Forwarding vs Memory Sourcing: Expanded Case Study

Expanded case study

Expanded Case Study for Forwarding vs Memory Sourcing explains how to reason from coherency invariant to measurable engineering decision.

Review a realistic incident end-to-end: symptom capture, mechanism isolation, corrective action, and long-tail prevention.

Evidence pack

STAFF REVIEW MEMO — Requests, Snoops, and Dataflow / Forwarding vs Memory Sourcing

1) Symptom
   - Tracked metric: forward-hit ratio and remote-supply tail latency
   - Workload and mode: <explicitly named>
   - First failing evidence: <artifact ID and timestamp>

2) Mechanism hypothesis
   - Candidate mechanism: Supplying data from an owner cache cuts memory latency but can amplify network hotspots and starvation if arbitration is weak.
   - Alternative explanations: ordering, backpressure, metadata staleness, or software misuse
   - Missing evidence required for decision: <list>

3) Action plan
   - Smallest reversible fix: <RTL, firmware, policy, or tooling>
   - Expected movement: <numeric trend expectation>
   - Risk of regression: performance, power, compatibility, or timing

4) Signoff gates
   - Primary artifact: source-selection report + hotspot map
   - Owners: architecture, soc-integration, performance
   - Decision: fix now, bounded waiver, or escalate

Cache coherency deep dive

Cache coherence is a correctness contract across caches, interconnect, and software ordering.

Concept diagram

requester -> coherence fabric -> owner or memory -> state update

Metric graph

traffic mix across request, snoop, response, data

Metrics and artifacts to collect

• coherence latency

• invalidation rate

• retry rate

• stale-read incidents

Mini case study

Anchor debug to first stale read and the exact line state transition.

Debug branches

• Track ownership

• Track ordering

• Track evidence

Senior review question

Ask: what is the first line state transition that deviates, and which ordering rule does it break?

Key takeaways

• Tie every coherency claim to one cache line, one transaction identity, and one measurable counter.

• Keep proof artifacts from simulation and silicon replay aligned by address, state, and ordering event.

Common pitfalls

• Chasing bandwidth regressions without checking false sharing and line bouncing first.

• Assuming coherence correctness implies memory consistency correctness.

• Declaring closure without litmus, stress, and post-silicon replay evidence.