Mechanism
Valid-Fraction-at-Reclaim (the WAF mechanism)
What it shows. The distribution of the valid fraction of each unit at the moment GC reclaims it. A unit reclaimed while 10% valid costs almost nothing to migrate; one reclaimed at 60% valid forces migrating most of its pages.
Why it's the mechanism. WAF is set almost entirely by this quantity: every valid page in a victim must be copied forward (an extra physical write). The green band marks near-empty (“free”) victims — the cheap reclaims a good policy maximises.
Key result. Across all 69 runs, WAF correlates with mean valid-at-reclaim at Pearson r = +0.92 — far stronger than any spatial/entropy measure. WAF is governed by when pages die (do a unit's pages die together?), a temporal property, not by the spatial layout.
What we found
- The correlation is tight: r = +0.92 between mean valid-at-reclaim and WAF across all 69 runs.
- Concrete points: prxy_1_3M FDP-PI reclaims units only 5.8% valid → WAF 1.06; 80/20 FDP-PI 13.6% valid → WAF 1.10; prxy_0 conv 37.7% valid → WAF 1.54; uniform conv 48.3% valid → WAF 1.56. Higher valid-at-reclaim, higher WAF, every time.
- Good policies push reclaims into the near-empty “free” band (<10% valid); uniform workloads can't, so they pay the most.
prxy_0 (MSR real)
prxy_1 (MSR real)
prxy_1 (3M writes, real)
TPC-C (W2K8, real)
uniform (synthetic)
Zipf-1.2 (synthetic)
80/20 (synthetic)
sequential (synthetic)
← Back to all graphs