Figure: HBM hit rate against HBM budget for four eviction policies, from experiment e1 of a trace-driven KV tiering simulation, two seeds each, no persistent metadata and no transfer-compute overlap. Mean hit rates in per cent, with the two seed values in brackets. At 8 GiB: LRU 8.52 (8.04, 9.01), LFU 8.76 (7.57, 9.95), EMA alpha 0.15 12.35 (12.70, 12.01), EMA alpha 0.01 12.12 (11.08, 13.17). At 16 GiB: LRU 17.87 (17.16, 18.57), LFU 12.70 (11.17, 14.22), EMA alpha 0.15 22.58 (23.28, 21.88), EMA alpha 0.01 18.70 (16.54, 20.86). At 32 GiB: LRU 38.34 (36.85, 39.83), LFU 16.09 (15.56, 16.62), EMA alpha 0.15 44.18 (44.33, 44.04), EMA alpha 0.01 35.91 (31.54, 40.29). At 64 GiB: LRU 67.12 (64.92, 69.31), LFU 23.00 (24.49, 21.50), EMA alpha 0.15 69.65 (67.99, 71.31), EMA alpha 0.01 66.83 (64.27, 69.40). At 128 GiB: LRU 88.01 (87.74, 88.28), LFU 32.70 (32.49, 32.91), EMA alpha 0.15 88.14 (87.86, 88.41), EMA alpha 0.01 91.59 (91.16, 92.02). The honest reading: the policies are close, LFU collapses, and the EMA lead over LRU is +3.8, +4.7, +5.8, +2.5 and +0.1 points across the five budgets — a few points, not tens of points. The seed spread is often as large as the gap between policies. These are simulated provisional results, not measurements.
Figure 11.1 · Simulated (provisional)
Hit rate against HBM budget — the policies are closer than the story wanted
Experiment e1: Llama-3 70B GQA, 16-token blocks, 200 sessions, Zipf(1.1) return frequency, 2500 turns, 421 GiB total footprint, HBM tier over a 512 GiB CXL tier. No persistent metadata, no transfer/compute overlap, two seeds — both plotted, because at this scale the seed spread is often as wide as the gap between two policies.
EMA α=0.15 leads at four of the five budgets and EMA α=0.01 leads at 128 GiB. The margin over LRU is +2.5 to +5.8 points, not the +25 points version 4.0 claimed — that figure is retracted. LFU is the one clear separation on the chart, and it separates downwards: it reaches 32.70% at 128 GiB where every other policy is near 90%. That is classical LFU aging once decode growth is modelled. At 128 GiB the LRU and EMA α=0.15 curves are 0.13 points apart, which is well inside the seed spread, so the chart does not distinguish them.
Show the data table (mean, and both seeds)
| HBM budget | LRU | LFU | EMA α=0.15 | EMA α=0.01 |
|---|---|---|---|---|
| 8 GiB | 8.52 (8.04 / 9.01) | 8.76 (7.57 / 9.95) | 12.35 (12.70 / 12.01) | 12.12 (11.08 / 13.17) |
| 16 GiB | 17.87 (17.16 / 18.57) | 12.70 (11.17 / 14.22) | 22.58 (23.28 / 21.88) | 18.70 (16.54 / 20.86) |
| 32 GiB | 38.34 (36.85 / 39.83) | 16.09 (15.56 / 16.62) | 44.18 (44.33 / 44.04) | 35.91 (31.54 / 40.29) |
| 64 GiB | 67.12 (64.92 / 69.31) | 23.00 (24.49 / 21.50) | 69.65 (67.99 / 71.31) | 66.83 (64.27 / 69.40) |
| 128 GiB | 88.01 (87.74 / 88.28) | 32.70 (32.49 / 32.91) | 88.14 (87.86 / 88.41) | 91.59 (91.16 / 92.02) |
Simulated (provisional) — results_v2.json experiment e1; see Appendix N for the standing caveat