NAND copyback: faster relocation is not fewer writes

Which boundary does the copy cross?

In the NAND organization described by Hu and colleagues, copyback moves a page through a local register without taking its payload across the external I/O bus. Their command has same-chip, same-die, same-plane and odd/even page-address restrictions. These are constraints of the studied interface, not universal rules for every NAND generation. [1, section 2.2]

At the controller boundary, Han and colleagues explain the reliability issue: an internal transfer can bypass the controller's ECC check. They describe checking alongside copying, or restricting consecutive copies, as possible responses. A host-level copy request alone does not reveal which physical path or checking policy was used. [2, section 2.1]

During garbage collection, live data must be preserved elsewhere before the victim block is erased. Copyback is one possible relocation mechanism; it is not an alternative to preserving live data.

A copy budget is conditional

Hong and colleagues characterized 81,920 pages from 20 chips of 1x-nm MLC NAND. Their restricted-copyback design bounds consecutive uncorrected moves using wear and retention requirements, then routes data through off-chip ECC. The proposed controller also reserves copyback opportunities for periods of higher demand. [3, sections 3-4]

The reliability characterization and performance evaluation are different evidence: throughput was evaluated in a configured 64-GB emulated system using four workload traces. Neither establishes a safe copy count for a current TLC or QLC product. We use the versioned 2018 manuscript, not an assumed product implementation. [3, section 5]

The overlooked cost: finding a destination

Hu's SSDsim study examines how satisfying copyback address constraints can waste free pages and affect erasures. A fast command can therefore interact badly with an allocation policy. Its performance results are simulator results, not measurements of arbitrary commercial drives. [1, section 4.3]

Han's ZNS+ research similarly distinguishes ordinary compaction from copyback-aware placement. Allocating extra destination space can increase copyback use but weaken space reclamation; the paper leaves a detailed treatment of that tradeoff outside its scope. ZNS+ is a research extension, not a synonym for standard ZNS. [2, section 3.3.1]

One controlled comparison, two different totals

This is Hesela's synthetic accounting example, not a device measurement. Fix 100 host-written pages and 60 relocated pages in one interval. Every page holds 16,384 payload bytes. Each host write and relocation programs exactly one page. Only the relocation path changes.

Let H be host-written pages, G relocated pages, C copyback pages and P payload bytes per page. Then programmed bytes = (H + G)P, write amplification = (H + G)/H, and relocation-channel bytes = 2(G - C)P. The factor two counts the outward and return payload transfers of each off-chip relocation.

The conclusion is deliberately narrow: changing only the path cannot change this ratio's numerator. A real policy can also change the number of relocations, placement, retries or metadata writes, so its amplification must be measured independently. Zero relocation-channel payload does not mean zero NAND activity, total device traffic or latency.

Excluded: ECC/spare bytes, compression, parity, metadata, retries and extra migrations. Eligibility is assumed; there is no error model, retention guarantee, confidence interval or performance prediction. This is exact arithmetic under stated assumptions, not a statistical estimate.

Download scenarios and assumptions (JSON) · Model source (JavaScript). The JSON records the model SHA-256. Regenerate in the corpus repository with node scripts/build-copyback-accounting.mjs; test with node --test scripts/copyback-accounting.test.mjs.

What an engineering report should preserve

Our reporting recommendation: record the NAND part and command rules, the exact ECC boundary, the consecutive-copy history, the source and destination placement policy, and any fallback copies. Separately report payload bytes, physical programs, erasures, errors and request latency over the same workload interval.

A host benchmark without internal telemetry cannot identify those mechanisms on its own. Do not turn a conditional firmware optimization into a universal endurance claim.

Structured definitions: NAND copyback, garbage collection and write amplification. Continue with LDPC sensing and decoding costs or ZNS resource limits.

Primary sources and scope

1. Hu, Jiang, Feng, Tian, Luo and Zhang. Performance Impact and Interplay of SSD Parallelism through Advanced Commands, Allocation Strategy and Data Granularity. ICS 2011, sections 2.2 and 4.3.
2. Han, Gwak, Shin and Hwang. ZNS+: Advanced Zoned Namespace Interface for Supporting In-Storage Zone Compaction. OSDI 2021, sections 2.1 and 3.3.1.
3. Hong, Kim, Park, Jung and Kim. Revitalizing Copybacks in Modern SSDs: Why and How. Author manuscript arXiv:1810.04603v1, 2018, sections 3-5.

Reviewed 11 October 2026. Full-text methods and relevant results examined. No paper figures, underlying chip data or performance results reproduced.