Skip to main content

Phase-1 Perf Playbook (First 5 Steps)

This playbook automates the first five optimization steps for plugin-test stress.

What it does​

  1. Freeze target metadata (target_throughput_pps=1000000).
  2. Run baseline/concurrency matrix:
    • p12 t2 b256
    • p6 t4 b256
    • p4 t6 b256
  3. Extract bottleneck diagnostics from worst_sample and worst_topk.
  4. Select best concurrency shape by throughput.
  5. Run batch neighborhood on the best shape (b128/b256/b512/b1024).

Script​

  • scripts/phase1_perf_playbook.py

Quick smoke (fast)​

python -u scripts/phase1_perf_playbook.py --total 20000 --repeats 1 --chain-mode e2e

Full run (heavy)​

python -u scripts/phase1_perf_playbook.py --total 1000000 --repeats 3 --chain-mode e2e

Optional custom config​

python -u scripts/phase1_perf_playbook.py --config data/benchmarks/Config_e2e_flush1024.json --total 1000000 --repeats 2 --chain-mode e2e

Output​

All artifacts are written under data/benchmarks/phase1/:

  • target_<timestamp>.json
  • bottleneck_<timestamp>.json
  • phase1_summary_<timestamp>.json
  • raw benchmark files per run

The summary includes:

  • best concurrency candidate
  • best batch candidate
  • throughput baseline stats
  • bottleneck decomposition (tracked vs untracked latency)