Click to edit title

Size: px

Start display at page:

Download "Click to edit title"

よしじろうそめや
5 years ago
Views:

1 インテル SSD データセンターファミリーの紹介と性能について力翠湖 (Midoriko Chikara) インテル株式会社技術本部プラットフォームアプリケーションエンジニア

2 Every day 1.5 GB Per day 1 3,000 GB Per day 2 4,000 GB Per day 3 40,000 GB Per day 3 1,000,000 GB Per day 3 Data is exploding

3 Explosion of Data: The need for real-time access Financial Services Fraud Detection Online Advertising Scientific Research Gaming Requires Memory Technology to Consistently Access More Information Faster 3

4 Data and CPU want to be closer together, but economics have kept them apart 4,000 GB Per day 3 3,000 GB Per day 2 Cold data warm data 40,000 GB Per day GB Per day 1 Hot data LOWER LESS DELAY MORE COST HIGHER 1,000,000 GB Per day 3 4

無断での引用転載を禁じます Intel インテル Intel ロゴはアメリカ合衆国および / またはその他の国における Intel Corporation の商標です * その他の社名製品名などは

5 Intel is investing in 2 technologies Intel 3D NAND technology Intel Optane technology lower cost & higher density Warm Data Higher Performance HOT DATA LOWER COST HIGHER LESS DELAY 2017 Intel Corporation. 無断での引用転載を禁じます Intel インテル Intel ロゴはアメリカ合衆国および / またはその他の国における Intel Corporation の商標です * その他の社名製品名などは一般に各社の表示商標または登録商標ですその他の社名製品名などは一般に各社の商標または登録商標ですインテル 5 ソフトウェア製品のパフォーマンス / 最適化に関する詳細は Optimization Notice ( 最適化に関する注意事項 ) を参照してください MORE

7 Leader in Flash Cell Technology NOR Flash Memory NAND Flash Memory 1.5µm 0.8µm 0.4µm 0.18µm 90nm 65nm 45nm 25nm 16nm 3D 1.0µm 0.6µm 0.25µm 0.13µm 50nm 34nm 20nm years of flash cell scaling experience 7

Replacement Gate Array Array Intel 3D NAND Technology

bigger capacities Smaller FG cell up to 20% higher

efficient array (CMOS under array) More GB per wafer

512 GB Intel 3D NAND to representative competitors

Conference papers citing Samsung Electronics and

8 Replacement Gate Array Array Intel 3D NAND Technology Optimized for Capacity and Cost Advantages Scale to bigger capacities Smaller FG cell up to 20% higher areal density 1 Manufacturing efficiency More efficient array (CMOS under array) More GB per wafer 1 Comparing areal density of Intel measured data on 512 GB Intel 3D NAND to representative competitors based on 2017 IEEE International Solid-State Circuits Conference papers citing Samsung Electronics and Western Digital/Toshiba die sizes for 64-stacked 3D NAND component.. 8

Proven Technology Accelerates Development 00 01 10 11 Accelerated Transitions 1 MLC to TLC TLC on 1 st Gen 3D NAND media 000

512 32-Layer to 64-Layer ~1 year Intel SSD 545s Series 1 Anand Tech 3D NAND Die Size Comparison http://www.anandtech.

Forward Insights QLC in the Datacenter, May 2017, http://www.forward-insights.com/reportslist.html; and http://www.anandtech.

9 Proven Technology Accelerates Development Accelerated Transitions 1 MLC to TLC TLC on 1 st Gen 3D NAND media First with commercially available 64-Layer TLC Gb and 512 Gb on 2 nd Gen 3D NAND media Layer to 64-Layer ~1 year Intel SSD 545s Series 1 Anand Tech 3D NAND Die Size Comparison Forward Insights QLC in the Datacenter, May 2017, and 2 Intel SSD 545s Series available on NewEgg June 27,

Solutions that Reach Across Multiple Segments Cloud and Data Center Internet of Things Devices PCIe* / NVMe* Intel SSD DC P4600 Series Intel SSD DC P4500 Series SATA Intel SSD DC S4600 Series Intel

10 Solutions that Reach Across Multiple Segments Cloud and Data Center Internet of Things Devices PCIe* / NVMe* Intel SSD DC P4600 Series Intel SSD DC P4500 Series SATA Intel SSD DC S4600 Series Intel SSD DC S3520 Series Embedded Intel SSD E 6000p Series Intel SSD E 5420s Series Client professional Intel SSD Pro 6000p Series Client consumer Intel SSD 600p Series Intel SSD DC P4600 Series Intel SSD DC S3520 Series Intel SSD E 6000p Series Intel SSD E 5420s Series Intel SSD Pro 6000p Series Intel SSD 600p Series *Other names and brands may be claimed as the property of others. 10

11 Intel Optane technology Unleashing Breakthrough Performance for a New Generation of Computing 11

12 3D XPoint MEMORY MEDIA In Pursuit of Large Memory Capacity / Word Access / Immediately Available Word (Cache Line) Crosspoint Structure Selectors allow dense packing and individual access to bits NVM Breakthrough Material Advances Compatible switch and memory cell materials Large Memory Capacity Crosspoint & Scalable Memory layers can be stacked in a 3D manner Immediately Available High Performance Cell and array architecture that can switch states 1000x faster than NAND 12

13 Intel Optane technology Big Memory and Fast Storage = Ideal Memory 3D XPoint Memory Media Intel Interconnect IP Intel Memory and Storage Controllers + = Intel Software Endurance latency Quality of service Throughput (IOPs) Technology claims are based on comparisons of latency, density and write cycling metrics amongst memory technologies recorded on published specifications of in-market memory products against internal Intel specifications. Intel Optane SSD prototype compared to the Intel SSD DC P3700 Series (NAND) 13

14 SSD Performance at Varying Queue Depths 4K Random Read 4K Random Write 4K Random 70/30 Mix Queue depth Queue depth Queue depth Intel Optane SSDs deliver high IOPs for a small # of Queue depth BUT this Measure ignores time per I/O Results measured by Intel based on the following configurations. Ubuntu LTS (GNU/Linux generic x86_64); Intel S2600WT motherboard with 2x Xeon 2.20GHz, 3.6GHz, 256GB RAM, fio , irqbalance off, smp affinity changed, cpu governor = performance; Prototype Intel Optane SSD: 187GB, FW: E , Intel DC P3700: 2 TB, FW: 8DV101F0 14

Predictably Fast Service up to 60x better at 99% QoS 1 Ideal for critical applications with aggressive latency requirements 1. Common Configuration Intel 2U Server System, OS CentOS 7.2, kernel 3.10.

15 Predictably Fast Service up to 60x better at 99% QoS 1 Ideal for critical applications with aggressive latency requirements 1. Common Configuration Intel 2U Server System, OS CentOS 7.2, kernel el7.x86_64, CPU 2 x Intel Xeon E GHz (22 cores), RAM 396GB 2133MHz. Configuration Intel Optane SSD DC P4800X 375GB and Intel SSD DC P GB. QoS measures 99% QoS under 4K workload at QD1 using fio Tests document performance of components on a particular test, in specific systems. Differences in hardware, software, or configuration will affect actual performance. 15

16 Responsive Under Load up to 40x faster response time under workload 1 Consistently amazing response time under load 1. Responsiveness defined as average read latency measured at queue depth 1 during 4k random write workload. Measured using FIO Common Configuration - Intel 2U Server System, OS CentOS 7.2, kernel el7.x86_64, CPU 2 x Intel Xeon E GHz (22 cores), RAM 396GB 2133MHz. Configuration Intel Optane SSD DC P4800X 375GB and Intel SSD DC P GB. Latency Average read latency measured at QD1 during 4K Random Write operations using fio Tests document performance of components on a particular test, in specific systems. Differences in hardware, software, or configuration will affect actual performance. 16

17 Queries per Second Intel Optane SSD Faster Transactions in MySQL* Sustained performance under load accelerates large batch Accelerating All-Flash Database Storage with Intel Optane SSD and Intel CAS Time to Complete 1 Million Transactions in MySQL With Intel Intel SATA SSD Optane RAID w/ SSD Intel and Optane Intel CAS SSD + Intel CAS Intel Without SATA Acceleration SSD RAID up to workloads x faster results Time (s) 1. System configuration - Intel Server System R2208WT2YS, 2x Intel Xeon E5 2699v4, 128 GB DDR4 DRAM, boot drive- 1x Intel SSD S3500 Series (240GB), database drives- 3x RAID 5 Intel SSD S3520 Series (1.6 TB) with 1x Intel RAID Controller RS3DC080, cache drive- Intel SSD P4800X Series (375 GB), CentOS 7.3, MySQL Server , Sysbench 0.5 configured for 70/30 R/W OLTP transaction split using a 1500GB database. Tests document performance of components on a particular test, in specific systems. Differences in hardware, software, or configuration will affect actual performance. *Other names and brands may be claimed as the property of others. 17

Endurance (DWPD) Ultra Endurance Architected for endurance scaling Write in place technology Non-destructive write process 30 10 Up to 2.8x more Total Bytes Written at similar capacity 1 0.

18 Endurance (DWPD) Ultra Endurance Architected for endurance scaling Write in place technology Non-destructive write process Up to 2.8x more Total Bytes Written at similar capacity MLC/TLC 2D/3D NAND SSD Intel Optane SSD 1. Comparing projected Intel Optane SSD 750GB specifications to actual Intel SSD DC P GB specifications. Total Bytes Written (TBW) calculated by multiplying specified or projected DWPD x specified or projected warranty duration x 365 days/year. Tests document performance of components on a particular test, in specific systems. Differences in hardware, software, or configuration will affect actual performance. 18

19 Extend Memory

Intel Optane SSD DC P4800X with Intel Memory Drive Technology Intel Memory Drive Technology Intel Optane SSD DC P4800X integrates transparently into memory subsystem 1 Middle layer SW boots prior to

20 Intel Optane SSD DC P4800X with Intel Memory Drive Technology Intel Memory Drive Technology Intel Optane SSD DC P4800X integrates transparently into memory subsystem 1 Middle layer SW boots prior to OS DRAM + Intel Optane SSD + Intel Memory Drive Technology emulate a single volatile memory pool No changes to OS and application required Supported on Intel Xeon processors 1. Maximum memory capacity varies by configuration. Maximum capacities measured binary are 320GiB for the 375GB drive, 640GiB for the 750GB drive and 1280GiB for the 1.5TB drive. 31

Range of Options to Extend Memory Intel Xeon E5 PCIe DDR PCIe* memory pool DRAM Intel Optane SSDs Intel 3D NAND SSDs Future: Intel DIMMs Based on 3D XPoint memory

21 Range of Options to Extend Memory Intel Xeon E5 PCIe DDR PCIe* memory pool DRAM Intel Optane SSDs Intel 3D NAND SSDs Future: Intel DIMMs Based on 3D XPoint memory media Today: Intel Optane SSDs with: OS Paging Optimize Applications Intel Memory Drive Technology *Other names and brands may be claimed as the property of others. 32

22 Intel Optane SSD + Intel Memory Drive Technology DRAM-like Performance in Select Applications All DRAM DRAM + Intel Optane SSD + Intel Memory Drive Technology 2322 GFLOPS 2605 GFLOPS 1077 TPS (transactions per second) up to 1.1x up to 870 TPS 80% faster matrix multiplication with optimized data locality 1 of the TPS in MySQL OLTP workloads 2 Performance varies by workload 1. Intel Optane + IMDT configuration 2 x Intel Xeon CPU E Ghz, Intel Server Board S2600WT, 128GB DDR4 + 4* Intel SSD Optane (SSDPED1K375GA), CentOS All DRAM configuration 2 x Intel Xeon CPU E Ghz, Intel Server Board S2600WT, 768GB DDR4 CentOS Test - GEMM, segment size 18689, factor 22, threads Intel Optane + IMDT configuration 2 x Intel Xeon CPU E Ghz, Intel Server Board S2600WT, 128GB DDR4 + 4* Intel Optane SSD (SSDPED1K375GA), CentOS All DRAM configuration 2 x Intel Xeon CPU E Ghz, Intel Server Board S2600WT, 768GB DDR4 CentOS Test - Sysbench 0.5 configured for 70/30 Read/Write OLTP transaction split using a 675GB database. Tests document performance of components on a particular test, in specific systems. Differences in hardware, software, or configuration will affect actual performance. 33

23 Intel Optane SSDs and Intel 3D NAND SSDs in the ruler form factor will come to market in the near future an ssd revolution ruler form factor

Optimized for Space Efficient Capacity per Drive 9.

5mm pitch enables denser server designs Designed

5in designed to share physical dimensions with

capacity U.2 SSDs require more complex flex PCB M.

systems, not intended to be serviced (hot-plug)

24 Optimized for Space Efficient Capacity per Drive 9.5mm drive width and 12.5mm pitch enables denser server designs Designed to accommodate more media sites (36) U.2 2.5in designed to share physical dimensions with HDDs, platter size dictated form factor High capacity U.2 SSDs require more complex flex PCB M.2 designed for thin and light client and mobile systems, not intended to be serviced (hot-plug) during operation. Limited NAND sites, 110mm has 6 Optimized dimensions for 1U server designs 38.6mm 35

Optimized for Space Efficient Capacity per Server 2U Server U.

Rack space 2U 1U Intel AF1000 Server SSD form factor U.

per drive 4TB 8TB TB/rack unit 48TB 256TB Ruler 8TB 1.

Intel Server Board S2600WP Family. 24 U.

25 Optimized for Space Efficient Capacity per Server 2U Server U.2 15mm 4TB 1U Server up to 96 TB up to 256 TB Storage Capacity 5.3x more tb per rack unit 1 System Type Standard 2U Server Rack space 2U 1U Intel AF1000 Server SSD form factor U.2 15mm Intel ruler Number of SSDs per server Capacity per drive 4TB 8TB TB/rack unit 48TB 256TB Ruler 8TB 1. Source Inte Comparing maximum capacity per 1 rack unit of Intel Server Board S2600WP Family. 24 U.2 bay option using 4TB U.2 15mm Intel SSD DC P4500 tobtb Intel AF1000 Server design. 32 ruler drive bays using BTB ruler form factor for Intel SSD DC P

Built in Serviceability Programmable LEDs to

Enclosure Management with slot level power

26 Built in Serviceability Programmable LEDs to quickly locate failed drives, offline drives, and unpopulated slots Carrier-less design with integrated latch removes need for drive carriers Enclosure Management with slot level power control enables single drive isolation or system level power loss 37

27 Roadmap to 1PB in 1U in 2018 Storage Capacity 1PB in 1U Intel 3d NAND SSD, 32TB ruler in PB IN 42u W/2 TB HDDs Opening up new use cases in warm storage with disruptive total cost of ownership 38

Intel Enables highly efficient / flexible All-PCIe * Solutions memory Caching and Fast Storage Mainstream storage + DRAM + Intel Optane SSD with Intel

insights from larger working sets New caching or fast storage for the most latency sensitive applications Accelerate cache tier and mixed workloads like

28 Intel Enables highly efficient / flexible All-PCIe * Solutions memory Caching and Fast Storage Mainstream storage + DRAM + Intel Optane SSD with Intel Memory Drive Technology Intel Optane SSD DC P4800X Intel SSD DC P4600 Series Intel SSD DC P4500 Series Intel SSD DC P4501 Series Bigger memory for new insights from larger working sets New caching or fast storage for the most latency sensitive applications Accelerate cache tier and mixed workloads like database for faster results and more capacity High-performance and massively scalable storage *Other names and brands may be claimed as the property of others. 39

29 法務上の注意書きと最適化に関する注意事項本資料の情報は現状のまま提供され本資料は明示されているか否かにかかわらずまた禁反言によるとよらずにかかわらずいかなる知的財産権のライセンスも許諾するものではありません製品に付属の売買契約書 Intel's Terms and Conditions of Sale に規定されている場合を除きインテルはいかなる責任を負うものではなくまたインテル製品の販売や使用に関する明示または黙示の保証 ( 特定目的への適合性商品性に関する保証第三者の特許権著作権その他知的財産権の侵害への保証を含む ) をするものではありません性能に関するテストに使用されるソフトウェアとワークロードは性能がインテルマイクロプロセッサー用に最適化されていることがあります SYSmark* や MobileMark* などの性能テストは特定のコンピューターシステムコンポーネントソフトウェア操作機能に基づいて行ったものです結果はこれらの要因によって異なります製品の購入を検討される場合は他の製品と組み合わせた場合の本製品の性能などほかの情報や性能テストも参考にしてパフォーマンスを総合的に評価することをお勧めします 2017 Intel Corporation. 無断での引用転載を禁じます Intel インテル Intel ロゴはアメリカ合衆国および / またはその他の国における Intel Corporation の商標です * その他の社名製品名などは一般に各社の商標または登録商標です最適化に関する注意事項インテルコンパイラーではインテルマイクロプロセッサーに限定されない最適化に関して他社製マイクロプロセッサー用に同等の最適化を行えないことがありますこれにはインテルストリーミング SIMD 拡張命令 2 インテルストリーミング SIMD 拡張命令 3 インテルストリーミング SIMD 拡張命令 3 補足命令などの最適化が該当しますインテルは他社製マイクロプロセッサーに関していかなる最適化の利用機能または効果も保証いたしません本製品のマイクロプロセッサー依存の最適化はインテルマイクロプロセッサーでの使用を前提としていますインテルマイクロアーキテクチャーに限定されない最適化のなかにもインテルマイクロプロセッサー用のものがありますこの注意事項で言及した命令セットの詳細については該当する製品のユーザーリファレンスガイドを参照してください改訂 #

30 Add Legal Disclaimers here Add optimization notice, performance claims, copyright year, and any branded Intel products as follows: For more complete information about compiler optimizations, see our Optimization Notice at Intel Corporation. All rights reserved. Intel and the Intel logo are trademarks of Intel Corporation or its subsidiaries in the U.S. and/or other countries. *Other names and brands may be claimed as the property of others. 41

単位、情報量、デジタルデータ、CPUと高速化～ICT用語集～

単位、情報量、デジタルデータ、CPUと高速化～ICT用語集～ CPU ICT mizutani@ic.daito.ac.jp 2014 SI: Systèm International d Unités SI SI 10 1 da 10 1 d 10 2 h 10 2 c 10 3 k 10 3 m 10 6 M 10 6 µ 10 9 G 10 9 n 10 12 T 10 12 p 10 15 P 10 15 f 10 18 E 10 18 a 10 21