Figure: paired bars of cache hit rate in per cent at a 32 gibibyte HBM budget, comparing eviction metadata that is reset on eviction against metadata that persists across the tier boundary. LFU goes from 15.56 per cent to 49.71 per cent, a gain of 34.15 points. EMA with alpha 0.01 goes from 31.54 per cent to 56.39 per cent, a gain of 24.85 points. EMA with alpha 0.15 goes from 44.33 per cent to 44.12 per cent, a change of minus 0.21 points, which is a wash. The conclusion is that persistence is worth more than the choice of policy: the 34.15-point gain exceeds the documented plus 2.5 to plus 5.8 point margin of the best scoring function over LRU. It is a protocol result and is testable with no CXL hardware. Simulated and provisional: a single data-structure defect in Revision 1 inverted four of five findings.

Figure 11.2 · Simulated (provisional)

Persistent identity-keyed policy state — the largest effect in the study

Hold the scoring function fixed and vary only whether policy state survives eviction from HBM. This is the one mechanism inside PLACE with an empirical result behind it.

Standing caveat — simulated, provisional, not settled. A single data-structure defect in Revision 1 of this simulator inverted four of five findings (99.6% wrong victims for EMA α=0.15). Revision 2 is the corrected rebuild: no Belady bound, no bandwidth/queueing model, whole-context residency assumed, synthetic workload, 2–3 seeds, no hardware validation. CANONICAL-NUMBERS.md §3.
0102030405060hit rate (%) — 32 GiB HBM budgetLFU · metadata reset on eviction · 15.56%15.56LFU · metadata persists · 49.71%49.71LFU+34.15 pts+34.15EMA α=0.01 · metadata reset on eviction · 31.54%31.54EMA α=0.01 · metadata persists · 56.39%56.39EMA α=0.01+24.85 pts+24.85EMA α=0.15 · metadata reset on eviction · 44.33%44.33EMA α=0.15 · metadata persists · 44.12%44.12EMA α=0.15−0.21 ptsExperiment: 32 GiB HBM over a 512 GiB tier, 421 GiB total footprint, Llama-3 70B GQA, 16-token blocks.
Metadata reset on eviction — the block is re-admitted as if newly createdMetadata persists — identity-keyed policy state survives the round trip
Table view
PolicyResetPersistsΔ
LFU15.56%49.71%+34.15 pts
EMA α=0.0131.54%56.39%+24.85 pts
EMA α=0.1544.33%44.12%−0.21 pts

Persistence is worth more than the choice of policy.The gain from letting identity-keyed policy state survive eviction — +34.15 points for LFU — is larger than the entire documented margin of the best scoring function over LRU at a fixed α, which is +2.5 to +5.8 points. It is also a protocol result: it is a claim about what state a controller chooses to preserve across an eviction event, not about the medium the evicted block lands on. It is testable with no CXL hardware.

+34.15 pts
LFU, 15.56% → 49.71%. The largest effect in the study.
+24.85 pts
EMA α=0.01, 31.54% → 56.39%. Persistence also removes the small-α collapse.
−0.21 pts
EMA α=0.15, 44.33% → 44.12%. A wash — this policy was not relying on the state it was losing.

Simulated (provisional) — CANONICAL-NUMBERS.md §3.3; formal_core.md §1.3