DPS Sample Dashboards#

DPS provides portable Grafana dashboard JSON files. They do not install Prometheus, Grafana, or a scrape configuration.

First configure collection through Prometheus Monitoring. Then import a JSON file in Grafana and select a Prometheus datasource for ${DS_PROMETHEUS}. The dashboards avoid hard-coded namespaces and cluster addresses.

Note

dps-sdk-dashboard.json is a scaffold with a placeholder panel and no metric query. It is not an operational sample dashboard.

Import the Samples#

In Grafana, select Dashboards > New > Import, upload a JSON file, and choose the Prometheus datasource for ${DS_PROMETHEUS}. Preserve the dashboard variables; they scope queries to a topology, domain, resource group, or node. The DPS Server Health Dashboard also requires configured log and trace data sources.

The Datacenter Dashboard uses marcusolsson-treemap-panel and ae3e-plotly-panel for treemap and grid views. Install these plugins when those panels are required. Basic Prometheus panels do not require the optional alerting datasource.

DPS Server Health Dashboard#

dps-server-trace.json provides DPS server availability and request-health signals. It includes replica availability, restart occurrence, management API latency and errors, method response and activation timing, and DPS control-loop latency, convergence, terminal outcomes, and tracker resets.

The following example shows the server-health and control-loop view.

DPS Server Health Dashboard with replica availability, restart occurrence, management API latency, control-loop latency and convergence, terminal outcomes, and tracker resets

Datacenter Dashboard#

dps-sdk-datacenter.json provides the fleet-level view. It refreshes every 10 seconds and filters by topology and domain.

Area	What it shows	Main metrics
Power states	Capacity, provisioned load, observed power, and active target	`topology_service_device_power_capacity_watts`, `topology_service_device_instance_power_load_watts`, `device_total_power_watts`, `nvgrid_system_status_load_target_watts`
Operating KPIs	Target utilization, provisioned-load coverage, pressure, stranded power, active-fleet coverage, and domains above target	Power-state metrics plus `nvgrid_system_status_calculated_load_watts`, `topology_service_device_instance_info`, and `resource_group_devices`
Redfish health	Failed-request count, failure rate, and failures by class and route	`redfish_request_duration_milliseconds_count`
Controller health	DPS and controller latency, convergence, and scheduler errors	DPS control-loop and controller duration and error series
Inventory and grid	Domain ratios, topology entities, resource-group membership, and grid schedules	`nvgrid_feed_schedule_load_target_watts` and topology-service inventory metrics

The following example shows the power-state trend, data center KPIs, and controller latency and convergence panels.

Datacenter Dashboard overview with power-state trends, data center KPIs, and controller latency and convergence panels

The following example focuses on Redfish request health. It shows failed requests, the failure rate, failures grouped by class and route, and the related controller latency and convergence signals.

Datacenter Dashboard Redfish request health with failed request count, failure rate, failures grouped by class and route, and controller latency and convergence signals

The following example shows a selected-domain drill-down with power ratios, utilization, resource-group membership, normalized power states, and device power distribution.

Datacenter Dashboard domain drill-down with power ratios, utilization, member nodes, normalized power states, and device power distribution

Resource Group Dashboard#

dps-sdk-resource-group.json is the workload-group view. It refreshes every 10 seconds and filters by topology, resource group, and member node.

Area	What it shows
Summary	Total and active resource groups, membership, and fleet observed-versus-provisioned power.
Power behavior	Observed power, observed-versus-provisioned deviation, average node provisioned limit, and 30-second provisioned-load changes.
Tables	Power and utilization by resource group, group configuration, and nodes above provisioned load.

It combines resource_group_count, topology_service_resource_group_active_count, resource_group_devices, device_total_power_watts, and topology-service provisioned-load metrics.

The following example shows the total and active resource-group summary, fleet observed-versus-provisioned power, observed-power and deviation trends, average provisioned limits, and 30-second provisioned-load changes.

Resource Group Dashboard with resource-group summary, observed-power and deviation trends, provisioned limits, and provisioned-load changes

MaxP Window Comparison Dashboard#

dps-sdk-maxp-comparison.json compares an operator-selected Grafana time range with an earlier range. Select a topology, the Grafana time range, baseline_offset, and the active-GPU threshold before interpretation.

It compares power envelope, active GPU counts, utilization, headroom, node power, and rack-level changes using machine_nvm_avg_power_budget_watts, DCGM_FI_DEV_POWER_USAGE, and DCGM_FI_DEV_GPU_UTIL.

Warning

This dashboard is an operator-selected diagnostic comparison. It does not create or retain release evidence. Record the time ranges, topology, and workload conditions separately.

The following example shows the diagnostic comparison view with comparison and current power envelopes, active GPU counts, utilization, headroom, and node-power panels.

MaxP Window Comparison Dashboard with comparison and current power envelopes, active GPUs, utilization, headroom, and node-power panels

Optional Alert Rules#

sdk-datacenter-alert-rules.yaml is an example Grafana alert rule file, not a dashboard-import manifest. It assumes a datasource UID named prometheus. Adapt the datasource, labels, routing, and thresholds before provisioning it.