Attempts / tuning notes

Experiment 3 — first design did not show the eviction contrast

The first cut of the calcification experiment (phase B = write 60,000 unique 8 KB items once, then sleep) did NOT reproduce the intended “evictions drop under automove” story, for two reasons:

1. 60,000 × 8 KB ≈ 480 MB into a 64 MB cache. The large workload can’t fit no matter how the pages are allocated, so both automove modes evict heavily — the eviction count is dominated by capacity, not by the allocator.
2. Global evictions actually went up under automove=2, because reassigning a page to the large class first evicts the ~5,461 stale small items living on it. Read as a global number it looks like automove made things worse.

Captured from that run (memcached 1.6.45, for the record):

automove=0  phase B: large pages 1,  large evicted 59,882, global evict since A  59,882
automove=2  phase B: large pages 63, large evicted 52,566, global evict since A 391,149

The page migration (1 → 63) was real and correct, but the eviction number told the wrong story.

What fixed it (the shipped version)

• Size the large working set to fit after a full rebalance (5,000 × 8 KB ≈ 44 MB, ~44 of 64 pages), so once the large class gets pages the set goes resident.
• Rewrite that working set repeatedly (15×, 1 s apart) to give the automove thread both the eviction pressure to react to and the wall-clock time to move pages — a single burst stalls the rebalance partway.
• Report large-class evictions (the new workload’s own thrashing), not global evictions (which include the one-time reclamation of the dead small items).

Result: 74,882 large-class evictions frozen at 1 page vs 1,976 once rebalanced to 44 pages. That is the version in ../exp3_calcification.csv.