Write Amplification
Subject: Computer and Technology noun /raɪt ˌæm.plə.fəˈkeɪ.ʃən/ 1 definition
- data-storage
- flash-memory
- storage-engines
- system-performance
-
Definition (noun)
Entry #313 · UpdatedWrite Amplification is a storage phenomenon in which the volume of physical data written to underlying storage media exceeds the volume of logical data originally submitted by the host or application. Quantified by the Write Amplification Factor (WAF)—the ratio of physical bytes written to logical bytes written—it occurs whenever hardware constraints or storage algorithms must pad, rewrite, relocate, or replicate surrounding data to persist a logical change. Because NAND flash memory programs data in pages but can only erase entire multi-page blocks, an SSD's Flash Translation Layer (FTL) must copy valid pages to clean blocks during garbage collection before erasing stale ones, while software layers such as log-structured merge-trees (LSM-trees) and copy-on-write filesystems add background compaction and metadata rewrites. Distinct from read amplification and space amplification—with which it routinely trades off—elevated write amplification consumes finite flash program/erase (P/E) cycles, degrades sustained write bandwidth, and increases I/O tail latency.
Common Forms
NAND Flash Garbage Collection: Occurs inside an SSD controller's Flash Translation Layer (FTL) when valid pages sharing a partially stale erase block must be copied to a clean block before the original block can be erased and reused.
LSM-Tree Compaction: Occurs in log-structured merge-tree storage engines (such as RocksDB and Apache Cassandra) as background compaction repeatedly merges and rewrites immutable SSTable files across disk levels to purge stale key versions and tombstones.
Page-Granularity and Parity Rewrites: Occurs in B-tree databases and parity-based RAID arrays (such as RAID 5 or RAID 6) when modifying a small logical record or partial stripe forces the system to flush an entire fixed-size page or perform read-modify-write parity updates.
Journaling and Copy-on-Write Cascades: Occurs when crash-safe systems persist updates twice via a write-ahead log (WAL) or journal (as in ext4 and relational databases), or when copy-on-write filesystems (such as ZFS and Btrfs) allocate new data blocks and rewrite parent metadata pointers up to the tree root.
Example
In a write-heavy RocksDB key-value store using leveled compaction, ingesting 100 GB of random 1 KB key-value updates triggers repeated multi-level SSTable merges totaling roughly 1 TB of disk writes—a software-layer Write Amplification Factor of 10. Below the filesystem, an NVMe SSD's Flash Translation Layer (FTL) performs garbage collection on fragmented erase blocks and may commit 2.5 TB of physical NAND flash writes for that 1 TB of host writes, yielding a combined end-to-end WAF of 25 and accelerating flash wear.