- M · Modified: the only copy, and newer than DRAM
- E · Exclusive: the only copy, matches DRAM
- S · Shared: a clean copy that other cores may also hold
- I · Invalid: stale, so the line must be re-fetched
- INV: an invalidation message
To keep the picture readable, this model:
- gives each core one cache line and no L2/L3 caches.
- always re-fetches an invalidated line from DRAM. Real CPUs usually forward a dirty line core-to-core through the shared L3 in about 40–80 ns. That is faster than DRAM, but still 40–80× slower than an L1 hit.
- lets a core do one increment each time it owns the line. Real cores often fit several in, so measured false-sharing slowdowns are usually 2–20× rather than hundreds, and are worse across sockets.