GNMON-001 M1: Metrikpfad auf der Workstation implementiert

llm-server/monitoring/ neu: amdgpu-textfile-collector.py (sysfs-only,
atomar, APU-tauglich), Rollen gnmon_collector (node_exporter v1.12.1,
SHA256SUMS-verifiziert, 10s-Timer) und gnmon_victoria_metrics (rootless
Quadlet, v1.154.1 digest-gepinnt, Scrape 10s, Retention 60d), Workstation-
Inventory. Verifikation: PromQL liefert gn_gpu_edge_temp_celsius von der
Laptop-APU; Idempotenz changed=0. GNMON-001 v2.0 in die Plattform-Struktur
übernommen, GNMON-000-Stand inkl. Alerting-Planlücke aktualisiert.

Generated by Mistral Vibe.
Co-Authored-By: Mistral Vibe <vibe@mistral.ai>
This commit is contained in:
2026-11-04 05:51:17 +01:00
parent 04afca49c9
commit aa8de57c62
18 changed files with 737 additions and 94 deletions
+12
View File
@@ -0,0 +1,12 @@
[defaults]
home = .ansible
host_key_checking = True
interpreter_python = auto_silent
local_tmp = .ansible/tmp
retry_files_enabled = False
roles_path = roles:../../../shared/ansible/roles
timeout = 15
[ssh_connection]
pipelining = True
ssh_args = -o ControlMaster=auto -o ControlPersist=60s
@@ -0,0 +1,10 @@
# GNMON-001 v2.0 workstation (development) inventory group vars.
# Role defaults carry the M1 development values; this file only pins the
# deployment-specific bits.
# VictoriaMetrics image: by tag during development. The digest of the
# verified pull is recorded by the role and pinned here before M5.
# Tag choice: v1.154.1 is the newest patch release of the second-newest
# line (the v1.136 LTS line has no bare OSS tags in the Docker registry).
gnmon_victoria_metrics_image_tag: v1.154.1
gnmon_victoria_metrics_image_digest: null
@@ -0,0 +1,10 @@
# GNMON-001 section 2: workstation inventory for the development phase.
# The monitoring suite is developed and first tested on the operator's
# CachyOS laptop (rootless, user-scope systemd) before the server rollout
# (ghostnet-ki) in milestone M5.
all:
hosts:
gnmon-workstation:
ansible_connection: local
ansible_user: ghost
# No become: everything in user scope during development.
@@ -0,0 +1,28 @@
---
# GNMON-001 v2.0: host monitoring deployment.
# M1 (workstation): metrics path (collector + node_exporter + VictoriaMetrics).
# Tags: `metrics` for the M1 path; log path (M2) and dashboard (M3) follow.
- name: GNMON metrics path
hosts: all
connection: local
roles:
- role: gnmon_collector
tags: [collector]
- role: gnmon_victoria_metrics
tags: [metrics]
post_tasks:
- name: Verify the GPU metrics are ingested
ansible.builtin.uri:
url: "http://127.0.0.1:{{ gnmon_victoria_metrics_port }}/api/v1/query?query=gn_gpu_edge_temp_celsius"
status_code: 200
register: gnmon_query_result
retries: 15
delay: 4
until: >-
gnmon_query_result.json is defined
and gnmon_query_result.json.data.result | length > 0
tags: [metrics]
- name: Show the ingested GPU temperature sample
ansible.builtin.debug:
msg: "gn_gpu_edge_temp_celsius = {{ gnmon_query_result.json.data.result[0].value[1] }} °C"
tags: [metrics]
@@ -0,0 +1,22 @@
# GNMON-001 v2.0 section 3: node_exporter + AMD-GPU textfile collector.
#
# M1 scope: user-scope deployment for the workstation development phase
# (CachyOS laptop). The system scope for ghostnet-ki (metrics account,
# host-systemd, journald groups) is added with milestone M5.
gnmon_collector_scope: user
gnmon_collector_user_home: "{{ lookup('ansible.builtin.env', 'HOME') }}"
gnmon_collector_textfile_dir: "{{ gnmon_collector_user_home }}/.local/state/gnmon/textfile"
gnmon_collector_binary_dir: "{{ gnmon_collector_user_home }}/.local/bin"
gnmon_collector_script_dir: "{{ gnmon_collector_user_home }}/.local/share/gnmon"
gnmon_collector_cache_dir: "{{ gnmon_collector_user_home }}/.cache/gnmon"
# Collection cadence (GNMON-001 section 3: host-systemd-Timer, 10 s).
gnmon_collector_interval_seconds: 10
# node_exporter: pinned upstream release, verified against the release
# SHA256SUMS file at download time.
gnmon_collector_node_exporter_version: "1.12.1"
gnmon_collector_node_exporter_url: "https://github.com/prometheus/node_exporter/releases/download"
gnmon_collector_node_exporter_port: 9100
gnmon_collector_node_exporter_listen: "127.0.0.1:{{ gnmon_collector_node_exporter_port }}"
@@ -0,0 +1,14 @@
# GNMON-001: collector role handlers (user scope, M1).
- name: Restart collector timer
ansible.builtin.systemd:
name: gnmon-amdgpu-collector.timer
scope: user
state: restarted
daemon_reload: true
- name: Restart node_exporter
ansible.builtin.systemd:
name: gnmon-node-exporter.service
scope: user
state: restarted
daemon_reload: true
@@ -0,0 +1,106 @@
# GNMON-001 v2.0 section 3: AMD-GPU textfile collector + node_exporter.
# M1: user-scope deployment on the workstation inventory.
- name: Validate collector declarations
ansible.builtin.assert:
that:
- gnmon_collector_scope == 'user'
- gnmon_collector_user_home is match('^/')
- gnmon_collector_interval_seconds | int >= 5
- gnmon_collector_node_exporter_version is match('^[0-9]+\\.[0-9]+\\.[0-9]+$')
fail_msg: >-
gnmon_collector only implements scope 'user' in milestone M1 and
requires absolute paths and a pinned node_exporter release.
- name: Create collector directories
ansible.builtin.file:
path: "{{ item }}"
state: directory
mode: "0755"
loop:
- "{{ gnmon_collector_textfile_dir }}"
- "{{ gnmon_collector_binary_dir }}"
- "{{ gnmon_collector_script_dir }}"
- "{{ gnmon_collector_cache_dir }}"
- name: Install the AMD GPU textfile collector
ansible.builtin.copy:
src: "{{ playbook_dir }}/../../scripts/amdgpu-textfile-collector.py"
dest: "{{ gnmon_collector_script_dir }}/amdgpu-textfile-collector.py"
mode: "0755"
notify: Restart collector timer
- name: Fetch the node_exporter release checksums
ansible.builtin.get_url:
url: "{{ gnmon_collector_node_exporter_url }}/v{{ gnmon_collector_node_exporter_version }}/sha256sums.txt"
dest: "{{ gnmon_collector_cache_dir }}/node_exporter-{{ gnmon_collector_node_exporter_version }}.SHA256SUMS"
mode: "0644"
- name: Record the pinned node_exporter tarball checksum
ansible.builtin.set_fact:
gnmon_collector_tarball_name: "node_exporter-{{ gnmon_collector_node_exporter_version }}.linux-amd64.tar.gz"
gnmon_collector_tarball_checksum: "{{ lookup('ansible.builtin.file',
gnmon_collector_cache_dir ~ '/node_exporter-' ~ gnmon_collector_node_exporter_version ~ '.SHA256SUMS')
| regex_findall('^([0-9a-f]{64})\\s+node_exporter-' ~ gnmon_collector_node_exporter_version ~ '\\.linux-amd64\\.tar\\.gz$', multiline=True) | first }}"
- name: Download node_exporter {{ gnmon_collector_node_exporter_version }}
ansible.builtin.get_url:
url: "{{ gnmon_collector_node_exporter_url }}/v{{ gnmon_collector_node_exporter_version }}/{{ gnmon_collector_tarball_name }}"
dest: "{{ gnmon_collector_cache_dir }}/{{ gnmon_collector_tarball_name }}"
checksum: "sha256:{{ gnmon_collector_tarball_checksum }}"
mode: "0644"
- name: Extract node_exporter
ansible.builtin.unarchive:
src: "{{ gnmon_collector_cache_dir }}/{{ gnmon_collector_tarball_name }}"
dest: "{{ gnmon_collector_cache_dir }}"
remote_src: true
creates: "{{ gnmon_collector_cache_dir }}/node_exporter-{{ gnmon_collector_node_exporter_version }}.linux-amd64"
- name: Install the node_exporter binary
ansible.builtin.copy:
src: "{{ gnmon_collector_cache_dir }}/node_exporter-{{ gnmon_collector_node_exporter_version }}.linux-amd64/node_exporter"
dest: "{{ gnmon_collector_binary_dir }}/node_exporter"
mode: "0755"
notify: Restart node_exporter
- name: Install the collector service unit
ansible.builtin.template:
src: gnmon-amdgpu-collector.service.j2
dest: "{{ gnmon_collector_user_home }}/.config/systemd/user/gnmon-amdgpu-collector.service"
mode: "0644"
notify: Restart collector timer
- name: Install the collector timer unit
ansible.builtin.template:
src: gnmon-amdgpu-collector.timer.j2
dest: "{{ gnmon_collector_user_home }}/.config/systemd/user/gnmon-amdgpu-collector.timer"
mode: "0644"
notify: Restart collector timer
- name: Install the node_exporter service unit
ansible.builtin.template:
src: gnmon-node-exporter.service.j2
dest: "{{ gnmon_collector_user_home }}/.config/systemd/user/gnmon-node-exporter.service"
mode: "0644"
notify: Restart node_exporter
- name: Run the collector once to seed the textfile
ansible.builtin.command:
cmd: "/usr/bin/python3 {{ gnmon_collector_script_dir }}/amdgpu-textfile-collector.py --textfile-dir {{ gnmon_collector_textfile_dir }}"
creates: "{{ gnmon_collector_textfile_dir }}/amdgpu.prom"
- name: Enable and start the collector timer
ansible.builtin.systemd:
name: gnmon-amdgpu-collector.timer
scope: user
enabled: true
state: started
daemon_reload: true
- name: Enable and start node_exporter
ansible.builtin.systemd:
name: gnmon-node-exporter.service
scope: user
enabled: true
state: started
daemon_reload: true
@@ -0,0 +1,9 @@
[Unit]
Description=GNMON AMD GPU textfile collector (GNMON-001)
Documentation=https://github.com/ghostnet/infrastructure
[Service]
Type=oneshot
ExecStart=/usr/bin/python3 {{ gnmon_collector_script_dir }}/amdgpu-textfile-collector.py --textfile-dir {{ gnmon_collector_textfile_dir }}
# A failing sample must surface in the journal, not silently skip cycles.
Nice=5
@@ -0,0 +1,10 @@
[Unit]
Description=GNMON AMD GPU textfile collector timer (GNMON-001)
[Timer]
OnBootSec=15s
OnUnitActiveSec={{ gnmon_collector_interval_seconds }}s
AccuracySec=1s
[Install]
WantedBy=timers.target
@@ -0,0 +1,11 @@
[Unit]
Description=GNMON node_exporter (GNMON-001)
Documentation=https://github.com/prometheus/node_exporter
[Service]
ExecStart={{ gnmon_collector_binary_dir }}/node_exporter --collector.textfile.directory={{ gnmon_collector_textfile_dir }} --web.listen-address={{ gnmon_collector_node_exporter_listen }}
Restart=on-failure
RestartSec=2s
[Install]
WantedBy=default.target
@@ -0,0 +1,27 @@
# GNMON-001 v2.0 sections 3-4: VictoriaMetrics (single) as rootless quadlet.
#
# M1 scope: user-scope quadlet under the current user on the workstation.
# The image runs by tag during development; M5 pins the digest of the
# verified pull in the server inventory (repo convention: digest pin).
gnmon_victoria_metrics_image: docker.io/victoriametrics/victoria-metrics
gnmon_victoria_metrics_image_tag: v1.154.1
# Digest of the verified pull on the workstation (2026-11-04, M1).
gnmon_victoria_metrics_image_digest: c7884f32920dd45a9d86542a717227e63af03e060e53bbf8031e6e6fd0e1e05d
gnmon_victoria_metrics_user_home: "{{ lookup('ansible.builtin.env', 'HOME') }}"
gnmon_victoria_metrics_quadlet_dir: "{{ gnmon_victoria_metrics_user_home }}/.config/containers/systemd"
gnmon_victoria_metrics_config_dir: "{{ gnmon_victoria_metrics_user_home }}/.config/gnmon"
gnmon_victoria_metrics_data_dir: "{{ gnmon_victoria_metrics_user_home }}/.local/share/gnmon/victoria-metrics-data"
gnmon_victoria_metrics_port: 8428
gnmon_victoria_metrics_retention_days: 60
gnmon_victoria_metrics_scrape_interval_seconds: 10
# node_exporter target (same host; NetworkMode=host in the quadlet).
gnmon_victoria_metrics_node_target: "127.0.0.1:9100"
gnmon_victoria_metrics_image_ref: >-
{{ (gnmon_victoria_metrics_image_digest is not none)
| ternary(gnmon_victoria_metrics_image ~ '@sha256:' ~ gnmon_victoria_metrics_image_digest,
gnmon_victoria_metrics_image ~ ':' ~ gnmon_victoria_metrics_image_tag) }}
@@ -0,0 +1,7 @@
# GNMON-001: VictoriaMetrics role handlers (user scope, M1).
- name: Restart VictoriaMetrics
ansible.builtin.systemd:
name: gnmon-victoria-metrics.service
scope: user
state: restarted
daemon_reload: true
@@ -0,0 +1,62 @@
# GNMON-001 v2.0 sections 3-4: VictoriaMetrics rootless quadlet (M1: user scope).
- name: Validate VictoriaMetrics declarations
ansible.builtin.assert:
that:
- gnmon_victoria_metrics_user_home is match('^/')
- gnmon_victoria_metrics_retention_days | int >= 1
- gnmon_victoria_metrics_node_target | length > 0
fail_msg: VictoriaMetrics declarations are incomplete or ambiguous.
- name: Create VictoriaMetrics directories
ansible.builtin.file:
path: "{{ item }}"
state: directory
mode: "0755"
loop:
- "{{ gnmon_victoria_metrics_quadlet_dir }}"
- "{{ gnmon_victoria_metrics_config_dir }}"
- "{{ gnmon_victoria_metrics_data_dir }}"
- name: Install the promscrape configuration
ansible.builtin.template:
src: scrape.yml.j2
dest: "{{ gnmon_victoria_metrics_config_dir }}/scrape.yml"
mode: "0644"
notify: Restart VictoriaMetrics
- name: Install the VictoriaMetrics quadlet
ansible.builtin.template:
src: gnmon-victoria-metrics.container.j2
dest: "{{ gnmon_victoria_metrics_quadlet_dir }}/gnmon-victoria-metrics.container"
mode: "0644"
notify: Restart VictoriaMetrics
- name: Enable and start VictoriaMetrics
ansible.builtin.systemd:
name: gnmon-victoria-metrics.service
scope: user
enabled: true
state: started
daemon_reload: true
- name: Record the running image digest
ansible.builtin.command:
cmd: "podman inspect --format '{{ '{{' }}.ImageDigest{{ '}}' }}' gnmon-victoria-metrics"
changed_when: false
register: gnmon_victoria_metrics_inspect
failed_when: false
- name: Show the running image digest (digest pin for M5)
ansible.builtin.debug:
msg: >-
gnmon-victoria-metrics runs {{ gnmon_victoria_metrics_image }}:{{ gnmon_victoria_metrics_image_tag }}
(digest {{ gnmon_victoria_metrics_inspect.stdout | default('n/a') }})
- name: Wait for the VictoriaMetrics health endpoint
ansible.builtin.uri:
url: "http://127.0.0.1:{{ gnmon_victoria_metrics_port }}/health"
status_code: 200
register: gnmon_victoria_metrics_health
retries: 30
delay: 2
until: gnmon_victoria_metrics_health.status == 200
@@ -0,0 +1,19 @@
[Unit]
Description=GNMON VictoriaMetrics (GNMON-001)
Documentation=https://docs.victoriametrics.com/
[Container]
ContainerName=gnmon-victoria-metrics
Image={{ gnmon_victoria_metrics_image_ref }}
# Rootless quadlet reaches node_exporter on the host loopback interface.
Network=host
Exec=-promscrape.config=/etc/gnmon/scrape.yml -promscrape.suppressDuplicateScrapeTargetErrors -retentionPeriod={{ gnmon_victoria_metrics_retention_days }}d
Volume={{ gnmon_victoria_metrics_config_dir }}/scrape.yml:/etc/gnmon/scrape.yml:ro,Z
Volume={{ gnmon_victoria_metrics_data_dir }}:/victoria-metrics-data:Z
[Service]
# Quadlet services are user units; pods are transient managed containers.
Restart=on-failure
[Install]
WantedBy=default.target
@@ -0,0 +1,9 @@
# GNMON-001 section 4: VictoriaMetrics promscrape configuration.
global:
scrape_interval: {{ gnmon_victoria_metrics_scrape_interval_seconds }}s
scrape_configs:
- job_name: node
static_configs:
- targets:
- "{{ gnmon_victoria_metrics_node_target }}"
+213
View File
@@ -0,0 +1,213 @@
#!/usr/bin/env python3
"""AMD GPU textfile collector for Prometheus node_exporter (GNMON-001 v2.0).
Reads amdgpu metrics directly from sysfs (no rocm-smi, no ROCm runtime
required) and writes them as a Prometheus textfile exposition for the
node_exporter textfile collector. Same data source as the thermal load
profile from 2026-11-01 (GNAI-010 section 10.8).
Per AMD GPU (vendor 0x1002) the following metrics are emitted when the
underlying sysfs file exists (APUs expose a subset, discrete cards all):
gn_gpu_up{card,gpu}
gn_gpu_utilization_percent{card,gpu}
gn_gpu_vram_used_bytes{card,gpu}
gn_gpu_vram_total_bytes{card,gpu}
gn_gpu_edge_temp_celsius{card,gpu}
gn_gpu_junction_temp_celsius{card,gpu}
gn_gpu_memory_temp_celsius{card,gpu}
gn_gpu_fan_rpm{card,gpu}
gn_gpu_fan_pwm_percent{card,gpu}
gn_gpu_power_watts{card,gpu}
The output file is written atomically (tmp + rename) with mode 0644 so the
node_exporter textfile directory never exposes a partial file.
Usage: amdgpu-textfile-collector.py --textfile-dir DIR [--drm-root /sys/class/drm]
"""
from __future__ import annotations
import argparse
import os
import sys
import tempfile
from pathlib import Path
AMD_VENDOR_ID = "0x1002"
def _read_int(path: Path) -> int | None:
try:
return int(path.read_text().strip())
except (OSError, ValueError):
return None
def _read_text(path: Path) -> str | None:
try:
return path.read_text().strip()
except OSError:
return None
def _read_hwmon_temps(device_dir: Path) -> dict[str, float]:
"""Map hwmon temperature labels (edge/junction/mem/mcd) to celsius."""
temps: dict[str, float] = {}
for hwmon_dir in sorted(device_dir.glob("hwmon/hwmon*")):
for input_path in sorted(hwmon_dir.glob("temp*_input")):
value = _read_int(input_path)
if value is None:
continue
label_path = input_path.with_name(
input_path.name.replace("_input", "_label")
)
label = _read_text(label_path) or ""
label = label.lower()
if "junction" in label:
temps["junction"] = value / 1000.0
elif "edge" in label:
temps["edge"] = value / 1000.0
elif "mem" in label or "mcd" in label:
temps.setdefault("memory", value / 1000.0)
return temps
def collect(drm_root: Path) -> list[str]:
lines: list[str] = []
for device_link in sorted(drm_root.glob("card*")):
if "-" in device_link.name:
continue # connector entries like card0-eDP-1
device_dir = device_link / "device"
if not device_dir.is_dir():
continue
vendor = _read_text(device_dir / "vendor")
if vendor != AMD_VENDOR_ID:
continue
card = device_link.name.removeprefix("card")
gpu = _read_text(device_dir / "product_name") or "unknown"
def emit(metric: str, value, help_text: str, metric_type: str) -> None:
lines.append(f"# HELP {metric} {help_text}")
lines.append(f"# TYPE {metric} {metric_type}")
if value is not None:
lines.append(f'{metric}{{card="{card}",gpu="{gpu}"}} {value}')
emit("gn_gpu_up", 1, "AMD GPU is visible in sysfs", "gauge")
emit(
"gn_gpu_utilization_percent",
_read_int(device_dir / "gpu_busy_percent"),
"GPU utilization in percent (amdgpu sysfs)",
"gauge",
)
emit(
"gn_gpu_vram_used_bytes",
_read_int(device_dir / "mem_info_vram_used"),
"Used VRAM in bytes",
"gauge",
)
emit(
"gn_gpu_vram_total_bytes",
_read_int(device_dir / "mem_info_vram_total"),
"Total VRAM in bytes",
"gauge",
)
temps = _read_hwmon_temps(device_dir)
emit(
"gn_gpu_edge_temp_celsius",
temps.get("edge"),
"GPU edge temperature in celsius",
"gauge",
)
emit(
"gn_gpu_junction_temp_celsius",
temps.get("junction"),
"GPU junction (hotspot) temperature in celsius",
"gauge",
)
emit(
"gn_gpu_memory_temp_celsius",
temps.get("memory"),
"GPU memory temperature in celsius",
"gauge",
)
for hwmon_dir in sorted(device_dir.glob("hwmon/hwmon*")):
for fan_path in sorted(hwmon_dir.glob("fan*_input")):
fan_index = fan_path.name.removeprefix("fan").removesuffix("_input")
rpm = _read_int(fan_path)
if rpm is None:
continue
lines.append(f"# HELP gn_gpu_fan_rpm GPU fan speed in rpm")
lines.append(f"# TYPE gn_gpu_fan_rpm gauge")
lines.append(
f'gn_gpu_fan_rpm{{card="{card}",gpu="{gpu}",fan="{fan_index}"}} {rpm}'
)
pwm = _read_int(hwmon_dir / "pwm1")
if pwm is not None:
lines.append(f"# HELP gn_gpu_fan_pwm_percent Requested fan PWM in percent")
lines.append(f"# TYPE gn_gpu_fan_pwm_percent gauge")
lines.append(
f'gn_gpu_fan_pwm_percent{{card="{card}",gpu="{gpu}"}} {pwm / 255.0 * 100:.2f}'
)
# AMD power sensors report microwatts (power1_average preferred).
for power_name in ("power1_average", "power1_input"):
microwatts = _read_int(hwmon_dir / power_name)
if microwatts is not None:
lines.append(f"# HELP gn_gpu_power_watts GPU board power in watts")
lines.append(f"# TYPE gn_gpu_power_watts gauge")
lines.append(
f'gn_gpu_power_watts{{card="{card}",gpu="{gpu}"}} {microwatts / 1e6:.3f}'
)
break
return lines
def write_textfile(lines: list[str], textfile_dir: Path) -> Path:
textfile_dir.mkdir(parents=True, exist_ok=True)
target = textfile_dir / "amdgpu.prom"
fd, tmp_name = tempfile.mkstemp(
dir=textfile_dir, prefix=".amdgpu.prom.", suffix=".tmp"
)
try:
with os.fdopen(fd, "w", encoding="ascii") as handle:
handle.write("\n".join(lines) + "\n")
os.chmod(tmp_name, 0o644)
os.replace(tmp_name, target)
except BaseException:
try:
os.unlink(tmp_name)
except OSError:
pass
raise
return target
def main() -> int:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument(
"--textfile-dir",
type=Path,
default=Path("/var/lib/node_exporter/textfile"),
help="node_exporter textfile directory (default: %(default)s)",
)
parser.add_argument(
"--drm-root",
type=Path,
default=Path("/sys/class/drm"),
help="sysfs DRM root (default: %(default)s)",
)
parser.add_argument(
"--stdout", action="store_true", help="print exposition to stdout instead"
)
args = parser.parse_args()
lines = collect(args.drm_root)
if args.stdout:
print("\n".join(lines))
return 0
write_textfile(lines, args.textfile_dir)
return 0
if __name__ == "__main__":
sys.exit(main())