Reports and metrics

Reports & Metrics for CHI Topology Basics focuses on request retry rate, directory occupancy, p99 fabric latency. The goal is to connect the observable symptom to protocol mechanism, ownership, and regression risk.

The job of a report is to turn request retry rate, directory occupancy, p99 fabric latency into a decision. A single average number is almost never enough; you need the distribution, the traffic class breakdown, and a clear gap between legal maximum and product target.

Metric movement

METRIC GRAPH — request retry rate, directory occupancy, p99 fabric latency

throughput / success
  ^
  |                         target
  |                       - - - - - - -
  |                  o after bounded fix
  |              o
  |         o baseline
  |    o failing run
  +--------------------------------------> experiment
    config A     isolated root cause     accepted change

Readout:
  - compare identical payload, clock, reset, traffic seed, and firmware setup
  - separate headline bandwidth from useful payload bandwidth
  - explain why the protocol mechanism moved the metric

Latency distribution

LATENCY HISTOGRAM — CHI Topology Basics

count
  |               ███
  |             ███████
  |          █████████████
  |        █████████████████        <- long tail = the real complaint
  |      ████████████████████████████
  +------------------------------------> latency
   p50      p90    p95       p99  (watch p99, not the average)

Average hides the tail; product pain lives at p95/p99.

• Track request retry rate, directory occupancy, p99 fabric latency by traffic class, payload size, and clock/reset mode.

• Report p50/p95/p99 latency when user-visible stalls matter.

• Include legal maximums and product targets; they are not the same thing.

• Always store the metric next to the artifact that produced it.

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.

How to read the numbers

request retry rate, directory occupancy, p99 fabric latency must be split by traffic class, payload size, and reset mode.