A coherence optimization needs a named cost and a preserved result. This part adds workload and storage accounting to the reviewed transactions without changing their permission or data obligations.
Lesson | After it you can... |
|---|---|
Separate logical data from physical permission units. | |
Count transfers and reject neighboring-value loss. | |
Check alignment, stride and footprint. | |
Design an address-level counterfactual experiment. | |
Extend the 4 / 66-bit baseline to large N. | |
Handle loss of precision conservatively. | |
Separate address capacity, set conflicts and pending work. | |
Preserve coverage across representation transitions. | |
Count lock traffic per spin and per handoff, from test-and-set to queue locks. | |
Compare a simple and a hashed home map and the stride that defeats each. | |
Place a page and its home, then cost local, remote and remote-dirty misses. | |
Choose a candidate with measurable rejection criteria. |
Suggested order: sharing and layout, metadata, then placement. Keep message counts, bytes, cycles and storage bits separate throughout. Every numerical example in this part is illustrative.
A changed required result, lost data, premature permission or broken synchronization contract. Performance improvements are compared only after correctness and useful work are preserved.