Deterministic accounting / no device measurements

RAID write costs: reads, writes and payload

56 deterministic RAID5/P+Q examples compare read-modify-write and reconstruct-write with explicit byte denominators. No device measurements or timing predictions.

JSON and provenance / Model code / Research explanation

Accounting assumptions

Healthy fixed-width parity rows with complete equal-size block replacements and cold cache. k data blocks, p parity blocks, u replacements. RMW reads u+p and writes u+p; reconstruct reads k-u and writes u+p. One transfer of one member block is one operation. Each row is independent. The two strategies have equal operation count when 2u+p=k.

At the same 16 KiB host payload, four isolated one-block RMW updates in a four-data-member RAID5 row use 16 member-block transfers. A coalesced full-row update uses 5. This comparison assumes no cache reuse between isolated updates; it is not a measured speedup.

Reproduce

The module needs only Node.js; it performs no device I/O and uses no random seed or dependencies. With the linked module in the current directory, this command prints the identical CSV.

node --input-type=module -e 'import { toCsv } from "./raid-write-costs.mjs"; process.stdout.write(toCsv())'

JSON includes SHA-256 hashes of the model and CSV. Generation time is not an observation period. Displayed ratios are rounded to four decimal places; CSV retains the numeric values.

Limits

• Synthetic arithmetic, not sampled observations, a benchmark or a paper replication. No confidence interval applies.
• All members are available. This is not an encoder, RAID simulator or degraded-write policy.
• RMW at u=k is forced as a comparison, not recommended or asserted controller behavior.
• Excludes caching, merged commands, journals, flushes, metadata, retries, CPU time and device-internal amplification.
• The 4 KiB block is an equal slice at the same offset across members, not a required configured RAID chunk size.
• Minimum transfer count is not a guaranteed minimum latency or maximum throughput.

Column definitions

k

Data blocks per parity row.

p

Parity blocks: 1 for RAID5, 2 for P+Q.

u

Complete data blocks replaced together in the same row.

blockBytes

Bytes per equal-size member block; illustrative 4096-byte input.

strategy

Forced rmw or reconstruct; not an observed controller choice.

reads

Member-block reads before parity calculation, with no cache hits.

writes

Changed-data plus parity member-block writes.

operations

Reads plus writes; counts, not seconds.

hostBytes

Replacement payload bytes: u times blockBytes.

readBytes

Member-interface bytes read.

writeBytes

Member-interface bytes written; excludes NAND-internal work.

operationsPerHostBlock

Member-block reads and writes divided by u.

writeBytesPerHostByte

Member write bytes divided by host payload bytes.

ioBytesPerHostByte

Member read plus write bytes divided by host payload bytes.

Sources and license

CC0-1.0 for Hesela's original model and generated rows. Cited works retain their own rights.

1. Peter M. Chen, Edward K. Lee, Garth A. Gibson, Randy H. Katz and David A. Patterson. RAID: High-Performance, Reliable Secondary Storage. ACM Computing Surveys 26(2), 145-185 (1994). Author manuscript: sections 3.2.5-3.3.2, Table 3 and 4.1. (accessed 2026-10-11)
2. Dan J. Williams. MD RAID Acceleration: Support for Asynchronous DMA/XOR Engines. Linux Symposium 2006, Volume Two, printed pp. 409-414; sections 3-5. The PDF filename uses different pagination. (accessed 2026-10-11)
3. Linux kernel documentation: RAID 4/5/6 cache, write-back mode and implementation. Rolling documentation; accessed 11 October 2026. (accessed 2026-10-11)

Definitions: read-modify-write, reconstruct-write and full-stripe write.