lossis 100-Gbit network phaib ntawm qhov nkag mus rau chav suav nrog GPU. Yog li, ultra-high-speed storage systems, nrog rau sab nraud NVMe ntawm Fabric, yuav mus los ntawm kev txawv txav mus rau lub ntsiab rau HPC daim ntawv thov. Ntxiv nrog rau kev tshawb fawb thiab kev suav nyiaj txiag, lawv yuav pom daim ntawv thov nyob rau hauv ntau qhov chaw ua haujlwm, xws li kev ruaj ntseg ntawm lub nroog nyab xeeb theem lossis cov chaw saib xyuas kev thauj mus los, qhov twg kev lees paub thiab kev txheeb xyuas nrawm ntawm tsheej lab ntawm HD duab ib ob yog xav tau, "saib kev lag luam niche ntawm sab saum toj Storage system

Xav paub ntau ntxiv txog Kingston cov khoom tuaj yeem nrhiav tau ntawm official website tuam txhab.

Tau qhov twg los: www.hab.com

Ntxiv ib saib