fix(agentic): pace snapshot warmup dispatches

[esmeetu]

Summary

• add a validated AIPERF_AGENTIC_SNAPSHOT_WARMUP_MIN_INTERVAL_S setting
• pace snapshot warmup dispatches in chronological order while preserving trace-derived offsets
• keep the feature disabled by default and document/test the behavior

Why

Agentic snapshot warmup can schedule many live trajectories at the same or nearly the same timestamp. At high concurrency this creates a synchronized connection burst, which can trigger transient TCP failures and abort the entire run as a terminal warmup failure even when the serving stack is healthy.

This change applies a configurable lower bound between initial warmup dispatches. It only extends the existing ramp when adjacent trace-derived offsets are closer than the configured interval.

Recommended configuration

For high-concurrency AgentX sweeps, start with:

export AIPERF_AGENTIC_SNAPSHOT_WARMUP_MIN_INTERVAL_S=0.05

The default is 0, which preserves the existing dispatch timing. Increase the value only when the warmup burst still causes connection pressure, since larger values extend warmup duration.

Validation

• uv run pytest tests/unit/common/test_environment.py tests/unit/timing/strategies/test_agentic_replay.py -q (100 passed)
• uv run pre-commit run (all hooks passed)

[github-actions]

Try out this PR

Quick install:

pip install --upgrade --force-reinstall git+https://github.com/ai-dynamo/aiperf.git@d0824e75212968bb49b179d9102df1637877c90e

Recommended with virtual environment (using uv):

uv venv --python 3.12 && source .venv/bin/activate
uv pip install --upgrade --force-reinstall git+https://github.com/ai-dynamo/aiperf.git@d0824e75212968bb49b179d9102df1637877c90e

---

Last updated for commit: d0824e7 • Browse code