Space-Time Occupancy Heatmaps
What it shows. Each panel is one engine. The vertical axis is the reclaim unit (block / RU / zone id), the horizontal axis is host-write time, and colour is the valid fraction of that unit (light = mostly invalid/empty, dark green = full of valid data).
Why it matters. The paper's central point is not just how much amplification (a scalar WAF) but how the wear and invalidity are distributed. These heatmaps make the distribution visible: you can see the device fill up (the diagonal ramp), reach steady state, and how each placement policy organizes hot vs cold data over time. An engine that keeps units either near-empty or near-full (sharp two-tone) reclaims cheaply; one that leaves everything half-valid pays a high GC cost.
How to read it. Compare the texture across engines: FDP-PI and ZNS hot/cold create distinct hot (churning) and cold (stable dark) bands, whereas conventional block GC shows a more uniform speckle.
- Placement is visible. FDP-PI and ZNS hot/cold form a stable dark cold band (cold data parked, never re-migrated) plus a small churning hot band — the separation the scalar WAF only hints at.
- Conventional block GC shows a fine, uniform speckle: hot and cold data mixed across all units, so every reclaim carries some valid pages.
- Sequential / TPC-C fill cleanly then stay dark and stable (no rewrites, WAF 1.0); uniform stays uniformly mid-toned (everything ~half-valid → expensive GC).
