HAL Meta-Scheduler — Public Demo

What this is: a safe, open demo that shows how an adaptive meta-layer can improve cluster workload placement using standard control signals (queue length, load coherence, task entropy). It includes a simulator, A/B comparison, charts, and a demo Helm chart.

What this is not: this demo does not contain any proprietary research formulae. Controllers here are simple (heuristic, PID, Bayesian) and meant for demonstration only.

Quick start

python -m venv .venv && .venv/Scripts/python -m pip install --upgrade pip
.venv/Scripts/pip install -r requirements.txt

# Run A/B comparison
.venv/Scripts/python simulate.py --steps 1500 --nodes 48 --arrival 1.2 --seed 42

# Plot metrics
.venv/Scripts/python plot_metrics.py

Outputs: out/metrics_baseline.csv, out/metrics_hal.csv, and a console summary.

Controllers

• heuristic: balances spread vs pack based on coherence (σ), queue drift (δ), and entropy (H).
• pid: PID loop on queue length + coherence term.
• bayes: simple stability posterior mapped to a single policy knob p ∈ [0,1].

Switch controller:

.venv/Scripts/python simulate.py --controller pid

Helm demo

This chart deploys a minimal metrics server that streams demo metrics (no scheduling hooks).

cd charts/halms
helm template halms .

Safe-by-design

• No proprietary math, no reverse path to any confidential models.
• Signals (σ, H, δ, Φ) are ordinary control/telemetry constructs.

License

Apache-2.0