From 91e648f55ed4fdabd09e5a75e04eb73e67225a3d Mon Sep 17 00:00:00 2001
From: alkatrazz <alkatrazz@localhost.localdomain>
Date: Tue, 2 Jun 2026 22:54:51 +0200
Subject: [PATCH] =?UTF-8?q?feat(monitoring):=20m=C3=A9triques=20conso=20?=
 =?UTF-8?q?=C3=A9lectrique=20CPU+GPU=20(RAPL=20+=20ROCm=20power)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- node_exporter : activer --collector.rapl (CPU power package, storage-01 + gpu-01 + nœuds k8s)
- rocm_scraper : ajouter rocm_gpu_power_watts + rocm_gpu_power_cap_watts via sysfs hwmon
- gpu-01 : activer rocm_exporter_enabled
- kube-prometheus-stack : propager --collector.rapl aux DaemonSet Talos via prometheus-node-exporter.extraArgs

Déployé sur storage-01 et gpu-01. ArgoCD prend en charge les nœuds Talos.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 ansible/host_vars/gpu-01/vars.yml                      |  3 +++
 .../node_exporter/templates/node_exporter.service.j2   |  3 ++-
 .../roles/node_exporter/templates/rocm_scraper.sh.j2   | 10 ++++++++++
 k8s/infra/monitoring/values.yaml                       |  4 ++++
 4 files changed, 19 insertions(+), 1 deletion(-)

diff --git a/ansible/host_vars/gpu-01/vars.yml b/ansible/host_vars/gpu-01/vars.yml
index 8acc216..986fada 100644
--- a/ansible/host_vars/gpu-01/vars.yml
+++ b/ansible/host_vars/gpu-01/vars.yml
@@ -6,6 +6,9 @@ llama_model_path: "/mnt/models/bartowski/Qwen3-8B-GGUF/Qwen3-8B-Q4_K_M.gguf"
 llama_model_alias: "qwen3-8b"
 llama_ctx_size: 32768
 
+# ROCm scraper — métriques GPU dans Prometheus
+rocm_exporter_enabled: true
+
 # Montages NFS (surcharge les defaults du rôle nfs_client)
 nfs_client_mounts:
   - local_path: /mnt/nfs
diff --git a/ansible/roles/node_exporter/templates/node_exporter.service.j2 b/ansible/roles/node_exporter/templates/node_exporter.service.j2
index 95fc896..0d39f6a 100644
--- a/ansible/roles/node_exporter/templates/node_exporter.service.j2
+++ b/ansible/roles/node_exporter/templates/node_exporter.service.j2
@@ -9,7 +9,8 @@ ExecStart=/usr/local/bin/node_exporter \
   --web.listen-address=":{{ node_exporter_port }}" \
   --collector.textfile.directory={{ node_exporter_textfile_dir }} \
   --collector.systemd \
-  --collector.processes
+  --collector.processes \
+  --collector.rapl
 
 Restart=on-failure
 RestartSec=5s
diff --git a/ansible/roles/node_exporter/templates/rocm_scraper.sh.j2 b/ansible/roles/node_exporter/templates/rocm_scraper.sh.j2
index ab9b9b9..7752222 100644
--- a/ansible/roles/node_exporter/templates/rocm_scraper.sh.j2
+++ b/ansible/roles/node_exporter/templates/rocm_scraper.sh.j2
@@ -12,6 +12,10 @@ TMPFILE="${OUTPUT}.tmp"
   echo "# TYPE rocm_vram_used_bytes gauge"
   echo "# HELP rocm_vram_total_bytes VRAM total in bytes"
   echo "# TYPE rocm_vram_total_bytes gauge"
+  echo "# HELP rocm_gpu_power_watts GPU power draw in watts"
+  echo "# TYPE rocm_gpu_power_watts gauge"
+  echo "# HELP rocm_gpu_power_cap_watts GPU power cap in watts"
+  echo "# TYPE rocm_gpu_power_cap_watts gauge"
 
   for card in /sys/class/drm/card[0-9]; do
     [ -d "$card/device" ] || continue
@@ -29,5 +33,11 @@ TMPFILE="${OUTPUT}.tmp"
 
     vram_total=$(cat "$card/device/mem_info_vram_total" 2>/dev/null)
     [ -n "$vram_total" ] && echo "rocm_vram_total_bytes{${labels}} ${vram_total}"
+
+    power_raw=$(cat "$card/device/hwmon/hwmon"*/power1_average 2>/dev/null | head -1)
+    [ -n "$power_raw" ] && echo "rocm_gpu_power_watts{${labels}} $(echo "scale=2; ${power_raw}/1000000" | bc)"
+
+    power_cap_raw=$(cat "$card/device/hwmon/hwmon"*/power1_cap 2>/dev/null | head -1)
+    [ -n "$power_cap_raw" ] && echo "rocm_gpu_power_cap_watts{${labels}} $(echo "scale=2; ${power_cap_raw}/1000000" | bc)"
   done
 } > "$TMPFILE" && mv "$TMPFILE" "$OUTPUT"
diff --git a/k8s/infra/monitoring/values.yaml b/k8s/infra/monitoring/values.yaml
index b129763..2d06485 100644
--- a/k8s/infra/monitoring/values.yaml
+++ b/k8s/infra/monitoring/values.yaml
@@ -162,6 +162,10 @@ nodeExporter:
     limits:
       memory: 64Mi
 
+prometheus-node-exporter:
+  extraArgs:
+    - --collector.rapl
+
 # Désactiver etcd/scheduler monitoring (Talos — non accessible)
 kubeEtcd:
   enabled: false