Performance Targets #

Baseline Policy #

This repository does not hardcode arbitrary throughput numbers up front. The first measured baseline is stored in benchmarks/results/baseline.json, and CI compares new results against that file.

Workload Profiles #

CI-safe suite #

• conn_churn_seq
• producer_contention

Command:

zig build bench -- --output benchmarks/results/latest.json

Scheduled or manual suite #

• conn_churn_parallel
• short_long_mix
• idle_burst_idle
• nested_submit
• blocking_lane_sim
• oversubscription

Command:

zig build scale-bench -- --output benchmarks/results/scale-current.json

Regression Gates #

CPU gate #

• compare-bench fails if throughput regresses by more than 15% against the stored baseline for a matched row.

Memory gate #

• compare-bench fails if rss_current_bytes grows by more than 15% against the stored baseline for a matched row.

Manual review expectations #

• Any sustained RSS increase under conn_churn_seq or conn_churn_parallel requires documenting the reason in docs/GAP_ANALYSIS.md.
• Any scheduler-default change requires refreshing docs/SCHEDULER_DECISION.md.

Reproduction Commands #

make bench-save
make bench-compare
make scale-bench

Threshold Updates #

Update thresholds only after:

1. capturing new baseline JSON,
2. recording the reason in a merge request,
3. updating docs/SCHEDULER_DECISION.md if scheduler-sensitive metrics changed materially,
4. updating docs/GAP_ANALYSIS.md evidence links.

Open Gaps #

• total_alloc_count counts scratch allocation events (incremented in src/memory/scratch.zig); per-size allocation histograms remain future work (RISK-06, #85).
• CPU time falls back to wall time when the OS sample returns zero or is unavailable.