Parallel Computing

Parallel Computing

Use this skill to convert parallel performance work into reproducible scaling decisions.

Workflow

1. Define scaling objective and constraints.

• Capture workload shape, data size, and latency/throughput targets.
• Define hardware assumptions (core count, SMT policy, NUMA context).

2. Choose parallel model and partitioning.

• Select task/data/pipeline parallelism intentionally.
• Set chunk size and scheduling strategy to minimize overhead and imbalance.
• Define shared-state boundaries before coding.

3. Diagnose bottlenecks.

• Check lock contention, false sharing, synchronization frequency, and memory bandwidth pressure.
• Separate algorithmic limits from runtime/scheduler overhead.

4. Validate scaling behavior.

• Compare baseline vs current throughput by thread count.
• Evaluate parallel efficiency and regressions at each thread level.
• Treat regressions above threshold as blockers.

5. Deliver implementation handoff.

• Include tuning deltas, tradeoffs, and reproducible benchmark commands.
• Provide clear patch plan for runtime/algorithm changes.

Commands

python3 scripts/compare_parallel_scaling.py \
  --baseline <baseline.json> \
  --current <current.json> \
  --regression-threshold-pct 5 \
  --efficiency-drop-threshold-pct 10

Treat non-zero exits as blocker regressions.

Output Contract

Return:

1. Scaling Context: workload and hardware assumptions.
2. Findings: thread-level throughput/speedup/efficiency deltas.
3. Optimization Plan: concrete runtime/algorithm changes.
4. Verification: benchmark commands and thresholds.
5. Residual Risks: unresolved contention or scaling ceilings.

References

• references/workflow.md: detailed parallel optimization sequence.
• references/scaling-playbook.md: common bottlenecks and remedies.
• references/signoff-template.md: concise scaling sign-off format.

Execution Rules

• Compare like-for-like workloads and environments only.
• Report both speedup and efficiency, not throughput alone.
• Flag thread-level regressions above thresholds as blockers.
• Avoid overfitting to one thread count; evaluate full scaling curve.