Alauda Build of HAMi

Monitor HAMi Usage

HAMi usage can be observed from ACP monitoring and, when installed, HAMi WebUI.

TOC

ACP dashboardsHAMi WebUIHAMi Ascend soft-slice metricsMetrics source

ACP dashboards

Use ACP dashboards to view accelerator metrics together with other cluster and workload metrics.

  1. Open Administrator -> Operations Center -> Monitor -> Dashboards.
  2. Select the HAMi or accelerator dashboard delivered for your environment.
  3. Check workload usage, device usage, and exporter scrape status.

Dashboard management belongs to ACP monitoring. For dashboard import and management, see Manage Dashboards.

For the HAMi NVIDIA monitoring path, this documentation site supplies a MonitorDashboard resource. Download hami-vgpu-metrics-dashboard-v1.0.2.yaml to a machine with access to the business cluster, then apply it:

kubectl apply -f hami-vgpu-metrics-dashboard-v1.0.2.yaml

The dashboard consumes HAMi and NVIDIA metrics. Install and verify NVIDIA metrics components through the NVIDIA GPU product documentation; the dashboard manifest does not install those components.

HAMi WebUI

If Alauda Build of HAMi-WebUI is installed, open {platform-url}/clusters/{cluster-name}/hami-webui. For Prometheus authentication, NodePort, and backend query validation, see Install HAMi-WebUI.

Use HAMi WebUI when you need a HAMi-specific resource overview. Use ACP monitoring when you need platform-wide dashboards, alerting, and metric management.

HAMi Ascend soft-slice metrics

HAMi Ascend Device Plugin v1.4.0 provides built-in Pod and container memory metrics for soft slicing. Enable ServiceMonitor in the HAMiAscendDevicePlugin instance to make these metrics available to ACP monitoring.

For HAMi-managed whole-card workloads, use the Ascend NPU Exporter monitoring path described below. HAMi-managed hard-slice workloads do not currently provide workload-level metrics. Soft slicing provides Pod and container memory metrics. The compute utilization shown for a soft-slice workload is the utilization of the underlying physical device, so workloads sharing that device report the same value.

Metrics source

HAMi does not install vendor metrics components.

  • For NVIDIA backend metrics, use Alauda Build of DCGM-Exporter and follow the NVIDIA GPU metrics documentation for installation and verification. DCGM-Exporter is an NVIDIA GPU monitoring component, not a HAMi component.
  • For Ascend backend metrics, Alauda Build of NPU Operator manages NPU Exporter when the exporter component is enabled in the NPUOperatorCtl instance. NPU Exporter is an Ascend NPU monitoring component, not a HAMi component. Use the Ascend NPU monitoring documentation for exporter verification and dashboard integration.

Use the backend pages to confirm which component owns the metrics path: