Cache Coherence
The single-writer invariant, MESI and MOESI state machines, and why snooping gives way to directories as core counts grow.
The single-writer invariant, MESI and MOESI state machines, and why snooping gives way to directories as core counts grow.
The tag-index-offset split, direct-mapped versus set-associative lookup, line size and index translation, and the numbers that decide hit rate and hit time.
The capacity-latency ladder from registers to disk, the AMAT identity that prices it, and how locality turns a hundred-nanosecond miss into a one-nanosecond average.
The three Cs plus coherence, non-blocking caches and MSHRs, hardware and software prefetch, and cache blocking as the software answer to a working set that will not fit.
Private L1 and L2 behind a shared L3: what inclusive, exclusive and non-inclusive caches cost in effective capacity, snoop filtering and back-invalidation traffic.
How a set chooses the line to evict and how a write reaches the next level: LRU and its approximations, write-back versus write-through, and the traffic each choice costs.