Measuring the Cost of Reliability in Archival Systems James Byron Center for Research in Storage Systems University of California, Santa Cruz
Overview ❖ Motivation & Goals ❖ Analysis: HDD Lifespan ❖ Archival Reliability and Cost ❖ Conclusion CRSS Confidential 2
Motivation ❖ Changing technology and demands for archival storage Changing storage technologies over time ➢ Expectation for predictably reliable and cost-effective storage ➢ ❖ Opportunity to meet future demand CRSS Confidential 3
Goals ❖ Provide insights on development of next-gen storage technologies ❖ Propose designs for better archival systems ❖ Create tool to aid long-term planning How to plan for reliability as devices change? ➢ CRSS Confidential 4
Overview ❖ Motivation & Goals ❖ Analysis: HDD Lifespan ❖ Archival Reliability and Cost ❖ Conclusion CRSS Confidential 5
Analysis: HDD Lifespan ❖ Inquire into changes in HDD usage patterns ❖ Backblaze data covers 2013 to 2020 Date drives added and removed ➢ Failure or retirement ➢ ❖ Observe how retirement rate changes as devices age ❖ Observe how device reliability affected by its age and generation CRSS Confidential 6
Analysis: HDD Lifespan ❖ Plot retirement and failure rates ❖ Failure rate increases after 5 years of operation ❖ Retirement rate not predicated by increase in failure rate CRSS Confidential 7
Analysis: HDD Lifespan ❖ Days to reach various drive failure rates by year ❖ Larger numbers (more days) are better ❖ Improvement in HDD reliability 2013-2017 Failure Rate 2013 2014 2015 2016 2017 2018 1% 67 143 168 268 332 252 2% 208 332 357 565 589 440 3% 394 503 554 948 745 582 4% 491 684 721 1349 1104 - 5% 618 816 941 - - - 10% 1195 - - - - - 15% 1793 - - - - - CRSS Confidential 8
Analysis: HDD Lifespan ❖ How would longer-lasting HDDs affect archival costs? ❖ How do different storage technologies compare in terms of cost and reliability? CRSS Confidential 9
Overview ❖ Motivation & Goals ❖ Analysis: HDD Lifespan ❖ Archival Reliability and Cost ❖ Conclusion CRSS Confidential 10
Simulation Model ❖ Implemented an archival system simulator ❖ Model devices, rates of development, device failures ❖ Simulate designs with various amounts of RAID parity for each device ❖ Compare systems based on cost, reliability, and device type ❖ 25 year simulation period CRSS Confidential 11
Cost of Reliability ❖ Assumptions used in simulation ❖ Archival data growth ➢ 1PB of data growing 30% annually ❖ Storage device capacity growth ➢ Each following historical trends ➢ Growth rate slows with time ➢ DNA’s growth continues unabated CRSS Confidential 12
Cost of Reliability ❖ Comparing cost for “nines” of reliability ❖ Tape ➢ Optimistic stable AFR: 0.000075% ➢ Pessimistic growing AFR: 0.3% CRSS Confidential 13
Cost of Reliability: HDD ❖ How much could be saved if HDDs lasted longer? ❖ How long should we keep HDDs? ❖ 7-9 years is optimal for HDDs with growing failure rates ❖ 11+ years is best for HDDs with stable failure rates ❖ 10% cost difference CRSS Confidential 14
Experiments: Device Design ❖ How will changes in design affect storage technologies for reliable archival storage? ❖ Example: Dysan Disk Pack (c. 1970), HDD (c. 1990), Iomega Jaz (c.1996) CRSS Confidential 15
HDD With Separable Platters ❖ How would HDD cost change with development of removable platters? ❖ 20% savings for 3-year upgrade cycle ❖ 42% savings for 1-year upgrade cycle CRSS Confidential 16
SSD Development and Reliability ❖ Is it more important to have high reliability or capacity growth for SSDs? ❖ Double AFR for SSDs (left) correlates with 8% increase in cost ❖ Extending SSD development apace for 3 additional years into future (right) reduces cost 52% CRSS Confidential 17
Tape With Higher AFR ❖ How will tape be affected if its reliability decreases with its increasing capacity? ❖ 10x increase in AFR correlates to an 81% cost difference CRSS Confidential 18
Archival Glass ❖ How is the cost of glass affected by drive cost? ❖ 10x increase in read drive cost increases cost by 78% ❖ Only slight cost increase for added reliability CRSS Confidential 19
Archival Synthetic DNA ❖ Is it better to have high capacity or a flexible design for DNA? ❖ DNA molecules ➢ Vary capacity ➢ Also vary forward compatibility of DNA with drives ❖ 100x capacity increase reduces cost by 68% ❖ Forward compatibility reduces cost by 60% CRSS Confidential 20
Overview ❖ Motivation & Goals ❖ Analysis: HDD Lifespan ❖ Archival Reliability and Cost ❖ Conclusion CRSS Confidential 21
Conclusion ❖ Improving HDD reliability could improve its competitive position ❖ Tape remains cost-effective across a range of reliability scenarios ❖ SSD becomes more competitive with higher capacity, even with lower reliability ❖ Glass and DNA are promising for archival storage CRSS Confidential 22
Thank You Contact James Byron jbyron@ucsc.edu CRSS Confidential 23
Recommend
More recommend
