Retention Optimized Reading (ROR)
Retention Optimized Reading learns per-block NAND read-reference settings to reduce retention-related errors and retry work.
The controller periodically calibrates the last-programmed page in each block, then uses its learned voltage as a starting bound for ordinary reads. Calibration explores lower and higher settings; the subsequent retry search moves downward. The proposed schedule runs daily and after power-on. [1]
The last-programmed page has the shortest retention age in the assumed sequentially filled block. Its higher optimum motivates the bound; a shared physical block does not establish a universal voltage optimum across 3D NAND layers. [2] [3]
In the nominal-lifetime evaluation, 10.1% lower modeled BCH decoding latency translates to about 2.4% lower flash-read latency when decoding contributes about 24% of the total. These are different denominators, not measured application speedups. [1]
Relations
- related
- NAND read-retry, NAND read-reference voltage, Flash data retention, Retention Failure Recovery (RFR), Retention Model Aware Reading (ReMAR)
Limitations
- The evaluation combines 2y-nm MLC characterization with a modeled 512 GB SSD, 14 workload traces and an assumed seven-day refresh period. [1]
- Calibration needs raw-NAND sensing control and extra reads; it does not replenish stored charge or guarantee recovery after ECC failure. [1]
- Layer variation and rapid early retention loss in the studied 3D MLC chips limit direct transfer of planar per-block tuning assumptions. [3]
Citations
- Cai et al. Data Retention in MLC NAND Flash Memory: Characterization, Optimization, and Recovery. HPCA 2015, sections 3.1 and 4.
- Cai et al. Experimental Characterization, Optimization, and Recovery of Data Retention Errors in MLC NAND Flash Memory. 2018, section 3.
- Luo et al. Improving 3D NAND Flash Memory Lifetime by Tolerating Early Retention Loss and Process Variation. POMACS 2018, sections 4.3, 6.1 and 6.3.
Sources & caveats
- Literature synthesis, not a hardware reproduction. The 2018 retrospective is by the same research group and is not an independent replication.
- HPCA 2015 section 4.1 calls the last-programmed page's retention age largest, contradicting time since programming; the 2018 retrospective explicitly states lowest.