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

monitoring(gpu): report time-weighted namespace usage

Browse Source
This commit is contained in:
jenkins 2026-08-02 03:08:20 -03:00
parent 26a5dde12d
commit e794ccf254
7 changed files with 81 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

51
scripts/dashboards_render_atlas.py
View File

@ -284,8 +284,14 @@ def dcgm_gpu_util_by_node():
)
def nvidia_gpu_util_by_node():
return "max by (node) (nvidia_gpu_device_utilization_percent)"
def gpu_util_by_node():
return f"{dcgm_gpu_util_by_node()} or {jetson_gpu_util_by_node()}"
process_exporter = nvidia_gpu_util_by_node()
dcgm_fallback = f"({dcgm_gpu_util_by_node()}) unless on(node) ({process_exporter})"
return f"{process_exporter} or {dcgm_fallback} or {jetson_gpu_util_by_node()}"
def gpu_util_by_hostname():
@ -357,6 +363,21 @@ def nvidia_process_gpu_usage_by_namespace(scope_var):
return f"(({usage}) > 0)"
def nvidia_process_gpu_usage_by_namespace_range(scope_var, window="$__range"):
namespace_samples = (
"sum_over_time(nvidia_namespace_gpu_sm_util_percent"
f"{{{namespace_gpu_selector(scope_var)}}}[{window}])"
)
device_samples = (
f"count_over_time(nvidia_gpu_device_utilization_percent[{window}])"
)
averaged = (
f"({namespace_samples}) / on(node,gpu,uuid,model) group_left() "
f"clamp_min({device_samples}, 1)"
)
return f"((sum by (namespace) ({averaged})) > 0)"
def nvidia_gpu_device_utilization():
return "max by (node,gpu,uuid,model) (last_over_time(nvidia_gpu_device_utilization_percent[5m]))"
@ -390,17 +411,41 @@ def unattributed_gpu_usage():
)
def legacy_gpu_util_range_without_process_exporter(window="$__range"):
dcgm_range = dcgm_gpu_util_by_node().replace(
"DCGM_FI_DEV_GPU_UTIL", f"avg_over_time(DCGM_FI_DEV_GPU_UTIL[{window}])"
)
jetson_range = (
f'max by (node) (avg_over_time(jetson_gr3d_freq_percent{{node!=""}}[{window}]))'
)
process_exporter_nodes = (
f"max by (node) (count_over_time(nvidia_gpu_device_utilization_percent[{window}]))"
)
return f"(({dcgm_range}) or ({jetson_range})) unless on(node) ({process_exporter_nodes})"
def unattributed_gpu_usage_range(window="$__range"):
total = f"(sum({legacy_gpu_util_range_without_process_exporter(window)}) or on() vector(0))"
return f'label_replace(({total} > 0), "namespace", "unattributed", "", "")'
def gpu_utilization_raw(scope_var):
return f"({nvidia_process_gpu_usage_by_namespace(scope_var)}) or ({unattributed_gpu_usage()})"
def gpu_utilization_range_raw(scope_var, window="$__range"):
attributed = nvidia_process_gpu_usage_by_namespace_range(scope_var, window)
unattributed = unattributed_gpu_usage_range(window)
return f"({attributed}) or ({unattributed})"
def gpu_pool_used_expr(scope_var):
raw_total = f"(sum({gpu_utilization_raw(scope_var)}) or on() vector(0))"
return f"100 * {raw_total} / clamp_min({gpu_capacity_percent()}, 1)"
def namespace_gpu_share_expr(scope_var):
activity = gpu_utilization_raw(scope_var)
activity = gpu_utilization_range_raw(scope_var)
total = f"(sum({activity}) or on() vector(0))"
share = f"100 * ({activity}) / clamp_min({total}, 1)"
idle = (
@ -1910,7 +1955,7 @@ 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": "Instant share of observed GPU compute activity by namespace. Host covers GPU work outside Kubernetes pods; idle appears only when observed GPU activity is zero.",
"Namespace GPU Utilization": "Time-weighted share of observed GPU compute activity across the selected range. Process-aware NVIDIA metrics attribute titan-22/24 work to namespaces; host is non-pod work, and Jetson titan-20/21 activity remains unattributed when process-level attribution is unavailable. Idle appears only when observed activity is zero.",
"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.",

19
scripts/tests/test_dashboards_render_atlas.py
View File

@ -157,6 +157,11 @@ def test_overview_uses_readable_quality_power_and_gitops_panels():
gpu_expr = panels_by_title["Namespace GPU Utilization"]["targets"][0]["expr"]
assert "nvidia_namespace_gpu_sm_util_percent" in gpu_expr
assert "nvidia_gpu_device_utilization_percent" in gpu_expr
assert "sum_over_time" in gpu_expr
assert "count_over_time" in gpu_expr
assert "avg_over_time(DCGM_FI_DEV_GPU_UTIL[$__range])" in gpu_expr
assert "avg_over_time(jetson_gr3d_freq_percent" in gpu_expr
assert "$__range" in gpu_expr
assert "sum by (namespace)" in gpu_expr
assert 'namespace", "shared"' not in gpu_expr
assert "kube_node_labels" not in gpu_expr
@ -165,6 +170,20 @@ def test_overview_uses_readable_quality_power_and_gitops_panels():
assert 'namespace", "unattributed"' in gpu_expr
assert 'namespace", "idle"' in gpu_expr
assert panels_by_title["Namespace GPU Utilization"]["targets"][0]["instant"] is True
assert "Time-weighted share" in panels_by_title["Namespace GPU Utilization"]["description"]
def test_gpu_node_panel_prefers_stable_process_metrics_and_covers_all_gpu_families():
mod = load_module()
dashboard = mod.build_gpu_dashboard()
panels_by_title = {panel["title"]: panel for panel in flatten_panels(dashboard["panels"])}
node_expr = panels_by_title["GPU Util by Node"]["targets"][0]["expr"]
assert "nvidia_gpu_device_utilization_percent" in node_expr
assert "DCGM_FI_DEV_GPU_UTIL" in node_expr
assert "jetson_gr3d_freq_percent" in node_expr
assert "unless on(node)" in node_expr
assert mod.GPU_NODES == ["titan-20", "titan-21", "titan-22", "titan-24"]
def test_overview_and_testing_panels_all_have_concise_descriptions():

6
services/monitoring/dashboards/atlas-gpu.json
View File

@ -20,7 +20,7 @@
},
"targets": [
{
"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\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) unless on(node) (max by (node,gpu,uuid,model) (last_over_time(nvidia_gpu_device_utilization_percent[5m]))))) or on() vector(0)) > 0), \"namespace\", \"unattributed\", \"\", \"\"))) / clamp_min((sum((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace(((sum(((avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) unless on(node) (max by (node,gpu,uuid,model) (last_over_time(nvidia_gpu_device_utilization_percent[5m]))))) or on() vector(0)) > 0), \"namespace\", \"unattributed\", \"\", \"\"))) or on() vector(0)), 1)) or (label_replace(vector(100), \"namespace\", \"idle\", \"\", \"\") and on() ((sum((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace(((sum(((avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) unless on(node) (max by (node,gpu,uuid,model) (last_over_time(nvidia_gpu_device_utilization_percent[5m]))))) or on() vector(0)) > 0), \"namespace\", \"unattributed\", \"\", \"\"))) or on() vector(0)) == 0))",
"expr": "(100 * ((((sum by (namespace) ((sum_over_time(nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu}[$__range])) / on(node,gpu,uuid,model) group_left() clamp_min(count_over_time(nvidia_gpu_device_utilization_percent[$__range]), 1))) > 0)) or (label_replace(((sum(((avg by (node) (label_replace(label_replace(avg_over_time(DCGM_FI_DEV_GPU_UTIL[$__range]), \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"})) or (max by (node) (avg_over_time(jetson_gr3d_freq_percent{node!=\"\"}[$__range])))) unless on(node) (max by (node) (count_over_time(nvidia_gpu_device_utilization_percent[$__range])))) or on() vector(0)) > 0), \"namespace\", \"unattributed\", \"\", \"\"))) / clamp_min((sum((((sum by (namespace) ((sum_over_time(nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu}[$__range])) / on(node,gpu,uuid,model) group_left() clamp_min(count_over_time(nvidia_gpu_device_utilization_percent[$__range]), 1))) > 0)) or (label_replace(((sum(((avg by (node) (label_replace(label_replace(avg_over_time(DCGM_FI_DEV_GPU_UTIL[$__range]), \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"})) or (max by (node) (avg_over_time(jetson_gr3d_freq_percent{node!=\"\"}[$__range])))) unless on(node) (max by (node) (count_over_time(nvidia_gpu_device_utilization_percent[$__range])))) or on() vector(0)) > 0), \"namespace\", \"unattributed\", \"\", \"\"))) or on() vector(0)), 1)) or (label_replace(vector(100), \"namespace\", \"idle\", \"\", \"\") and on() ((sum((((sum by (namespace) ((sum_over_time(nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu}[$__range])) / on(node,gpu,uuid,model) group_left() clamp_min(count_over_time(nvidia_gpu_device_utilization_percent[$__range]), 1))) > 0)) or (label_replace(((sum(((avg by (node) (label_replace(label_replace(avg_over_time(DCGM_FI_DEV_GPU_UTIL[$__range]), \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"})) or (max by (node) (avg_over_time(jetson_gr3d_freq_percent{node!=\"\"}[$__range])))) unless on(node) (max by (node) (count_over_time(nvidia_gpu_device_utilization_percent[$__range])))) or on() vector(0)) > 0), \"namespace\", \"unattributed\", \"\", \"\"))) or on() vector(0)) == 0))",
"refId": "A",
"legendFormat": "{{namespace}}",
"instant": true
@ -72,7 +72,7 @@
"targetBlank": false
}
],
"description": "Instant share of observed GPU compute activity by namespace. Host covers GPU work outside Kubernetes pods; idle appears only when observed GPU activity is zero."
"description": "Time-weighted share of observed GPU compute activity across the selected range. Process-aware NVIDIA metrics attribute titan-22/24 work to namespaces; host is non-pod work, and Jetson titan-20/21 activity remains unattributed when process-level attribution is unavailable. Idle appears only when observed activity is zero."
},
{
"id": 2,
@ -128,7 +128,7 @@
},
"targets": [
{
"expr": "label_replace(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!=\"\"}), \"Hostname\", \"$1\", \"node\", \"(.*)\")",
"expr": "label_replace(max by (node) (nvidia_gpu_device_utilization_percent) or (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\"})) unless on(node) (max by (node) (nvidia_gpu_device_utilization_percent)) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"}), \"Hostname\", \"$1\", \"node\", \"(.*)\")",
"refId": "A",
"legendFormat": "{{Hostname}}"
}

4
services/monitoring/dashboards/atlas-overview.json
View File

@ -3728,7 +3728,7 @@
},
"targets": [
{
"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\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) unless on(node) (max by (node,gpu,uuid,model) (last_over_time(nvidia_gpu_device_utilization_percent[5m]))))) or on() vector(0)) > 0), \"namespace\", \"unattributed\", \"\", \"\"))) / clamp_min((sum((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace(((sum(((avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) unless on(node) (max by (node,gpu,uuid,model) (last_over_time(nvidia_gpu_device_utilization_percent[5m]))))) or on() vector(0)) > 0), \"namespace\", \"unattributed\", \"\", \"\"))) or on() vector(0)), 1)) or (label_replace(vector(100), \"namespace\", \"idle\", \"\", \"\") and on() ((sum((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace(((sum(((avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) unless on(node) (max by (node,gpu,uuid,model) (last_over_time(nvidia_gpu_device_utilization_percent[5m]))))) or on() vector(0)) > 0), \"namespace\", \"unattributed\", \"\", \"\"))) or on() vector(0)) == 0))",
"expr": "(100 * ((((sum by (namespace) ((sum_over_time(nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu}[$__range])) / on(node,gpu,uuid,model) group_left() clamp_min(count_over_time(nvidia_gpu_device_utilization_percent[$__range]), 1))) > 0)) or (label_replace(((sum(((avg by (node) (label_replace(label_replace(avg_over_time(DCGM_FI_DEV_GPU_UTIL[$__range]), \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"})) or (max by (node) (avg_over_time(jetson_gr3d_freq_percent{node!=\"\"}[$__range])))) unless on(node) (max by (node) (count_over_time(nvidia_gpu_device_utilization_percent[$__range])))) or on() vector(0)) > 0), \"namespace\", \"unattributed\", \"\", \"\"))) / clamp_min((sum((((sum by (namespace) ((sum_over_time(nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu}[$__range])) / on(node,gpu,uuid,model) group_left() clamp_min(count_over_time(nvidia_gpu_device_utilization_percent[$__range]), 1))) > 0)) or (label_replace(((sum(((avg by (node) (label_replace(label_replace(avg_over_time(DCGM_FI_DEV_GPU_UTIL[$__range]), \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"})) or (max by (node) (avg_over_time(jetson_gr3d_freq_percent{node!=\"\"}[$__range])))) unless on(node) (max by (node) (count_over_time(nvidia_gpu_device_utilization_percent[$__range])))) or on() vector(0)) > 0), \"namespace\", \"unattributed\", \"\", \"\"))) or on() vector(0)), 1)) or (label_replace(vector(100), \"namespace\", \"idle\", \"\", \"\") and on() ((sum((((sum by (namespace) ((sum_over_time(nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu}[$__range])) / on(node,gpu,uuid,model) group_left() clamp_min(count_over_time(nvidia_gpu_device_utilization_percent[$__range]), 1))) > 0)) or (label_replace(((sum(((avg by (node) (label_replace(label_replace(avg_over_time(DCGM_FI_DEV_GPU_UTIL[$__range]), \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"})) or (max by (node) (avg_over_time(jetson_gr3d_freq_percent{node!=\"\"}[$__range])))) unless on(node) (max by (node) (count_over_time(nvidia_gpu_device_utilization_percent[$__range])))) or on() vector(0)) > 0), \"namespace\", \"unattributed\", \"\", \"\"))) or on() vector(0)) == 0))",
"refId": "A",
"legendFormat": "{{namespace}}",
"instant": true
@ -3780,7 +3780,7 @@
"targetBlank": false
}
],
"description": "Instant share of observed GPU compute activity by namespace. Host covers GPU work outside Kubernetes pods; idle appears only when observed GPU activity is zero."
"description": "Time-weighted share of observed GPU compute activity across the selected range. Process-aware NVIDIA metrics attribute titan-22/24 work to namespaces; host is non-pod work, and Jetson titan-20/21 activity remains unattributed when process-level attribution is unavailable. Idle appears only when observed activity is zero."
},
{
"id": 13,

5
services/monitoring/dcgm-exporter.yaml
View File

@ -65,7 +65,10 @@ spec:
resources:
requests:
cpu: 50m
memory: 64Mi
memory: 512Mi
limits:
cpu: 500m
memory: 1Gi
volumeMounts:
- name: pod-resources
mountPath: /var/lib/kubelet/pod-resources

6
services/monitoring/grafana-dashboard-gpu.yaml
View File

@ -29,7 +29,7 @@ data:
},
"targets": [
{
"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\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) unless on(node) (max by (node,gpu,uuid,model) (last_over_time(nvidia_gpu_device_utilization_percent[5m]))))) or on() vector(0)) > 0), \"namespace\", \"unattributed\", \"\", \"\"))) / clamp_min((sum((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace(((sum(((avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) unless on(node) (max by (node,gpu,uuid,model) (last_over_time(nvidia_gpu_device_utilization_percent[5m]))))) or on() vector(0)) > 0), \"namespace\", \"unattributed\", \"\", \"\"))) or on() vector(0)), 1)) or (label_replace(vector(100), \"namespace\", \"idle\", \"\", \"\") and on() ((sum((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace(((sum(((avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) unless on(node) (max by (node,gpu,uuid,model) (last_over_time(nvidia_gpu_device_utilization_percent[5m]))))) or on() vector(0)) > 0), \"namespace\", \"unattributed\", \"\", \"\"))) or on() vector(0)) == 0))",
"expr": "(100 * ((((sum by (namespace) ((sum_over_time(nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu}[$__range])) / on(node,gpu,uuid,model) group_left() clamp_min(count_over_time(nvidia_gpu_device_utilization_percent[$__range]), 1))) > 0)) or (label_replace(((sum(((avg by (node) (label_replace(label_replace(avg_over_time(DCGM_FI_DEV_GPU_UTIL[$__range]), \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"})) or (max by (node) (avg_over_time(jetson_gr3d_freq_percent{node!=\"\"}[$__range])))) unless on(node) (max by (node) (count_over_time(nvidia_gpu_device_utilization_percent[$__range])))) or on() vector(0)) > 0), \"namespace\", \"unattributed\", \"\", \"\"))) / clamp_min((sum((((sum by (namespace) ((sum_over_time(nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu}[$__range])) / on(node,gpu,uuid,model) group_left() clamp_min(count_over_time(nvidia_gpu_device_utilization_percent[$__range]), 1))) > 0)) or (label_replace(((sum(((avg by (node) (label_replace(label_replace(avg_over_time(DCGM_FI_DEV_GPU_UTIL[$__range]), \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"})) or (max by (node) (avg_over_time(jetson_gr3d_freq_percent{node!=\"\"}[$__range])))) unless on(node) (max by (node) (count_over_time(nvidia_gpu_device_utilization_percent[$__range])))) or on() vector(0)) > 0), \"namespace\", \"unattributed\", \"\", \"\"))) or on() vector(0)), 1)) or (label_replace(vector(100), \"namespace\", \"idle\", \"\", \"\") and on() ((sum((((sum by (namespace) ((sum_over_time(nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu}[$__range])) / on(node,gpu,uuid,model) group_left() clamp_min(count_over_time(nvidia_gpu_device_utilization_percent[$__range]), 1))) > 0)) or (label_replace(((sum(((avg by (node) (label_replace(label_replace(avg_over_time(DCGM_FI_DEV_GPU_UTIL[$__range]), \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"})) or (max by (node) (avg_over_time(jetson_gr3d_freq_percent{node!=\"\"}[$__range])))) unless on(node) (max by (node) (count_over_time(nvidia_gpu_device_utilization_percent[$__range])))) or on() vector(0)) > 0), \"namespace\", \"unattributed\", \"\", \"\"))) or on() vector(0)) == 0))",
"refId": "A",
"legendFormat": "{{namespace}}",
"instant": true
@ -81,7 +81,7 @@ data:
"targetBlank": false
}
],
"description": "Instant share of observed GPU compute activity by namespace. Host covers GPU work outside Kubernetes pods; idle appears only when observed GPU activity is zero."
"description": "Time-weighted share of observed GPU compute activity across the selected range. Process-aware NVIDIA metrics attribute titan-22/24 work to namespaces; host is non-pod work, and Jetson titan-20/21 activity remains unattributed when process-level attribution is unavailable. Idle appears only when observed activity is zero."
},
{
"id": 2,
@ -137,7 +137,7 @@ data:
},
"targets": [
{
"expr": "label_replace(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!=\"\"}), \"Hostname\", \"$1\", \"node\", \"(.*)\")",
"expr": "label_replace(max by (node) (nvidia_gpu_device_utilization_percent) or (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\"})) unless on(node) (max by (node) (nvidia_gpu_device_utilization_percent)) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"}), \"Hostname\", \"$1\", \"node\", \"(.*)\")",
"refId": "A",
"legendFormat": "{{Hostname}}"
}

4
services/monitoring/grafana-dashboard-overview.yaml
View File

@ -3737,7 +3737,7 @@ data:
},
"targets": [
{
"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\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) unless on(node) (max by (node,gpu,uuid,model) (last_over_time(nvidia_gpu_device_utilization_percent[5m]))))) or on() vector(0)) > 0), \"namespace\", \"unattributed\", \"\", \"\"))) / clamp_min((sum((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace(((sum(((avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) unless on(node) (max by (node,gpu,uuid,model) (last_over_time(nvidia_gpu_device_utilization_percent[5m]))))) or on() vector(0)) > 0), \"namespace\", \"unattributed\", \"\", \"\"))) or on() vector(0)), 1)) or (label_replace(vector(100), \"namespace\", \"idle\", \"\", \"\") and on() ((sum((((sum by (namespace) (nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu})) > 0)) or (label_replace(((sum(((avg by (node) (label_replace(label_replace(DCGM_FI_DEV_GPU_UTIL, \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"}) or max by (node) (jetson_gr3d_freq_percent{node!=\"\"})) unless on(node) (max by (node,gpu,uuid,model) (last_over_time(nvidia_gpu_device_utilization_percent[5m]))))) or on() vector(0)) > 0), \"namespace\", \"unattributed\", \"\", \"\"))) or on() vector(0)) == 0))",
"expr": "(100 * ((((sum by (namespace) ((sum_over_time(nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu}[$__range])) / on(node,gpu,uuid,model) group_left() clamp_min(count_over_time(nvidia_gpu_device_utilization_percent[$__range]), 1))) > 0)) or (label_replace(((sum(((avg by (node) (label_replace(label_replace(avg_over_time(DCGM_FI_DEV_GPU_UTIL[$__range]), \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"})) or (max by (node) (avg_over_time(jetson_gr3d_freq_percent{node!=\"\"}[$__range])))) unless on(node) (max by (node) (count_over_time(nvidia_gpu_device_utilization_percent[$__range])))) or on() vector(0)) > 0), \"namespace\", \"unattributed\", \"\", \"\"))) / clamp_min((sum((((sum by (namespace) ((sum_over_time(nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu}[$__range])) / on(node,gpu,uuid,model) group_left() clamp_min(count_over_time(nvidia_gpu_device_utilization_percent[$__range]), 1))) > 0)) or (label_replace(((sum(((avg by (node) (label_replace(label_replace(avg_over_time(DCGM_FI_DEV_GPU_UTIL[$__range]), \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"})) or (max by (node) (avg_over_time(jetson_gr3d_freq_percent{node!=\"\"}[$__range])))) unless on(node) (max by (node) (count_over_time(nvidia_gpu_device_utilization_percent[$__range])))) or on() vector(0)) > 0), \"namespace\", \"unattributed\", \"\", \"\"))) or on() vector(0)), 1)) or (label_replace(vector(100), \"namespace\", \"idle\", \"\", \"\") and on() ((sum((((sum by (namespace) ((sum_over_time(nvidia_namespace_gpu_sm_util_percent{namespace!=\"\",pod!=\"\",$namespace_scope_gpu}[$__range])) / on(node,gpu,uuid,model) group_left() clamp_min(count_over_time(nvidia_gpu_device_utilization_percent[$__range]), 1))) > 0)) or (label_replace(((sum(((avg by (node) (label_replace(label_replace(avg_over_time(DCGM_FI_DEV_GPU_UTIL[$__range]), \"pod\", \"$1\", \"Hostname\", \"(.*)\"), \"namespace\", \"monitoring\", \"\", \"\") * on(namespace,pod) group_left(node) kube_pod_info{namespace=\"monitoring\"})) or (max by (node) (avg_over_time(jetson_gr3d_freq_percent{node!=\"\"}[$__range])))) unless on(node) (max by (node) (count_over_time(nvidia_gpu_device_utilization_percent[$__range])))) or on() vector(0)) > 0), \"namespace\", \"unattributed\", \"\", \"\"))) or on() vector(0)) == 0))",
"refId": "A",
"legendFormat": "{{namespace}}",
"instant": true
@ -3789,7 +3789,7 @@ data:
"targetBlank": false
}
],
"description": "Instant share of observed GPU compute activity by namespace. Host covers GPU work outside Kubernetes pods; idle appears only when observed GPU activity is zero."
"description": "Time-weighted share of observed GPU compute activity across the selected range. Process-aware NVIDIA metrics attribute titan-22/24 work to namespaces; host is non-pod work, and Jetson titan-20/21 activity remains unattributed when process-level attribution is unavailable. Idle appears only when observed activity is zero."
},
{
"id": 13,