Skip to main content

AI Insight Overview

KakaoCloud AI Insight is a GPU monitoring service that lets you check the status and key metrics of GPU resources by cluster, node, and GPU. You can understand the overall GPU status at a glance, quickly identify GPUs with abnormal signs, and analyze causes on detailed pages.

AI Insight provides metrics required for GPU operations, such as GPU utilization, GPU memory usage, temperature, idle ratio, ECC Error, XID Event Code, and Throttling. You can monitor both Kubernetes Engine (KE)-based GPU nodes and Virtual Machine (VM)-based GPU nodes. In environments with MIG configured, you can also check status by MIG instance.

Note

To check metrics in AI Insight, Metric Exporter or the monitoring agent must be installed in the target environment. If it is not installed or is not working properly, the resource may appear in Agent Missing status and GPU metrics may not be collected.

Key Features

FeatureDescription
Check overall GPU statusCheck total GPUs, clusters, nodes, average GPU utilization, average memory usage, average temperature, and ECC Error count on the Overview page.
Check GPU statusCheck the number of GPUs in Active, Idle, Warning, Critical, Pending, and Agent Missing status.
GPU MapVisualize resources by GPU, cluster, or node and explore resources by status.
GPU ExplorerCheck detailed metrics and events by Cluster, Node, and GPU.
GPU event analysisIdentify abnormal causes using ECC Error, XID Event Code, Throttling, and Overheat information.
Check MIG instancesCheck utilization and status by instance for GPUs with MIG enabled.
Check node system metricsCheck CPU, memory, disk, and network metrics for VM or KE nodes.

GPU Status Criteria

AI Insight displays GPU status based on collected GPU metrics and node status. If multiple status conditions are met at the same time, the status with the highest severity is displayed first.

StatusDescription
ActiveNormal operating status where GPU compute or memory is in use.
IdleIdle status where both GPU compute and memory usage are low.
WarningStatus where abnormal signs are detected, such as increased GPU temperature, SBE ECC Error, or minor Thermal/Power Throttling.
CriticalStatus requiring immediate inspection, such as excessive GPU temperature increase, DBE ECC Error, severe Thermal Throttling, or Reliability Violation.
PendingStatus where the node containing the GPU is in an inactive lifecycle state, such as stopped, booting, rebooting, or resizing.
Agent MissingStatus where metrics cannot be collected because Metric Exporter or the monitoring agent is not installed or is not working properly.
Note

XID Event Code is displayed as an informational metric on the GPU details page. Currently, XID Event Code is not reflected in Warning or Critical status determination.

Console Menu

AI Insight consists of the following pages.

MenuDescription
OverviewPage for checking summary status, GPU counts by status, and GPU Map for all GPU resources.
GPU Explorer > ClusterPage for checking status, metrics, outliers, and correlations for GPU resources in a specific cluster.
GPU Explorer > NodePage for checking GPU status and node system metrics, such as CPU, memory, disk, and network, for a specific node.
GPU Explorer > GPUPage for checking detailed utilization, memory usage, temperature, idle ratio, Throttling, and ECC Error trends for an individual GPU or MIG instance.

Workflow

You can use AI Insight with the following flow.

  1. Install Metric Exporter or the monitoring agent in the target environment.
  2. Check the overall GPU status and GPU counts by status on the AI Insight Overview page.
  3. If Warning, Critical, or Agent Missing status exists, select the target resource from GPU Map or the list.
  4. In GPU Explorer, check detailed metrics and events by Cluster, Node, and GPU.
  5. Depending on the cause, check GPU temperature, ECC Error, Throttling, and node system resource status together. Use XID Event Code as reference information.

Prerequisites

AI Insight metric collection methods differ depending on the target environment.

Target EnvironmentRequired ConfigurationReference
Kubernetes EngineInstall Metric Exporter based on GPU Operator and DCGM ExporterMetric Exporter Installation
Virtual MachineInstall DCGM, DCGM Exporter, and the monitoring agent, and configure Prometheus inputMetric Exporter Installation
Caution

If Metric Exporter or the monitoring agent is not installed or is not working properly, metrics such as GPU utilization, GPU memory usage, temperature, and ECC Error are not collected.

DocumentDescription
Key ConceptsDescribes AI Insight components, GPU status, key metrics, and event metrics.
Metric Exporter InstallationDescribes how to install components for collecting GPU metrics in KE and VM environments.
Check Overall GPU StatusDescribes how to check the overall GPU status on the Overview page.
View GPU Resource DetailsDescribes how to view Cluster, Node, and GPU details.