Design Tradeoffs for SSD Reliability

on YouTube ↗24m

" summary

Measured raw bit error rates rise sharply across generations, and uncoordinated use of data re-read, intra-SSD redundancy, and scrubbing drives big slowdowns. Repeated re-reads with a 75-bit ECC incur about 35% overhead at end-of-life; striping-based redundancy increases write/read amplification and worsens tail latency for worn devices; background scrubbing is least harmful. A selective scheme that migrates cold data to parity, conditionally re-reads hot reads, and targets scrubs matches an ideal oracle on many workloads but can hurt very uniform read patterns.

AI-written summary · about

" chapters (16)

1. 0:00Intro
2. 0:32High-level objectives
3. 1:29How bad is it?
4. 2:27SSD's reliability issue
5. 3:05Flash memory errors
6. 4:03Flash memory error modeling RBER (cycles, time, reads)
7. 5:32Error model: 3x-nm MLC (2011)
8. 7:20Error model: 3D TLC (2018)
9. 8:25Error correction code
10. 10:36Evaluation: data re-read
11. 12:18Why is data re-read bad?
12. 14:27Evaluation: redundancy
13. 15:44Observations
14. 16:38Holistic reliability management . Cold data
15. 20:26The bright side of flash memory
16. 20:43The dark side of flash memory