Write combining
Write combining in storage buffers writes so that a later transfer can coalesce or reorder them before reaching the underlying storage.
Combining is not a universal reduction in NAND writes. BPLRU can add padding traffic while reducing merge cost. Its power-loss remedies were proposed, not validated. [1]
Skylight measured two SMR drives, 5 TB and 8 TB, whose persistent caches deferred band cleaning. It reported 20/25 GiB caches, 15-40 MiB bands and 0.6-1.6 seconds per modified-band cleaning; these are not universal drive specifications. The dynamically mapped emulator is a separate configuration, not a measured capability of those drives. [2]
In the measured SMR design, extent-map entries can run out before the disk cache fills. Journal-entry packing also depends on write size and queue depth. Cache capacity in GiB is therefore not a fixed host-write budget before cleaning. [2]
Engineering implication: distinguish incoming writes, buffer-to-device traffic and media work. A shorter request stream does not by itself establish lower write amplification, crash safety or better throughput.
Relations
- related
- Write amplification, Shingled magnetic recording (SMR), Solid-state drive (SSD), Flash translation layer (FTL), Read-modify-write, SMR persistent cache, Write-back cache, Block Padding LRU (BPLRU), Restricted Insertion Priority Queue (RIPQ)
Citations
- Hyojun Kim and Seongjun Ahn. BPLRU. FAST 2008, sections 3-5.
- Abutalib Aghayev and Peter Desnoyers. Skylight. FAST 2015, sections 1, 3-4 and Table 3.
Sources & caveats
- Storage buffering only, not the CPU memory-type meaning.
- c2 is manufacturer-authored historical research; c3 measures two Seagate drives and separately evaluates emulators.
- The final paragraph is an engineering inference, not a new measurement.