bstein/titan-iac
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

monitoring(gpu): normalize utilization pie to pool capacity

Browse Source
This commit is contained in:
jenkins 2026-05-22 02:55:24 -03:00
parent ea21e106cf
commit 570b1212d7
6 changed files with 36 additions and 21 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

26
scripts/dashboards_render_atlas.py
View File

@ -376,17 +376,20 @@ def unattributed_gpu_usage():
def namespace_gpu_share_expr(scope_var):
utilization = f"({nvidia_process_gpu_usage_by_namespace(scope_var)}) or ({unattributed_gpu_usage()})"
utilization_raw = f"({nvidia_process_gpu_usage_by_namespace(scope_var)}) or ({unattributed_gpu_usage()})"
total_raw = f"(sum({utilization_raw}) or on() vector(0))"
capacity = gpu_capacity_percent()
utilization = f"100 * ({utilization_raw}) / clamp_min({capacity}, 1)"
total = f"(sum({utilization}) or on() vector(0))"
unused = (
'label_replace(clamp_min('
f"{gpu_capacity_percent()} - {total}"
f"100 - {total}"
', 0), "namespace", "unused", "", "") '
f"and on() ({total} > 0)"
f"and on() ({total_raw} > 0)"
)
idle = (
'label_replace(vector(100), "namespace", "idle", "", "") '
f"and on() ({total} == 0)"
f"and on() ({total_raw} == 0)"
)
return f"({utilization}) or ({unused}) or ({idle})"
@ -1393,15 +1396,18 @@ def table_panel(
return panel
def pie_panel(panel_id, title, expr, grid, *, links=None, description=None):
def pie_panel(panel_id, title, expr, grid, *, links=None, description=None, instant=False):
"""Return a pie chart panel with readable namespace labels."""
target = {"expr": expr, "refId": "A", "legendFormat": "{{namespace}}"}
if instant:
target["instant"] = True
panel = {
"id": panel_id,
"type": "piechart",
"title": title,
"datasource": PROM_DS,
"gridPos": grid,
"targets": [{"expr": expr, "refId": "A", "legendFormat": "{{namespace}}"}],
"targets": [target],
"fieldConfig": {
"defaults": {
"unit": "percent",
@ -1845,7 +1851,7 @@ OVERVIEW_PANEL_DESCRIPTIONS = {
"Postgres Connections Used": "Current Postgres connections; lower leaves room for apps during spikes.",
"Postgres Hottest Connections": "Database with the most active connections; high values identify the pressure source.",
"Namespace CPU Share": "CPU share by namespace in the selected scope; big slices show who is using compute.",
"Namespace GPU Utilization": "Current NVIDIA process-level GPU utilization by namespace. Host covers non-Kubernetes processes; unused fills remaining capacity while active; idle appears only at zero activity.",
"Namespace GPU Utilization": "Instant NVIDIA process-level GPU utilization normalized to the monitored GPU pool. Host covers non-Kubernetes processes; unused fills remaining capacity; idle appears only at zero activity.",
"Namespace RAM Share": "Memory share by namespace in the selected scope; big slices show who may drive pressure.",
"Worker Node CPU": "Worker CPU over time; lower is calmer, sustained high load may need rescheduling.",
"Worker Node RAM": "Worker memory over time; lower is safer, sustained high use risks evictions.",
@ -2862,7 +2868,8 @@ def build_overview():
namespace_gpu_share_expr(gpu_scope),
{"h": 9, "w": 8, "x": 8, "y": 23},
links=namespace_scope_links("namespace_scope_gpu"),
description="Current NVIDIA process-level GPU utilization by namespace. Host covers non-Kubernetes processes; unused fills remaining capacity while active; idle appears only at zero activity.",
description="Instant NVIDIA process-level GPU utilization normalized to the monitored GPU pool. Host covers non-Kubernetes processes; unused fills remaining capacity; idle appears only at zero activity.",
instant=True,
)
)
panels.append(
@ -5383,7 +5390,8 @@ def build_gpu_dashboard():
namespace_gpu_share_expr(gpu_scope),
{"h": 8, "w": 12, "x": 0, "y": 0},
links=namespace_scope_links("namespace_scope_gpu"),
description="Current NVIDIA process-level GPU utilization by namespace. Host covers non-Kubernetes processes; unused fills remaining capacity while active; idle appears only at zero activity.",
description="Instant NVIDIA process-level GPU utilization normalized to the monitored GPU pool. Host covers non-Kubernetes processes; unused fills remaining capacity; idle appears only at zero activity.",
instant=True,
)
)
panels.append(

3
scripts/tests/test_dashboards_render_atlas.py
View File

@ -160,8 +160,11 @@ def test_overview_uses_readable_quality_power_and_gitops_panels():
assert "sum by (namespace)" in gpu_expr
assert 'namespace", "shared"' not in gpu_expr
assert "kube_node_labels" not in gpu_expr
assert "100 *" in gpu_expr
assert "100 -" in gpu_expr
assert 'namespace", "unused"' in gpu_expr
assert 'namespace", "idle"' in gpu_expr
assert panels_by_title["Namespace GPU Utilization"]["targets"][0]["instant"] is True
def test_overview_and_testing_panels_all_have_concise_descriptions():

7
services/monitoring/dashboards/atlas-gpu.json
View File

@ -20,9 +20,10 @@
},
"targets": [
{
"expr": "((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) or (label_replace(clamp_min(((100 * count(nvidia_gpu_device_utilization_percent)) or (100 * count(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) unless on() nvidia_gpu_device_utilization_percent) or on() vector(0)) - (sum((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) or on() vector(0)), 0), \"namespace\", \"unused\", \"\", \"\") and on() ((sum((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) or on() vector(0)) > 0)) or (label_replace(vector(100), \"namespace\", \"idle\", \"\", \"\") and on() ((sum((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) or on() vector(0)) == 0))",
"expr": "(100 * ((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) / clamp_min(((100 * count(nvidia_gpu_device_utilization_percent)) or (100 * count(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) unless on() nvidia_gpu_device_utilization_percent) or on() vector(0)), 1)) or (label_replace(clamp_min(100 - (sum(100 * ((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) / clamp_min(((100 * count(nvidia_gpu_device_utilization_percent)) or (100 * count(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) unless on() nvidia_gpu_device_utilization_percent) or on() vector(0)), 1)) or on() vector(0)), 0), \"namespace\", \"unused\", \"\", \"\") and on() ((sum((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) or on() vector(0)) > 0)) or (label_replace(vector(100), \"namespace\", \"idle\", \"\", \"\") and on() ((sum((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) or on() vector(0)) == 0))",
"refId": "A",
"legendFormat": "{{namespace}}"
"legendFormat": "{{namespace}}",
"instant": true
}
],
"fieldConfig": {
@ -71,7 +72,7 @@
"targetBlank": false
}
],
"description": "Current NVIDIA process-level GPU utilization by namespace. Host covers non-Kubernetes processes; unused fills remaining capacity while active; idle appears only at zero activity."
"description": "Instant NVIDIA process-level GPU utilization normalized to the monitored GPU pool. Host covers non-Kubernetes processes; unused fills remaining capacity; idle appears only at zero activity."
},
{
"id": 2,

7
services/monitoring/dashboards/atlas-overview.json
View File

@ -3728,9 +3728,10 @@
},
"targets": [
{
"expr": "((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) or (label_replace(clamp_min(((100 * count(nvidia_gpu_device_utilization_percent)) or (100 * count(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) unless on() nvidia_gpu_device_utilization_percent) or on() vector(0)) - (sum((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) or on() vector(0)), 0), \"namespace\", \"unused\", \"\", \"\") and on() ((sum((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) or on() vector(0)) > 0)) or (label_replace(vector(100), \"namespace\", \"idle\", \"\", \"\") and on() ((sum((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) or on() vector(0)) == 0))",
"expr": "(100 * ((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) / clamp_min(((100 * count(nvidia_gpu_device_utilization_percent)) or (100 * count(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) unless on() nvidia_gpu_device_utilization_percent) or on() vector(0)), 1)) or (label_replace(clamp_min(100 - (sum(100 * ((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) / clamp_min(((100 * count(nvidia_gpu_device_utilization_percent)) or (100 * count(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) unless on() nvidia_gpu_device_utilization_percent) or on() vector(0)), 1)) or on() vector(0)), 0), \"namespace\", \"unused\", \"\", \"\") and on() ((sum((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) or on() vector(0)) > 0)) or (label_replace(vector(100), \"namespace\", \"idle\", \"\", \"\") and on() ((sum((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) or on() vector(0)) == 0))",
"refId": "A",
"legendFormat": "{{namespace}}"
"legendFormat": "{{namespace}}",
"instant": true
}
],
"fieldConfig": {
@ -3779,7 +3780,7 @@
"targetBlank": false
}
],
"description": "Current NVIDIA process-level GPU utilization by namespace. Host covers non-Kubernetes processes; unused fills remaining capacity while active; idle appears only at zero activity."
"description": "Instant NVIDIA process-level GPU utilization normalized to the monitored GPU pool. Host covers non-Kubernetes processes; unused fills remaining capacity; idle appears only at zero activity."
},
{
"id": 13,

7
services/monitoring/grafana-dashboard-gpu.yaml
View File

@ -29,9 +29,10 @@ data:
},
"targets": [
{
"expr": "((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) or (label_replace(clamp_min(((100 * count(nvidia_gpu_device_utilization_percent)) or (100 * count(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) unless on() nvidia_gpu_device_utilization_percent) or on() vector(0)) - (sum((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) or on() vector(0)), 0), \"namespace\", \"unused\", \"\", \"\") and on() ((sum((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) or on() vector(0)) > 0)) or (label_replace(vector(100), \"namespace\", \"idle\", \"\", \"\") and on() ((sum((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) or on() vector(0)) == 0))",
"expr": "(100 * ((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) / clamp_min(((100 * count(nvidia_gpu_device_utilization_percent)) or (100 * count(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) unless on() nvidia_gpu_device_utilization_percent) or on() vector(0)), 1)) or (label_replace(clamp_min(100 - (sum(100 * ((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) / clamp_min(((100 * count(nvidia_gpu_device_utilization_percent)) or (100 * count(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) unless on() nvidia_gpu_device_utilization_percent) or on() vector(0)), 1)) or on() vector(0)), 0), \"namespace\", \"unused\", \"\", \"\") and on() ((sum((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) or on() vector(0)) > 0)) or (label_replace(vector(100), \"namespace\", \"idle\", \"\", \"\") and on() ((sum((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) or on() vector(0)) == 0))",
"refId": "A",
"legendFormat": "{{namespace}}"
"legendFormat": "{{namespace}}",
"instant": true
}
],
"fieldConfig": {
@ -80,7 +81,7 @@ data:
"targetBlank": false
}
],
"description": "Current NVIDIA process-level GPU utilization by namespace. Host covers non-Kubernetes processes; unused fills remaining capacity while active; idle appears only at zero activity."
"description": "Instant NVIDIA process-level GPU utilization normalized to the monitored GPU pool. Host covers non-Kubernetes processes; unused fills remaining capacity; idle appears only at zero activity."
},
{
"id": 2,

7
services/monitoring/grafana-dashboard-overview.yaml
View File

@ -3737,9 +3737,10 @@ data:
},
"targets": [
{
"expr": "((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) or (label_replace(clamp_min(((100 * count(nvidia_gpu_device_utilization_percent)) or (100 * count(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) unless on() nvidia_gpu_device_utilization_percent) or on() vector(0)) - (sum((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) or on() vector(0)), 0), \"namespace\", \"unused\", \"\", \"\") and on() ((sum((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) or on() vector(0)) > 0)) or (label_replace(vector(100), \"namespace\", \"idle\", \"\", \"\") and on() ((sum((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) or on() vector(0)) == 0))",
"expr": "(100 * ((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) / clamp_min(((100 * count(nvidia_gpu_device_utilization_percent)) or (100 * count(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) unless on() nvidia_gpu_device_utilization_percent) or on() vector(0)), 1)) or (label_replace(clamp_min(100 - (sum(100 * ((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) / clamp_min(((100 * count(nvidia_gpu_device_utilization_percent)) or (100 * count(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) unless on() nvidia_gpu_device_utilization_percent) or on() vector(0)), 1)) or on() vector(0)), 0), \"namespace\", \"unused\", \"\", \"\") and on() ((sum((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) or on() vector(0)) > 0)) or (label_replace(vector(100), \"namespace\", \"idle\", \"\", \"\") and on() ((sum((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace((sum(avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) or on() vector(0)), \"namespace\", \"unattributed\", \"\", \"\") unless on() ((count(nvidia_gpu_device_utilization_percent) or on() vector(0)) > 0))) or on() vector(0)) == 0))",
"refId": "A",
"legendFormat": "{{namespace}}"
"legendFormat": "{{namespace}}",
"instant": true
}
],
"fieldConfig": {
@ -3788,7 +3789,7 @@ data:
"targetBlank": false
}
],
"description": "Current NVIDIA process-level GPU utilization by namespace. Host covers non-Kubernetes processes; unused fills remaining capacity while active; idle appears only at zero activity."
"description": "Instant NVIDIA process-level GPU utilization normalized to the monitored GPU pool. Host covers non-Kubernetes processes; unused fills remaining capacity; idle appears only at zero activity."
},
{
"id": 13,