PCIe SSDs Sequential performance bottleneck removed Protocol overhead / bridging removed: Better latency NVMe will enable PCIe to be ubiquitous SATA 2.4X.. SATA 6Gb/s Limit..[r]
(1)Understanding Why Demand for SSDs is Soaring Ryan Smith Sr Manager – SSD Product Marketing, Samsung Semiconductor, Inc YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (2) What is a HDD? milliseconds HDD = Hard Disk Drive Block Based Device (e.g., 512-byte) Capacities in Base 10 Drivers are common and ubiquitous RPM has peaked Avg Latency 10 3,600 4,200 4,500 4,900 5,200 5,400 7,200 10,000 12,000 15,000 YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (3) What is Flash? Non-volatile computer storage chip that can be electronically programm ed and erased Definition and Ingot Image Source : http://en.wikipedia.org/wiki/Silicon YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (4) What is an SSD? SSD = Solid State Drive RAM-based introduced in 1970’s Flash-based version in 1990’s Today, it typically uses NAND Flash Don’t complicate it it’s just a really fast drive! 140,000 120,000 100,000 80,000 60,000 40,000 20,000 2011 2012 2013 PC 2014 2015 2016 2017 Enterprise Source : Forward Insights, 3Q13 YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (5) Why an SSD? Three things that dictate the speed of your PC/Server: • CPU, DRAM, and HDD Everything is speeding up Except the HDD Memory: • Larger footprint • Higher bandwidth Processor: • Multi-core • Higher bandwidth Performance Closing the gap with Solid State Storage Storage: • Minor throughput improvements • Currently solved with spindles Time YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (6) Why an SSD? Lower response times (latency) Higher IOPS and Throughput ~3,000K RPM HDD Lower Power No RVI Issues, More reliable http://tiny.cc/yellatstorage Random Performance (IOPS) Power Consumption (Watt) SV843 15K RPM HDD SV843 15K RPM HDD 88K X200 12.6 X75 8.5 29K X38 -80% -88% 2.5 14K 1.0 Read 70:30 Test Environment : IOMeter2008 Write Idle Active Test Environment : IOMeter2008 / 4KB RND R70:W30 Source : Samsung YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (7) SSD Performance SSD Performance cannot be matched 10% 80K 15 x 73K 5K HDD PM830 PM841 YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (8) SSD Performance 90,000 IOPS 900 Hard Drives Samsung SSD YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (9) So what’s there to know about an SSD? SSD Key Characteristics SSD Components NAND Characteristics P/E Cycles WAF TBW / WPD SMART Host Interface Sustained vs Peak Performance Benchmarking 3,000 ML C SSD Influencers TRIM Over-provisioning Changing Workload User Area O/P YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 Reserved (10) SSD Key Characteristics YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (11) SSD Components Host/NAND Controller Firmware NAND Flash DRAM Capacitors (optional) NAND Controller DRAM Firmware Host Interface DRAM Controller Firmware NAND Flash SSD Image Source : Anandtech All components work closely together YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (12) NAND Characteristics Types of NAND 1 ML C 1 1 3-5K P/E Cycles year retention 10-30K P/E Cycles month retention SLC 90-100K P/E Cycles mo – yr retention 1 NAND Hierarchy • Pages: Smallest unit that can be read/written (e.g., 8KB) • Erase block: Groups of pages (e.g., 64 pages @ 8KB = 512KB) YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 Enterprise 1 500-1K P/E Cycles year retention EMLC 1 PC TLC TLC MLC E-MLC SLC Geometry / Lithography • 4xnm, 3xnm, 2xnm • Smaller = Less Cost • • • • (13) P/E Cycles Program / Erase Cycles The # of times a given NAND cell can be programmed & erased As geometries shrink, error correction must get better It’s like a car warranty! • years or 50,000 miles • years or 3,000 P/E Cycles Not a useful characteristic by itself ECC Requirements 3,000 3xnm 2xnm YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 2ynm (14) Write Amplification Factor (WAF) Write Amplification Factor Bytes written to NAND versus bytes written from PC/Server Bytes written to NAND Bytes written from Host WAF = WAF means 1MB from host writes 1MB to NAND WAF means 1MB from host writes 5MB to NAND Factors that can affect WAF: Flash Translation Layer (FTL) Controller Wear Leveling Over-provisioning Garbage Collection Host Application Write Profile (Ran vs Seq) Free user space / TRIM YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (15) Write Amplification (WAF) Example Below example illustrates WAF of 4KB from Host Z Cache LBA Flash SSD Host Host wants to update LBA Z A C E LBA Z C E B D F No more free pages A C E B D F Z C E B D F B D F Z C E Z C E Need to erase entire block Read existing data to Cache Time Erase block B D F B D F Write modified page and old pages back to Flash YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 24KB to NAND (16) TBW TeraBytes Written # of terabytes you can write to the drive over it’s useful life TBW = (Capacity GB/1000) x PE Cycles WAF Examples: ((128GB / 1000) * 3000 PE) / = 76.8 TBW ((128GB / 1000) * 3000 PE) / 2.5 = 153.6 TBW ((256GB / 1000) * 3000 PE) / = 153.6 TBW ((128GB / 1000) * 30000 PE) / = 768 TBW YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (17) WPD Writes Per Day Same as TBW but normalized on capacity and lifetime How many complete drive writes can be done over a period of time WPD = TBW * 1000 / Capacity GB Years * 365 Examples: (1,730 TBW * 1,000 / 480GB) / (5yr * 365) (1,730 TBW * 1,000 / 400GB) / (5yr * 365) (1,730 TBW * 1,000 / 400GB) / (3yr * 365) = 1.97 WPD = 2.37 WPD = 3.95 WPD YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (18) SMART Attributes Look at health and various statistics Allows for predictable maintenance windows Calculate WAF, TBW Host GB written = [ID241] / (2/1024/1024) NAND GB written = [ID177] * Capacity GB WAF = NAND GB / Host GB Expected Life (yrs) = Warranty PE * ([ID9]/24/365) / [ID177] ID Attribute Name Reallocated Sector Count Power-on Hours 12 Power-on Count 177 Wear Leveling Count 179 Used Reserved Block Count 180 Unused Reserved Block Count 181 Program Fail Count 182 Erase Fail Count 187 Uncorrectable Error Count 195 ECC Error Count 199 CRC Error Count 241 Total LBA Written YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (19) Host Interface This is how you communicate to the SSD So many choices • SATA • SAS • PCIe (NVMe, SCSIe, SATAe, Proprietary) Which is right for you? PC Server Current SATA PCIe SATA SAS PCIe Future SATA PCIe SATA SAS PCIe External Storage SATA + SAS bridge SAS SAS PCIe YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (20) PCIe SSDs Sequential performance bottleneck removed Protocol overhead / bridging removed: Better latency NVMe will enable PCIe to be ubiquitous SATA 2.4X PCIe 3.1X SATA 6Gb/s Limit 2.7X 1.8X YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (21) Sustained vs Peak Performance There can be significant differences in sustained vs peak Run enterprise benchmark (e.g., SNIA RTP 2.0) Or even better, run your own workload (or simulated) [IOPS] Samsung Samsungvs vsVendor Vendor“X” “X” 11x Sustained Random Writes 11x Sustained Random Writes [Ran Performance @ 4KB] 94% 94% below below Peak Peak Samsung 128GB 4KB Ran R/W 100/0 (NCQ=16) Samsung 200GB 4KB Ran R/W 65/35 (NCQ=16) [MBs] Samsung Samsungvs vsVendor Vendor“X” “X” 2x Sustained Sequential 2x Sustained SequentialWrites Writes [Seq Performance @ 1MB] 95% 95% below below Peak Peak 99% 99% below below Peak Peak Vendor “X” 160GB 4KB Ran R/W 0/100 (NCQ=16) Samsung 128GB 1MB Seq R/W 100/0 (NCQ=16) Samsung 200GB 1MB Seq R/W 65/35 (NCQ=16) 95% 95% below below Peak Peak Vendor “X” 160GB 1MB Seq R/W 0/100 (NCQ=16) Source : Samsung / SNIA RTP2.0 Benchmark YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (22) Benchmarking Synthetic or actual workload & take measurements Benchmark URL SNIA RTP 2.0 http://www.snia.org/tech_activities/standards/curr_standards/pts Iometer http://sourceforge.net/projects/iometer/ ATTO Disk http://www.attotech.com/products/product.php?sku=Disk_Benchmark CrystalDiskMark http://crystalmark.info/software/CrystalDiskMark/index-e.html HD Tune Pro http://www.hdtune.com/ AS SSD (SSD) http://alex-is.de/PHP/fusion/downloads.php?download_id=9 Anvil (SSD) http://thessdreview.com/latest-buzz/anvil-storage-utilities-releases-new-storage-and-ssd-benchmark/ Scripts Have multiple “dd” running with best guess workload, capturing timing/speeds Real Workload Capture trace during real workload and playback (ioapps, blktrace/btereplay) YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (23) SSD Reviewers Good SSD Review sites available YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (24) SSD Influencers YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (25) TRIM SSD Host Helps the SSD know which blocks aren’t used Widely supported standard: Windows, Mac OS X, Linux, hdparm Better sustained performance and extends TBW Without TRIM, SSD only knows block isn’t used once the same LBA is writt en to No TRIM needed Bye Hi LBA LBA Hi Hi LBA Time TRIM makes SSD aware Bye LBA LBA Hi LBA Hi ? Hi LBA Bye LBA TRIM LBA Bye LBA Hi Bye LBA Time YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (26) Base-2 versus Base-10 Base Units KiB, MiB, GiB, TiB 10 KB, MB, GB, TB Difference KB MB GB TB 1,024 1,048,576 1,073,741,824 1,099,511,627,776 1,000 2.4% 1,000,000 4.9% 1,000,000,000 7.4% 1,000,000,000,000 10.0% 238GB Base-2 7.4% difference can be used for wear-leveling, bad blocks, etc 255GB Base-10 YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (27) Over-Provisioning Helps a few things: • Improves Write Performance, Reduces WAF, Increases TBW 512GB User Area O/P Reserved 112GB 400GB 28% O/P Sample 128GB SSD 480GB 400GB Over-Provisioning 7% 28% Random Read (8K) IOPS 89K 89K Random Write (8K) IOPS 13,000 Sequential Read (64K) MB/s 530 530 Sequential Write (64K) MB/s 360 360 4KB Random WAF 5.4 -58% 2.25 4KB Random TBW 1,730 2.4x 4,096 4KB Random WPD (5yr) 1.8 3.1x 5.6 3.5x 512GB Base-2 to Base-10 conversion: 549,755,813,888 to 512,000,000,0000 (7.4%) 35,000 These performance numbers are fictitious but represent the actual benefits seen during tests YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (28) Over-Provisioning However, there can be diminishing returns on OP levels YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (29) Change Write Workload Write sequentially instead of random to reduce WAF • If you have control of the I/O to the disk, this will pay of Random MLC 480GB SSD 1,730 TBW Align your2writes the page boundaries (e.g., 8KB) Pages with needed Sequential 5.8x 10,000 TBW Only Page needed Host LBA LBA 16 8K 8K Change block alignment SSD 8K LBA 8K LBA 16 8K 8K LBA 8K 8K LBA 16 YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (30) Applications of SSDs YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (31) HDD Replacement Replace boot drive or main storage Fastest and easiest way to experience SSDs Server Storage HDD HDD SSD SSD YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (32) Caching Appliance Read and/or Write Cache Caching will take place in one of the following: • Between servers and storage, typically in a SAN • Inside server Used to speed up legacy or slower storage Servers Cache SSD SSD Storage HDD YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (33) Tiered Storage An external storage device (NAS, SAN) Only puts “hot” or “critical” data on SSD Most of the storage is still on HDD Servers SSD HDD HDD Storage YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (34) All Flash Storage External storage based on 100% SSD/Flash Typically uses MLC and de-duplication/compression to achieve better prici ng Designers of these systems are Flash experts Servers SSD SSD SSD Storage YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (35) Datacenter Trends YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (36) Mainstream 2TB 2.5” SSDs 2014 brings similar 1U server capacity and unmatched performance HDD SSD 3+1 RAID (5TB) RAID Card x 3.5” HDDs <1K IOPS 15TB Usable IOPS/GB 0.05 vs 20 RAID (2TB) HBA x 2.5” SSDs 320K IOPS* 16TB Usable * 6Gb x8 HBA bottleneck YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (37) SATAe/NVMe in the PC & Server Protocol Standardization PC-adoption is key to bringing mainstream SATA will start to shift to PCIe in PC/Server Form factors: HHHL, 2.5” SFF8639, M.2 YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (38) Educated End-Users End-users’ questions are more advanced Customers are shaping next generation Tailoring workloads specifically for SSDs Random Sequential Random MLC 480GB SSD 1,730 TBW Sequential 5.8x 10,000 TBW Looking closely at SMART attributes Aligning writes to page boundaries YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (39) IOPS/GB + Large Capacities For a given unit of content, need minimum level of performance Requirement of >10 IOPS/GB was difficult to use high capacity 10K RPM HDD Capacity IOPS IOPS/GB 2011-12 SSD 2013 SSD 300GB 600GB 120GB 240GB 480GB 960GB 200 200 1,500 1,500 14,500 14,500 0.66 0.33 12 28 14 Other factors 10Gb networking Server density Rack optimization YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (40) Density at Component and System Level SSDs more dense by 2015; However, already there! 2013 2014 2015 2016 3.5” 4TB 5TB 6TB 8TB 2.5” 2TB 4TB 8TB 16TB 2U12 3.5” 2U24 2.5” 2U48 2.5” 5TB 60TB 48TB* N/A 2TB 24TB 48TB 96TB 4TB 48TB 96TB 192TB * 2TB largest 2.5” HDD YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (41) QoS – Quality of Service If consistency is important, SSDs can vary depending on degree 2000 90% 1800 80% Write latency (us) 1600 70% 1400 60% 1200 50% 1000 40% 800 30% 600 20% 400 10% 200 Samsung 10 Competitor 11 12 13 14 15 16 17 0% Difference YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (42) Datacenter Write Workload Datacenter workload 0-10 WPD Competitors 1 0.3WPD Heavy Read ML C Heavy Write EMLC 1 10WPD TLC 1 ML C 1 EMLC 1 0.3WPD 2-4WPD 10WPD YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (43) Dell’s customers prefer Samsung A B C Samsung is the preferred vendor for Dell’s customers YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (44) Why Samsung? 100% Vertically Integrated Samsung produces all critical components that make up a SSD, resulting in: Increased Performance, Superior Reliability, Ownership of design Market Leader provides highest quality products #1 #1 #1 Source: Gartner 2Q13 Report YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (45) Why Samsung? “It’s what’s inside that counts” Only the best Flash makes it into Samsung SSDs Samsung SSD Flash is 20 times better than conventional Flash! YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (46) Why Samsung? Highest Quality NAND Server grade NAND is used for all Samsung SSDs Samsung SSDs prove to have the lowest failure rate compared to any competitor SSD Failure Rate * from Oct’12 to Apr’13 [Source :HardWare.fr, May’13] YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (47) Why Samsung? Innovation Samsung is leading the SSD industry with innovative products 2006 2008 Samsung SSD Introduction World’s 2-bit SSD st 2012 1st 3-bit SSD 1H2013 2H2013 World’s PC PCIe SSD World’s 1st 3D-NAND st Clear Market leader in SSD 2012 [Source : iSuppli, Q1 2013] YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (48) Planar NAND Vertical NAND Over 30nm Over 30nm Over 30nm Cell e- e20nm e e10nm Cell-to-Cell Interference Patterning eV-NAND Cell - Over 30nm ee- e- e- eYYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (49) V-NAND Era for the Future Design Rule (nm) 2D Planar 3D V-NAND / No Patterning Limitation stack 24 stack 16Gb 128Gb 1Tb 128Gb ‘03 ‘05 ‘07 ‘09 ‘11 ‘13 ‘15 YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 ‘17 Year (50) Innovation Result: 128Gb V-NAND Chip “The World’s 1st 3D V-NAND Flash Mass Production” Compared to 20nm planar NAND Flash: • 2X Density and Write Speed • ½ Power Consumption • 10X Endurance 128Gb V-NAND Flash 24 Layer Cell Structure YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (51) Dell / Samsung - Partnership Partners for 25 years Samsung technology ofered in leading edge Latitude and XPS notebooks YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (52) Dell / Samsung – Solving Real World Problems Dell PowerEdge R720n Server Solution with Samsung SSDs and Windows Server 2012 VDI performance comparison of Samsung SSDs vs SAS drives in a Dell PowerEdge R720 Dell PowerEdge R720 Server Solution with Samsung SSDs and Windows Server 2012: Supporting Virtual Desktops in Remote Offices Enhancing server performance and scalability with solid-state drives http://www.samsung.com/us/dellcio YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (53) SSD Resources Dell PCs: www.dell.com/business/ssd Dell Servers: www.dell.com/business/ssd/server Samsung SSDs: www.samsung.com/us/dellpromo YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (54) Thank you! Visit our booth or email us any questions at ssd@ssi.samsung.com YYYY.MM.DD / 홍길동 책임 / xxxxxx 팀 (55)