A filter is useful only if its absence claim is trustworthy. Start by counting messages, then keep the invariant that every holder is covered. Finally inspect how a full filter reclaims an entry.
Lesson | After it you can... |
|---|---|
Count probes and responses, separately from latency. | |
State the inclusive-filter invariant and where false positives come from. | |
Match acknowledgements to a fixed expected set. | |
Free an entry only after clean holders ack or dirty data reaches memory. |
Prerequisite: Hybrids and Snoop Filters, which introduced the inclusive filter F, false positives and negatives, and back-invalidation, plus the Part 3 upgrade and clean-eviction lessons. Scope: the traces keep two cores, so a requester has at most one peer; larger fanout appears only as arithmetic, and H's back-invalidation of both cores gives the two-target example.
No. Unknown requires a conservative coverage path or retained tracking. Known empty is a stronger fact justified by completed protocol work.