enable-xpu-monitoring

enable-xpu-monitoring

Deploy a single-node monitoring stack on an Intel GPU host from intel/xe-full-stack-monitoring. The stack is pure Docker Compose — no CLI, no Python package, no tarball. That repo's AGENT.md is the source of truth; this skill gives the short path and defers detail to it.

What gets deployed

Container	Port	Purpose
xpumd (XPU Manager v2.0)	9471	GPU hardware metrics via Level Zero
node-exporter	9100	CPU, RAM, disk, network
intel-xpu-prometheus	9090	Scrapes xpumd, node, and vLLM
intel-xpu-grafana	3000	4 dashboards: Quickview, GPU, vLLM, System

All listeners bind to 127.0.0.1. Grafana has no login (anonymous, read-only Viewer).

Step 1 — Preflight

ls /dev/dri/renderD* >/dev/null 2>&1 && echo "[OK] GPU render nodes" || echo "[FAIL] no render nodes"
docker info >/dev/null 2>&1 && echo "[OK] docker" || echo "[FAIL] docker not reachable"
docker compose version >/dev/null 2>&1 && echo "[OK] compose" || echo "[FAIL] compose plugin missing"
for p in 9471 9100 9090 3000; do
  (ss -ltn 2>/dev/null || netstat -ltn) | grep -q ":$p " && echo "port $p IN USE" || echo "port $p free"
done

Fix any [FAIL]. Ports held by a previous run of this same stack are fine. If 9471, 9100 or 9090 is held by anything else, it must be freed; those ports cannot be changed.

If port 3000 is held by another app, do not stop that app. Move Grafana to a free port instead:

GRAFANA_PORT=3000
if (ss -ltn 2>/dev/null || netstat -ltn) | grep -q ":3000 " \
   && ! docker ps --format '{{.Names}}' | grep -qx intel-xpu-grafana; then
  for p in $(seq 3001 3099); do
    (ss -ltn 2>/dev/null || netstat -ltn) | grep -q ":$p " || { GRAFANA_PORT=$p; break; }
  done
fi
echo "Grafana port: $GRAFANA_PORT"

If GRAFANA_PORT is not 3000, tell the user clearly: