ModelArts Metric Collector

Description

ModelArts Metric Collector, a default built-in plug-in of ModelArts, runs as a node daemon to collect node and job metrics and report them to AOM. For details about the metrics, see Viewing Lite Cluster Metrics on AOM.

Figure 1 ModelArts Metric Collector
Click to enlarge

Constraints

The plug-in is automatically installed during resource pool creation and cannot be uninstalled.
This plug-in is automatically installed if ModelArts Node Agent is upgraded to the latest version for an existing resource pool.
During the plug-in upgrade, the pod for metric collection restarts. As a result, metrics may not be reported for a short period of time. Exercise caution when performing the operation.

Components

Component	Description	Resource Type
modelarts-metric-collector	Node and container metrics collection	DaemonSet

Parameters

Parameter	Description
Standby Node Metric Reporting	Whether the standby node of a dedicated pool reports metrics. The default value is false.
Enable Exporter	Third-party monitoring systems such as Prometheus can obtain metrics collected by ModelArts. If this function is disabled, third-party monitoring systems such as Prometheus cannot collect metrics. This function is enabled by default. Dedicated pool: Enable this function if you want to use inference job metrics for scaling.
Report Metrics to a Custom Common Prometheus Instance on AOM	By default, metrics are reported to the Prometheus_AOM_Default instance of AOM. If this function is enabled, metrics are reported to the custom Prometheus common instance, as shown in Figure 2. If this function is disabled, metrics are reported to the default Prometheus instance, that is, the Prometheus_AOM_Default instance, as shown in Figure 3.

Parameter

Description

Standby Node Metric Reporting

Whether the standby node of a dedicated pool reports metrics. The default value is false.

Enable Exporter

Third-party monitoring systems such as Prometheus can obtain metrics collected by ModelArts. If this function is disabled, third-party monitoring systems such as Prometheus cannot collect metrics. This function is enabled by default.

Dedicated pool: Enable this function if you want to use inference job metrics for scaling.

Report Metrics to a Custom Common Prometheus Instance on AOM

By default, metrics are reported to the Prometheus_AOM_Default instance of AOM.

If this function is enabled, metrics are reported to the custom Prometheus common instance, as shown in Figure 2. If this function is disabled, metrics are reported to the default Prometheus instance, that is, the Prometheus_AOM_Default instance, as shown in Figure 3.

Figure 2 Custom Prometheus common instance
Click to enlarge