Knowing the mechanisms of Parts 1 to 12 is not the same as explaining them under pressure or deciding with them on a real program. This part turns the course into practice: a four-step answer framework (Assumption, Mechanism, Trade-off, Evidence) and a whiteboard routine for drawing a transaction, how a senior engineer reviews a protocol change and writes a design memo, how coherence risk is tracked and escalated before tape-out, and a question bank graded on a 0 to 4 rubric. Nothing here adds new protocol content: every answer links back to the lesson that teaches it.
Lesson | After it you can... |
|---|---|
Structure any coherence answer as assumption, mechanism, trade-off, evidence. | |
Draw a correct transaction with its state/value table in minutes. | |
Explain who wins, who waits and why the data stays safe. | |
Grade an answer from 0 to 4. | |
Find the new races and costs a protocol change brings. | |
Run a coherence design review with evidence for each item. | |
Write a short memo that a team can decide from. | |
Answer senior trade-off questions with the deciding conditions. | |
Rank coherence risks on a chip program. | |
Choose between a fix, a workaround and an erratum. | |
Run a late hang investigation to a decision. | |
State the evidence a coherent design needs before tape-out. | |
Answer invariant, state and snoop-versus-directory questions. | |
Answer transaction, race and controller questions. | |
Answer memory-model, fence and atomic questions. | |
Answer checking, formal, coverage and triage questions. | |
Answer scaling, hierarchy, protocol and integration questions. |
Attempt each question on paper before reading its answer, then grade yourself with the rubric. A strong answer always says which model it assumes: the same question has different correct answers under MSI and MOESI, with acks at the home or at the requester, and under TSO or a weak memory model.
A 2 explains the mechanism: a request is in flight, so the cache is neither in its old nor its new stable state and must stall or handle certain messages. A 4 also states the assumption (an unordered network, acks collected at the home), names a concrete race the transient state resolves (an Inv arriving in SM^D), gives the trade-off (more states to verify), and says how it would be checked (state-by-event coverage, a model checker). The mechanism alone is not enough for a 4.