GPU Lab

GPU Lab is an observability environment for Apple Silicon inference workloads. Phase 1 establishes a macOS host-metrics baseline for comparison with later GPU and application telemetry.

Phase 1 architecture

macOS node_exporter (:9100)
        |
        v
Grafana Alloy (:12345)
        | remote write
        v
Prometheus (:9090)
        |
        v
Grafana (:3000)

node_exporter runs natively so it can access Darwin host metrics. Alloy, Prometheus, and Grafana run in Docker Compose. Container ports bind to 127.0.0.1 and are not exposed to the local network.

Prerequisites

• Apple Silicon Mac
• Docker Desktop
• Homebrew
• Native node_exporter

Install and start the exporter:

brew install node_exporter
brew services start node_exporter
curl -fsS http://localhost:9100/metrics >/dev/null

Start the observability stack:

docker compose -f observability/compose.yaml up -d

Open Grafana at http://localhost:3000. A new Grafana data volume uses Grafana's initial administrator setup flow.

Repository layout

• 00-baseline/: sanitized hardware and toolchain baseline
• observability/alloy/: host scrape and remote-write configuration
• observability/prometheus/: Prometheus configuration
• observability/grafana/: provisioned dashboards, metric inventory, and dashboard documentation

See observability/grafana/README.md for dashboard details and validation commands.

Scope

Phase 1 covers host identity, CPU, load, memory, storage, filesystems, networking, and telemetry health. GPU, Metal, MLX, llama.cpp, and inference-request metrics are reserved for later phases.