SSD Applications in the Enterprise Area Tony Kim Samsung Semiconductor January 8, 2010
Outline Part I: SSD Market Outlook Application Trends Part II: Challenge of Enterprise MLC SSD Understanding SSD Lifetime
Market outlook for Enterprise SSD SSD market is expected to be $4.0B in 2012 SSD adoption ratio is increasing by replacing 15Krpm HDD products In 2013, 14.3% of A/R, 5.0M Set of SSD in Server & Storage Market [ Server & Storage ] 5.1M 3.4M 2.0M 0.8M 0.2M 2009 2010 2011 2012 2013 [Source : Gartner 2Q 09]
Only two I/F standards would remain SAS 6Gb/s & SATA 3Gb/s or Greater HDD/SSD Interface Trend 50,000 45,000 40,000 FC 35,000 SAS 6Gb/s 30,000 25,000 20,000 SAS 3Gb/s 15,000 10,000 SATA 5,000 0 2009 2010 2011 2012 2013 FC 4Gb/s SAS 6Gb/s or Greater SAS 3Gb/s SCSI SATA 3Gb/s or Greater [Source : Gartner, June 09]
Key Advantages of SSD (vs. HDD) Higher Performance More Reliability Less Power Less Cooling Reduced hardware SSDs offer outstanding performance and reliability SSDs significantly decrease datacenter power and cooling SSDs provide the best TCO and Green performance solution
New Paradigm for Enterprise market with SSD Enterprise market started to utilize the benefits of SLC SSD on High IOPS, Low Power & Reliability Streaming and Web servers especially with Read-centric applications are considering use of MLC SSDs for their lower price and higher capacity More Opportunity in Enterprise Market
Challenge of Enterprise MLC SSD Cost: Flash memory dominates the BOM; Cost reduction slow down (3Xnm 2Xnm 1Xnm) Reliability: P/E cycle and Data retention are key to determining SSD lifetime Performance: lifetime trade-offs (IOPS, bandwidth need to define real application usage) More Challenge on Enterprise class SSD
Challenge(1) High Performance Performance Impacts Workload (random vs sequential Read/Write) Spare Area: over-provisioning helps reduce WAF Firmware Efficiency NAND Flash performance SDR DDR (Toggle-mode, ONFI) tprog increase (vs. SLC)
Challenge(2) - Extended Reliability Increased Mapping Efficiency Lower WAF at same over-provisioning ratio Enhance both lifetime and performance via advanced logging scheme Managing decreased P/E cycle Extended MLC NAND adoption Extending P/E cycle, sacrificing data retention time and performance
Performance / Lifetime Trade-Offs Lifetime inversely proportional to Performance Random performance is more important in Server Application More impact on Lifetime of SSD Only limited IOPS can meet 5 year warranty Performance(%) Less capacity would be less performance due to lifetime guaranty Life Time
Impact of Workload on Performance Lifetime highly depends on daily workload (especially Random Write) Assuming 24hr, 365day workload (always working under worst scenarios) Read:Write workload ratio as 90:10, 80:20, 70:30, ~ 0:100 (@4KB) Assuming 10% write portion can provide 5year lifetime with reasonable performance Relative Random Write IOPS % of Daily Write Workload Higher capacity can enable 2x performance or 2x write portion increase
Lifetime Expectation on Drive Capacity The guaranteed year can be extended with larger capacity With the same Random Write IOPS, 800GB MLC SSD can be guaranteed ~8times more than 100GB MLC SSD Reduced performance helps extend life time Assumption : Random Write performance (@8KB size) 24hrs, 365 days usage scenario
Understanding SSD Lifetime(1) SSD Lifetime = f{wai, Performance(IOPS or MB/s)} Estimated Lifetime ( year) = ( Host write per P / E) P / E Limit Daily usage 365day / year where Host write per P/E = # of Host writes inducing Erase cycle of NAND Block = 1/WAI where Daily usage depends on performance (i.e., IOPS or MB/s) - WAI (Wear Acceleration Index) WAI Erase cycle( P / E) # of HostWrites( IOPS) = i.e. WAI depends on Host workload pattern & SSD behavior Lower WAI can extend SSD lifetime for the same workload/performance condition!
Understanding SSD Lifetime(2) - WAF (Write Amplification Factor) WAF = Total amount of PhysicalWrite Total amount of Host Write WAI/WAF is desirable in comparing the efficiency of the mapping algorithm, however highly depends on host workload SSD lifetime depends on 1) Host workload 2) Performance 3) WAI/WAF(FTL Efficiency) WAF can not represent SSD Lifetime correctly because WAF can not directly translate into # of P/E counts WAI is preferable to evaluate the lifetime of SSD
What else can we do? De-duplication / Compression can reduce WAF/WAI? The efficiency of these techniques highly depends on data patterns, not consistent performance/lifetime for different server applications Need to understand host data pattern for applying these concepts Host (FS/driver) can perform better than drive for these technologies
Needs of SSD-friendly Storage Infrastructure - System architecture, OS optimization, File System optimization and application, etc. - Current HBA/RAID technology a bottleneck for SSD performance, because it has been optimized only for HDDs Need to observe some HBA/RAID limit in certain kinds of conditions like RAID configuration IOps 4 Drive File Server with 64K Stripe A B C D E F Q1 Q4 Q16 Q64 Queue Depth File Server : 100% Random, 80% Read
Summary Cost reduction is main drive for accommodating MLC NAND into Server/Storage systems The balance among speed, safety and $$ becoming more important than more IOPS MLC SSD need to be designed to provide breakthrough performance and extend the lifetime System architectures must advance to fully exploit SSD The industry needs JEDEC standard to commercialize more applications
Thank you very much