Mechanism
What Predicts WAF — Mechanism Scatters
What it shows. Every (engine, trace) run as a point, WAF on the y-axis. Left: against valid-fraction-at-reclaim (the temporal GC cost). Right: against DoIPD (the spatial invalidity dispersion).
The finding. Valid-at-reclaim predicts WAF tightly (r = +0.92); the spatial dispersion only loosely (r = +0.55). This is the empirical basis for the paper's claim that write amplification is a temporal phenomenon (pages dying together) rather than a spatial one (how mixed the layout looks) — and it is the evidence that refutes the entropy model of FTL behaviour.
What we found
- Temporal beats spatial: WAF vs valid-at-reclaim r = +0.92; WAF vs DoIPD r = +0.55 — over the same 69 runs.
- This is the single figure that carries the paper's thesis: write amplification is set by when pages die (do a unit's pages die together?), not by how disordered the layout looks at any instant.
- Consistent with the separate entropy-model test, where an entropy-optimal GC policy gave the worst WAF (+42% to +124%).
