Invalidity Distribution (DoIPD)
What it shows. Left: the Degree of Invalid-Page Distribution (DoIPD, the spread of invalid pages across reclaim units) measured two ways — before every GC trigger (the steady-state, methodologically-correct measurement) versus at end of trace. Right: histograms of the invalid-fraction per unit.
Why measure before GC. Measuring only at the end of the trace samples a cooled-down device and understates the real dispersion the FTL works against. Snapshotting the distribution at each GC trigger captures the hot steady state that actually drives reclaim decisions.
Relation to WAF. Dispersion of invalidity is a spatial property. It correlates with WAF but only weakly (see Mechanism Scatters, r ≈ +0.55) — the real driver is the temporal valid-fraction-at-reclaim.
- Measuring at end-of-trace understates dispersion. The at-GC-trigger DoIPD is consistently higher than the end-of-trace value on active workloads — the device looks calmer once writing stops than it is during steady-state GC.
- Dispersion alone doesn't set WAF. Across all runs DoIPD correlates with WAF at only r ≈ +0.55 — useful context, but weaker than the temporal predictor (r = +0.92).
- Skewed workloads (Zipf, 80/20) show wide invalidity spread yet low WAF, because their hot pages die together regardless of how mixed the layout looks.
