Temesi Tibor: Az adatok optimális helye HPC környezetben [DKA metaadatok]

K é p a l á í r á s : Az adatok optimális helye HPC környezetben

N y e r s v a g y O C R - e s s z ö v e g : Az adatok optimális helye HPC környezetben Temesi Tibor tibor@silicon.hu Silicon Computers "The confluence of AI with traditional simulations is going to transform the very nature of high performance computing. That's the thing that I think is going be yet another sea change in how we do science." "We're not doing our grandfather's HPC here." Rick Stevens, Associate Lab Director, Argonne National Laboratory www.nextplatform.com/2019/10/22/exascale-is-not-your-grandfathers-hpc SIMULATION + AI = OVER- PROPORTIONAL STORAGE GROWTH HPC Storage spending is forecasted to grow with a 40% higher CAGR than HPC Servers through 2024. Projected Compound Annual Growth Rate for customer spending 2019-2024 Source: Hyperion Research, 2019 Market Results, New Forecasts and HPC Trends, April 2020 INPUT/OUTPUT PROFILES COULD NOT BE MORE DIFFERENT Era of convergence of traditional simulation and AI requires NEW HPC storage Traditional simulation Mainly WRITING PETAbytes of LARGE FILES in SEQUENTIAL order @ extreme speed INPUT DATA + ALGORITHM CPU nodes OUTPUT Machine Learning Mainly READING TERAbytes of FILES OF ALL SIZES in RANDOM order @ extreme speed INPUT DATA + HISTORICAL OUTPUT GPU nodes ALGORITHM START ANYWHERE, GO TO WHEREVER YOU Nearly 10 TB/sec and > 700 PB capacity* simulations for quantum computers, nuclear energy systems, fusion reactors, and precision medicines. 2021-2022 > 4 TB/sec and 30PB (All Flash) capacity* NESAP enhance simulation, data processing, and machine learning applications 2.5 TB/sec and ~ 400 PB capacity* nuclear stockpile, secondary national security missions, nuclear nonproliferation and counterterrorism 2023 2.5 TB/sec and ~ 100 PB capacity* LUMI - Large Unified Modern Infrastructure 2021 1 TB/sec and ~ 200 PB capacity* cancer research, materials science, climate science, and cosmology 2H 2021 1 TB/sec and ~ 1 PB (All Flash) capacity* IT4I - Astrophysics, Eng., Chemistry, Material Earth Live science 1H 2021 240 GB/sec and ~ 14.5 PB capacity* 2H 2020 All Flash entry point (6U) Up to 80/50 GB/sec read/write and 115 TB capacity* Disk entry point (10U): 15 GB/sec and 315 TB capacity* All Flash base rack: > 1TB/sec and up to 4.5 PB capacity* Expansion rack: > 2 TB/sec and 4.6 PB* Disk base rack: 90 GB/sec and 7.5 PB capacity* Expansion rack: 120 GB/sec and 10 PB* EXASCALE KÖVETELMÉNYEK Erősebb "motorok" Jobb kommunikáció: protokoll, hálózat, ... Alkalmazás-orientált architektúrák Az adatok feldolgozása A cél, hogy az adat minél közelebb legyen a processzorhoz, amikor éppen dolgozik vele. Az adatok olvasási/írási sebessége: nem optimális, ha motorok "gyorsan várnak" az adatra. Nyilván nem lehet minden adat a processzor chipek belső cache memóriájában, akkor hol legyen? "Ki" mozgassa az adatokat? MEMÓRIÁK Processzor belső cache Külső memóriák: DRAM (32 bit bus), HBM (4096 bit bus): sok adat (AI) -> szélesebb adatutak (HBM memóriák), kisebb késleltetések TÁROLÓK Közös, megosztott, többrétegű tárolórendszer Az éppen feldolgozandó adatok számára maximális írás/olvasási teljesítmény: SCRATCH: all- flash, NVMe,... Tárolási kapacitás az operatív adatok számára: HOME: HDD Archív réteg: szalagkönyvtár Automatizált adatmenedzsment NON-VOLATILE MEMORY EXPRESS (NVMe) A new storage protocol focused on SSDs, replacet traditional I/O stacks built on SCSI protocol, which is optimized for spinning media. It leverages PCIe instead of SAS/SATA for greater bandwidth, IOPS and reduced latency Designed to move beyond HDDs Standard interface for Solid State Media A new protocol Media? TÁROLÓK Közös, megosztott, többrétegű tárolórendszer Az éppen feldolgozandó adatok számára maximális írás/olvasási teljesítmény: SCRATCH: all- flash, NVMe,... Tárolási kapacitás az operatív adatok számára: HOME: HDD Archív réteg: szalagkönyvtár Automatizált adatmenedzsment THREE BASIC FILE SYSTEM CONFIGURATION OPTIONS All HDD file system /lustre Specifications per rack: Up to 120 GB/sec (read/write) Up to 10 PB usable capacity (16 TB HDD) Scalable Storage Units (SSU) SSU-D2 (10U) Up to 30 GB/sec from 212 HDD SSU-D4 (18U) Up to 40 GB/sec from 424 HDD Hybrid file system /lustre /lustre/disk /lustre/flash Disk and Flash pools in the same namespace with Cray ClusterStor data services providing capacity from Disk and performance from Flash All Flash file system /lustre Specifications per rack: Up to 1,600 GB/sec (read) Up to 1,000 GB/sec (write) Scalable Storage Unit (SSU) SSU-F (2U) Up to 80 GB/sec sequential read from 24 SSD Up to 50 GB/sec sequential write from 24 SSD BASIC LUSTRE Typical Lustre components CLUSTERSTOR E1000 LUSTRE FILE SYSTEM & DMF V7 SZALAGKÖNYVTÁR Az adatok megőrzése, archiválása, biztonsági másolata TFINITY EXASCALE - ENTERPRISE ARCHIVE TFinity Delivers Over 1 Exabyte native capacity with LTO-9 Up to 858 PB native capacity with TS1160 24 x 7 x 365 operations Dual robotics for availability and performance Flexible Configurations 3 to 45 frames 100 to over 56,400 LTO slots/42,930 TS slots 2 to 144 tape drives LTO-9, LTO-8, LTO-7, TS1160, TS1155, TS1150, T10K Technology Upgrade in 10 LTO or 9 TS slot increments 400 mounts per hour TERAPACK ARCHITECTURE Industry leading density Industry’s smallest footprint Reduce floor space requirements Reduce tape handling 10 LTO or 9 TS11xx tapes per TeraPack TeraPack design allows TFinity to maximize the floor space needed for the tape archive. TeraPacks make moving media much more efficient Both in the TFinity and when transporting or moving tapes outside the library STANDARD BLUESCALE SOFTWARE FEATURES Media Lifecycle Management (MLM) - Media Health reporting Drive Lifecycle Management (DLM) - Drive Health reporting and diagnostics Hardware Health Management (HHM) - Library Hardware reporting and diagnostics Standard Encryption - Built in Encryption Key Management Data Integrity Verification (DIV) - Verify data on tape automatically as a background task without host system involvement AutoSupport - The library will automatically create a support ticket if any critical event is encountered Partitioning - Support for up to 16 partitions Remote Library Connection (RLC) - Manage the library from anywhere through the same interface that is presented on the front panel touchscreen Auto Drive Clean - Let the library automatically clean tape drives when they request to be cleaned Compatible with all major ISV software Spectra Logic HPC Customer References Europe HPC/AI COMPUTE & STORAGE HPC Storage Supports an Optimized Compute Experience for Active Workloads INTRODUCING DATA MANAGEMENT FRAMEWORK VERSION 7 DMF Enables Data Curation, Placement and Protection Over Time & Distance DMF INTEGRATION WITH TAPE STORAGE Still the world's most cost effective Storage @ 1¢ per GB DMF is certified with libraries from HPE, as well as Spectra Logic, IBM and Oracle (StorageTek) Streams to tape drive at native rates, even for small files Block ID positioning for fast seek Support for latest LTO-9 and Enterprise-class drive technology Advanced feature support for accelerated retrieval and automated library management Supports Data Integrity Verification (DIV) and Logical Block Protection (LBP) available with Oracle T10k as well as IBM LTO and TS drives Recommended Access Order (RAO) and SpectraLogic's TAOS DATA MANAGEMENT FRAMEWORK VERSION 7 Resolve Data Challenges with Data Management Optimize Storage Utilization Hierarchical storage management: Integrated policy engine allows tiered data model with automated data movement to appropriate data tier, e.g. SSD/Flash tier for active data, tape and cloud for dormant/archive data Maximize resources with automatic archival of stale or cold data Data Protection Strategy Safeguard petabytes of data with integrated, continuous backups Maintain data integrity via self-describing media format and data checksums Streamline Workflows Accelerate workflows: Simplify management of large file collections with data set labeling Construct data sets using file metadata including extended attributes Automate data workflows with integrated job scheduling Streamline data recovery with automatic file versioning and Point-in-Time restore Reduce manual intervention: Reduce script writing and script maintenance Reduce operator errors with automatic data movement Manage Costs Control Storage Costs: - Minimize TCO by moving data to most cost-effective tier for desired access level, i.e. from flash to HDD to tape to cloud Scale storage acquisitions incrementally depending on need, i.e. more flash for active data, more tape for archive depending on demand Protect and future-proof data with automated migration to new storage infrastructure Köszönöm

D o k u m e n t u m n y e l v e : magyar