Locality experiments and defensible performance claims
This required 18–26-hour package replaces the former 2–3-hour matrix traversal lab: lessons and prediction 5–7, experiments/assembly/evidence 8–12, assessment and delayed transfer 5–7. Original ISA, datapath, pipeline, virtual-memory and device concepts remain required; See the reconciled semester workload.
Sequence
- Address models.
- Benchmark contracts.
- Variability and causal limits.
- Prediction practice, runnable lab, and PE1–PE6 assessment.
Carry forward Module 2’s allocation and size reasoning. The supplied harness uses bounded dimensions; do not remove those bounds without revisiting arithmetic and allocation checks. A passing measurement report may conclude that the evidence is inconclusive.
Selective resources
Reuse Computer Organization and Design readings for caches and performance. MIT 6.004 Computation Structures provides optional performance/cache lectures and worksheets. Intel’s manuals are a target-specific reference, not a requirement for ARM learners. Linux perf documentation explains counter-access constraints; counters are optional here. Sources checked on 2026-10-23. No purchases or full-book conversion required.
The original lab measures repeated warm-cache traversals using process CPU time. It does not measure network latency, energy, cache-miss counts, cold-start behavior or concurrent throughput. The separate required specialized-platform benchmark remains part of the semester strand. Independent review and learner timing pilots remain pending.