cozystack/dashboards/gpu/gpu-fleet.json
Arsolitt 950c5dd669
fix(fleet): guard TDP division and document DCGM dependency
Signed-off-by: Arsolitt <arsolitt@gmail.com>
2026-04-19 11:01:40 +03:00

1157 lines
26 KiB
JSON

{
"uid": "gpu-fleet",
"title": "GPU Fleet Overview",
"description": "Cluster-wide GPU inventory, capacity, utilization and health — admin view across the whole fleet.",
"tags": [
"gpu",
"fleet",
"admin"
],
"timezone": "browser",
"editable": true,
"graphTooltip": 1,
"time": {
"from": "now-6h",
"to": "now"
},
"fiscalYearStartMonth": 0,
"schemaVersion": 42,
"panels": [
{
"type": "row",
"collapsed": false,
"title": "Cluster snapshot",
"gridPos": {
"h": 1,
"w": 24,
"x": 0,
"y": 0
},
"id": 1,
"panels": []
},
{
"type": "stat",
"id": 2,
"targets": [
{
"expr": "cluster:gpu_count:total",
"refId": "A"
}
],
"title": "Total GPUs",
"description": "Physical GPUs discovered by DCGM across all nodes.",
"transparent": false,
"datasource": {
"type": "prometheus",
"uid": "$ds_prometheus"
},
"gridPos": {
"h": 4,
"w": 4,
"x": 0,
"y": 1
},
"repeatDirection": "h",
"options": {
"graphMode": "none",
"colorMode": "value",
"justifyMode": "auto",
"textMode": "value",
"wideLayout": true,
"showPercentChange": false,
"reduceOptions": {
"calcs": [
"lastNotNull"
]
},
"percentChangeColorMode": "standard",
"orientation": ""
},
"fieldConfig": {
"defaults": {
"unit": "short",
"thresholds": {
"mode": "absolute",
"steps": [
{
"value": null,
"color": "blue"
}
]
}
},
"overrides": []
}
},
{
"type": "stat",
"id": 3,
"targets": [
{
"expr": "cluster:gpu_count:allocated",
"refId": "A"
}
],
"title": "Allocated",
"description": "Sum of kube GPU requests across all pods. Billable view — includes Pending pods.",
"transparent": false,
"datasource": {
"type": "prometheus",
"uid": "$ds_prometheus"
},
"gridPos": {
"h": 4,
"w": 4,
"x": 4,
"y": 1
},
"repeatDirection": "h",
"options": {
"graphMode": "none",
"colorMode": "background",
"justifyMode": "auto",
"textMode": "value",
"wideLayout": true,
"showPercentChange": false,
"reduceOptions": {
"calcs": [
"lastNotNull"
]
},
"percentChangeColorMode": "standard",
"orientation": ""
},
"fieldConfig": {
"defaults": {
"unit": "short",
"thresholds": {
"mode": "absolute",
"steps": [
{
"value": null,
"color": "blue"
},
{
"value": 1,
"color": "green"
}
]
}
},
"overrides": []
}
},
{
"type": "stat",
"id": 4,
"targets": [
{
"expr": "cluster:gpu_count:free",
"refId": "A"
}
],
"title": "Free",
"description": "Unallocated capacity available for new tenants.",
"transparent": false,
"datasource": {
"type": "prometheus",
"uid": "$ds_prometheus"
},
"gridPos": {
"h": 4,
"w": 4,
"x": 8,
"y": 1
},
"repeatDirection": "h",
"options": {
"graphMode": "none",
"colorMode": "background",
"justifyMode": "auto",
"textMode": "value",
"wideLayout": true,
"showPercentChange": false,
"reduceOptions": {
"calcs": [
"lastNotNull"
]
},
"percentChangeColorMode": "standard",
"orientation": ""
},
"fieldConfig": {
"defaults": {
"unit": "short",
"thresholds": {
"mode": "absolute",
"steps": [
{
"value": null,
"color": "red"
},
{
"value": 1,
"color": "yellow"
},
{
"value": 2,
"color": "green"
}
]
}
},
"overrides": []
}
},
{
"type": "stat",
"id": 5,
"targets": [
{
"expr": "count(count by (Hostname) (DCGM_FI_DEV_GPU_UTIL))",
"refId": "A"
}
],
"title": "Nodes with GPU",
"description": "Nodes reporting at least one GPU to DCGM. Counted via DCGM_FI_DEV_GPU_UTIL because kube-state-metrics does not expose node labels by default.",
"transparent": false,
"datasource": {
"type": "prometheus",
"uid": "$ds_prometheus"
},
"gridPos": {
"h": 4,
"w": 4,
"x": 12,
"y": 1
},
"repeatDirection": "h",
"options": {
"graphMode": "none",
"colorMode": "value",
"justifyMode": "auto",
"textMode": "value",
"wideLayout": true,
"showPercentChange": false,
"reduceOptions": {
"calcs": [
"lastNotNull"
]
},
"percentChangeColorMode": "standard",
"orientation": ""
},
"fieldConfig": {
"defaults": {
"unit": "short",
"thresholds": {
"mode": "absolute",
"steps": [
{
"value": null,
"color": "blue"
}
]
}
},
"overrides": []
}
},
{
"type": "stat",
"id": 6,
"targets": [
{
"expr": "cluster:gpu_util:avg",
"refId": "A"
}
],
"title": "Average utilization",
"description": "Legacy NVML utilization across all tenant GPUs.",
"transparent": false,
"datasource": {
"type": "prometheus",
"uid": "$ds_prometheus"
},
"gridPos": {
"h": 4,
"w": 4,
"x": 16,
"y": 1
},
"repeatDirection": "h",
"options": {
"graphMode": "area",
"colorMode": "value",
"justifyMode": "auto",
"textMode": "value",
"wideLayout": true,
"showPercentChange": false,
"reduceOptions": {
"calcs": [
"lastNotNull"
]
},
"percentChangeColorMode": "standard",
"orientation": ""
},
"fieldConfig": {
"defaults": {
"unit": "percent",
"min": 0,
"max": 100,
"thresholds": {
"mode": "absolute",
"steps": [
{
"value": null,
"color": "blue"
},
{
"value": 30,
"color": "green"
},
{
"value": 80,
"color": "orange"
}
]
}
},
"overrides": []
}
},
{
"type": "stat",
"id": 7,
"targets": [
{
"expr": "cluster:gpu_power_watts:sum",
"refId": "A"
}
],
"title": "Total power",
"description": "Live power draw summed across the fleet.",
"transparent": false,
"datasource": {
"type": "prometheus",
"uid": "$ds_prometheus"
},
"gridPos": {
"h": 4,
"w": 4,
"x": 20,
"y": 1
},
"repeatDirection": "h",
"options": {
"graphMode": "area",
"colorMode": "value",
"justifyMode": "auto",
"textMode": "value",
"wideLayout": true,
"showPercentChange": false,
"reduceOptions": {
"calcs": [
"lastNotNull"
]
},
"percentChangeColorMode": "standard",
"orientation": ""
},
"fieldConfig": {
"defaults": {
"unit": "watt",
"decimals": 0,
"thresholds": {
"mode": "absolute",
"steps": [
{
"value": null,
"color": "green"
}
]
}
},
"overrides": []
}
},
{
"type": "row",
"collapsed": false,
"title": "Per-node GPU activity",
"gridPos": {
"h": 1,
"w": 24,
"x": 0,
"y": 5
},
"id": 10,
"panels": []
},
{
"type": "bargauge",
"id": 11,
"targets": [
{
"expr": "avg by (Hostname) (DCGM_FI_DEV_GPU_UTIL)",
"legendFormat": "{{Hostname}}",
"refId": "A"
}
],
"title": "GPU utilization per node",
"description": "Average NVML utilization per node — spot idle or saturated hosts at a glance.",
"transparent": false,
"datasource": {
"type": "prometheus",
"uid": "$ds_prometheus"
},
"gridPos": {
"h": 8,
"w": 12,
"x": 0,
"y": 6
},
"repeatDirection": "h",
"options": {
"displayMode": "gradient",
"valueMode": "color",
"namePlacement": "auto",
"showUnfilled": true,
"sizing": "auto",
"minVizWidth": 8,
"minVizHeight": 16,
"legend": {
"displayMode": "list",
"placement": "bottom",
"showLegend": false,
"calcs": []
},
"reduceOptions": {
"calcs": [
"lastNotNull"
]
},
"maxVizHeight": 300,
"orientation": "horizontal"
},
"fieldConfig": {
"defaults": {
"unit": "percent",
"min": 0,
"max": 100,
"thresholds": {
"mode": "absolute",
"steps": [
{
"value": null,
"color": "blue"
},
{
"value": 30,
"color": "green"
},
{
"value": 80,
"color": "orange"
}
]
}
},
"overrides": []
}
},
{
"type": "bargauge",
"id": 12,
"targets": [
{
"expr": "sum by (Hostname) (DCGM_FI_DEV_POWER_USAGE) / clamp_min(sum by (Hostname) (DCGM_FI_DEV_POWER_MGMT_LIMIT), 1) * 100",
"legendFormat": "{{Hostname}}",
"refId": "A"
}
],
"title": "Power draw per node (% of TDP)",
"description": "GPU power usage as a fraction of the node's combined TDP cap. Works across GPU SKUs without per-model thresholds. Requires DCGM_FI_DEV_POWER_MGMT_LIMIT from custom DCGM CSV (see packages/system/gpu-operator/examples/dcgm-custom-metrics.yaml).",
"transparent": false,
"datasource": {
"type": "prometheus",
"uid": "$ds_prometheus"
},
"gridPos": {
"h": 8,
"w": 12,
"x": 12,
"y": 6
},
"repeatDirection": "h",
"options": {
"displayMode": "gradient",
"valueMode": "color",
"namePlacement": "auto",
"showUnfilled": true,
"sizing": "auto",
"minVizWidth": 8,
"minVizHeight": 16,
"legend": {
"displayMode": "list",
"placement": "bottom",
"showLegend": false,
"calcs": []
},
"reduceOptions": {
"calcs": [
"lastNotNull"
]
},
"maxVizHeight": 300,
"orientation": "horizontal"
},
"fieldConfig": {
"defaults": {
"unit": "percent",
"min": 0,
"max": 100,
"thresholds": {
"mode": "absolute",
"steps": [
{
"value": null,
"color": "green"
},
{
"value": 60,
"color": "yellow"
},
{
"value": 80,
"color": "orange"
}
]
}
},
"overrides": []
}
},
{
"type": "row",
"collapsed": false,
"title": "Usage trends",
"gridPos": {
"h": 1,
"w": 24,
"x": 0,
"y": 14
},
"id": 20,
"panels": []
},
{
"type": "timeseries",
"id": 21,
"targets": [
{
"expr": "avg by (Hostname) (DCGM_FI_DEV_GPU_UTIL)",
"legendFormat": "{{Hostname}}",
"refId": "A"
}
],
"title": "GPU utilization per node (trend)",
"description": "Per-node NVML utilization over time — shows shifting workload across hosts.",
"transparent": false,
"datasource": {
"type": "prometheus",
"uid": "$ds_prometheus"
},
"gridPos": {
"h": 8,
"w": 12,
"x": 0,
"y": 15
},
"repeatDirection": "h",
"options": {
"legend": {
"displayMode": "table",
"placement": "bottom",
"showLegend": false,
"calcs": [
"mean",
"max"
]
},
"tooltip": {
"mode": "multi",
"sort": ""
}
},
"fieldConfig": {
"defaults": {
"unit": "percent",
"min": 0,
"max": 100,
"custom": {
"drawStyle": "line",
"lineInterpolation": "linear",
"fillOpacity": 20,
"showPoints": "never"
}
},
"overrides": []
}
},
{
"type": "timeseries",
"id": 22,
"targets": [
{
"expr": "sum by (Hostname) (DCGM_FI_DEV_POWER_USAGE)",
"legendFormat": "{{Hostname}}",
"refId": "A"
}
],
"title": "Power draw per node (trend)",
"description": "Per-node GPU power consumption over time.",
"transparent": false,
"datasource": {
"type": "prometheus",
"uid": "$ds_prometheus"
},
"gridPos": {
"h": 8,
"w": 12,
"x": 12,
"y": 15
},
"repeatDirection": "h",
"options": {
"legend": {
"displayMode": "table",
"placement": "bottom",
"showLegend": false,
"calcs": [
"mean",
"max"
]
},
"tooltip": {
"mode": "multi",
"sort": ""
}
},
"fieldConfig": {
"defaults": {
"unit": "watt",
"custom": {
"drawStyle": "line",
"lineInterpolation": "linear",
"fillOpacity": 20,
"showPoints": "never"
}
},
"overrides": []
}
},
{
"type": "row",
"collapsed": false,
"title": "Health",
"gridPos": {
"h": 1,
"w": 24,
"x": 0,
"y": 23
},
"id": 30,
"panels": []
},
{
"type": "stat",
"id": 31,
"targets": [
{
"expr": "count((max by (Hostname, gpu) (DCGM_FI_DEV_XID_ERRORS)) \u003e 0) or vector(0)",
"refId": "A"
}
],
"title": "GPUs with XID errors",
"description": "Count of GPUs reporting a non-zero XID code. XID = NVIDIA hardware/driver error — any non-zero value needs investigation.",
"transparent": false,
"datasource": {
"type": "prometheus",
"uid": "$ds_prometheus"
},
"gridPos": {
"h": 5,
"w": 6,
"x": 0,
"y": 24
},
"repeatDirection": "h",
"options": {
"graphMode": "none",
"colorMode": "background",
"justifyMode": "auto",
"textMode": "value",
"wideLayout": true,
"showPercentChange": false,
"reduceOptions": {
"calcs": [
"lastNotNull"
]
},
"percentChangeColorMode": "standard",
"orientation": ""
},
"fieldConfig": {
"defaults": {
"unit": "short",
"thresholds": {
"mode": "absolute",
"steps": [
{
"value": null,
"color": "green"
},
{
"value": 1,
"color": "red"
}
]
}
},
"overrides": []
}
},
{
"type": "stat",
"id": 32,
"targets": [
{
"expr": "avg(gpu:power_throttle_fraction:rate5m)",
"refId": "A"
}
],
"title": "Power throttling (cluster avg)",
"description": "Fraction of time the fleet spent power-throttled. \u003e5% = billed FLOPS are being underdelivered.",
"transparent": false,
"datasource": {
"type": "prometheus",
"uid": "$ds_prometheus"
},
"gridPos": {
"h": 5,
"w": 6,
"x": 6,
"y": 24
},
"repeatDirection": "h",
"options": {
"graphMode": "none",
"colorMode": "background",
"justifyMode": "auto",
"textMode": "value",
"wideLayout": true,
"showPercentChange": false,
"reduceOptions": {
"calcs": [
"lastNotNull"
]
},
"percentChangeColorMode": "standard",
"orientation": ""
},
"fieldConfig": {
"defaults": {
"unit": "percentunit",
"decimals": 2,
"thresholds": {
"mode": "absolute",
"steps": [
{
"value": null,
"color": "green"
},
{
"value": 0.05,
"color": "yellow"
},
{
"value": 0.2,
"color": "red"
}
]
}
},
"overrides": []
}
},
{
"type": "stat",
"id": 33,
"targets": [
{
"expr": "avg(gpu:thermal_throttle_fraction:rate5m)",
"refId": "A"
}
],
"title": "Thermal throttling (cluster avg)",
"description": "Fraction of time the fleet spent thermal-throttled. Indicates cooling/datacenter issues.",
"transparent": false,
"datasource": {
"type": "prometheus",
"uid": "$ds_prometheus"
},
"gridPos": {
"h": 5,
"w": 6,
"x": 12,
"y": 24
},
"repeatDirection": "h",
"options": {
"graphMode": "none",
"colorMode": "background",
"justifyMode": "auto",
"textMode": "value",
"wideLayout": true,
"showPercentChange": false,
"reduceOptions": {
"calcs": [
"lastNotNull"
]
},
"percentChangeColorMode": "standard",
"orientation": ""
},
"fieldConfig": {
"defaults": {
"unit": "percentunit",
"decimals": 2,
"thresholds": {
"mode": "absolute",
"steps": [
{
"value": null,
"color": "green"
},
{
"value": 0.05,
"color": "yellow"
},
{
"value": 0.2,
"color": "red"
}
]
}
},
"overrides": []
}
},
{
"type": "stat",
"id": 34,
"targets": [
{
"expr": "max(DCGM_FI_DEV_GPU_TEMP)",
"refId": "A"
}
],
"title": "Max GPU temperature",
"description": "Hottest GPU in the fleet right now. A10 sustained target \u003c83°C.",
"transparent": false,
"datasource": {
"type": "prometheus",
"uid": "$ds_prometheus"
},
"gridPos": {
"h": 5,
"w": 6,
"x": 18,
"y": 24
},
"repeatDirection": "h",
"options": {
"graphMode": "none",
"colorMode": "background",
"justifyMode": "auto",
"textMode": "value",
"wideLayout": true,
"showPercentChange": false,
"reduceOptions": {
"calcs": [
"lastNotNull"
]
},
"percentChangeColorMode": "standard",
"orientation": ""
},
"fieldConfig": {
"defaults": {
"unit": "celsius",
"thresholds": {
"mode": "absolute",
"steps": [
{
"value": null,
"color": "green"
},
{
"value": 75,
"color": "yellow"
},
{
"value": 85,
"color": "red"
}
]
}
},
"overrides": []
}
},
{
"type": "timeseries",
"id": 35,
"targets": [
{
"expr": "DCGM_FI_DEV_GPU_TEMP",
"legendFormat": "{{Hostname}}/{{gpu}}",
"refId": "A"
}
],
"title": "Temperature per GPU",
"description": "Per-GPU temperature trace. Sustained red zone = cooling issue or aggressive workload.",
"transparent": false,
"datasource": {
"type": "prometheus",
"uid": "$ds_prometheus"
},
"gridPos": {
"h": 8,
"w": 12,
"x": 0,
"y": 29
},
"repeatDirection": "h",
"options": {
"legend": {
"displayMode": "table",
"placement": "bottom",
"showLegend": false,
"calcs": [
"mean",
"max"
]
},
"tooltip": {
"mode": "multi",
"sort": ""
}
},
"fieldConfig": {
"defaults": {
"unit": "celsius",
"min": 20,
"max": 100,
"thresholds": {
"mode": "absolute",
"steps": [
{
"value": null,
"color": "green"
},
{
"value": 75,
"color": "yellow"
},
{
"value": 85,
"color": "red"
}
]
},
"custom": {
"drawStyle": "line",
"lineInterpolation": "linear",
"fillOpacity": 10,
"showPoints": "never"
}
},
"overrides": []
}
},
{
"type": "timeseries",
"id": 36,
"targets": [
{
"expr": "gpu:power_throttle_fraction:rate5m",
"legendFormat": "power {{Hostname}}/{{gpu}}",
"refId": "A"
},
{
"expr": "gpu:thermal_throttle_fraction:rate5m",
"legendFormat": "thermal {{Hostname}}/{{gpu}}",
"refId": "B"
}
],
"title": "Throttle fraction per GPU",
"description": "Red = power throttle, blue = thermal throttle. Both on same chart to compare causes.",
"transparent": false,
"datasource": {
"type": "prometheus",
"uid": "$ds_prometheus"
},
"gridPos": {
"h": 8,
"w": 12,
"x": 12,
"y": 29
},
"repeatDirection": "h",
"options": {
"legend": {
"displayMode": "table",
"placement": "bottom",
"showLegend": false,
"calcs": [
"mean",
"max"
]
},
"tooltip": {
"mode": "multi",
"sort": ""
}
},
"fieldConfig": {
"defaults": {
"unit": "percentunit",
"min": 0,
"max": 1,
"custom": {
"drawStyle": "line",
"lineInterpolation": "linear",
"fillOpacity": 25,
"showPoints": "never"
}
},
"overrides": []
}
},
{
"type": "row",
"collapsed": false,
"title": "Hardware inventory",
"gridPos": {
"h": 1,
"w": 24,
"x": 0,
"y": 37
},
"id": 40,
"panels": []
},
{
"type": "table",
"id": 41,
"targets": [
{
"expr": "group by (Hostname, modelName, gpu, UUID) (DCGM_FI_DEV_GPU_UTIL)",
"instant": true,
"range": false,
"format": "table",
"refId": "A"
}
],
"title": "GPU inventory",
"description": "One row per physical GPU. Model / index / UUID sourced from DCGM labels.",
"transparent": false,
"datasource": {
"type": "prometheus",
"uid": "$ds_prometheus"
},
"gridPos": {
"h": 6,
"w": 24,
"x": 0,
"y": 38
},
"repeatDirection": "h",
"transformations": [
{
"id": "organize",
"options": {
"excludeByName": {
"Time": true,
"Value": true,
"__name__": true,
"cluster": true,
"container": true,
"device": true,
"endpoint": true,
"instance": true,
"job": true,
"pci_bus_id": true,
"prometheus": true,
"service": true,
"tenant": true,
"tier": true,
"uid": true,
"unit": true
},
"indexByName": {
"Hostname": 0,
"UUID": 3,
"gpu": 2,
"modelName": 1
},
"renameByName": {
"Hostname": "Node",
"UUID": "UUID",
"gpu": "Index",
"modelName": "GPU Model"
}
}
}
],
"options": {
"frameIndex": 0,
"showHeader": true,
"showTypeIcons": false,
"footer": {
"show": false,
"reducer": []
},
"cellHeight": "sm"
}
}
],
"templating": {
"list": [
{
"type": "datasource",
"name": "ds_prometheus",
"label": "Prometheus",
"skipUrlSync": false,
"query": "prometheus",
"current": {
"selected": false,
"text": "default",
"value": "default"
},
"multi": false,
"allowCustomValue": true,
"includeAll": false,
"regex": "",
"auto": false,
"auto_min": "10s",
"auto_count": 30
}
]
},
"annotations": {}
}