Software and programmer view

False sharing, fences, and cache maintenance ops dominate software-visible behavior.

What programmers feel

• Timeouts with healthy-looking hardware counters

• Data corruption without obvious ECC/CRC

• Ordering surprises under multi-threaded drivers

• Performance cliffs when payload size changes

API / driver implications

• Descriptor alignment and cache line sharing

• Fence/barrier placement around DMA

• IRQ type (level vs edge) and clear sequence

• Memory-mapped register access ordering

Compiler and runtime interaction

• Volatile and barrier semantics for device memory

• Struct padding affecting burst efficiency

• Batching policy in userspace drivers

Software-side mitigations

• Pad structures to cache lines

• Pin buffers and use coherent DMA where required

• Expose hardware counters to software profilers

• Document legal outstanding depth and ordering

SOFTWARE EXAMPLE — Coherence Fabric Debug

// Bad: assumes ordering across unrelated IDs without fence
dma_start(ch0); dma_start(ch1); cpu_read(result); // may see stale

// Better: document which completions are ordered and insert barrier
dma_start(ch0); wait_completion(ch0); cpu_read(result);

Layer the driver touches

LAYER RESPONSIBILITY — Coherence Fabric Debug

layer          owns                         common failure
-----------    --------------------------   -----------------------
software       intent, ordering needs       wrong assumption
transaction    id/addr/len/attributes       ordering / outstanding
link/channel   handshake, credits, retry    backpressure / deadlock
physical       clock/reset/lanes/PHY        timing / training / SI
observability  waveform/log/counter         missing evidence

Protocol deep dive

Coherence extends memory transactions with snoop and state — traffic multiplies when software shares cache lines.

Concept diagram

COHERENCE TRAFFIC FLOW

RN issues coherent read
   -> HN looks up directory
   -> snoops to sharers
   -> data + state update returned

False sharing: different variables, same cache line -> coherence storm.

Metric graph

COHERENCY TRAFFIC STACK

data fetch        ████████
snoop responses   ██████████████
writebacks        ██████
maintenance ops   ████

High snoop stack with good IPC -> suspect line sharing before faster NoC.

Metrics and artifacts to collect

• snoop rate

• intervention latency

• coherency transaction mix

• false sharing indicators

Mini case study

Benchmark IPC looked fine but system power spiked: per-core counters were on one cache line. Padding counters fixed coherency traffic without any NoC change.

Debug branches

• If snoop latency high, check home node placement and directory policy.

• If ordering bug, run litmus sequences before microarch changes.

• If traffic storm, profile cache line sharing in software layout.

Senior review question

Ask: what is the first transaction that deviates, and which spec rule does it test?

Key takeaways

• Connect every protocol claim to a transaction identity and measurable metric.

• Store the artifact (waveform, log, counter) next to every signoff decision.

Common pitfalls

• Debugging timeouts without finding the first bad transaction.

• Quoting peak bus width without payload efficiency and retry overhead.

• Treating VIP compliance as a substitute for system integration replay.

Principal review addendum

Re-read Coherence Fabric Debug against one concrete product workload, not a synthetic directed test.

debug requires correlating transaction IDs, cache-line addresses, state transitions, and fabric backpressure.