GPU Miscount Alert - Troubleshooting Guide

GPU Miscount Alert - Troubleshooting Guide

Purpose

Provide troubleshooting steps when a GPU Miscount alert is triggered, indicating that the system detects an incorrect number of GPUs on a node.


Alert Name

GPU Miscount


What the Alert Is

This alert indicates that the number of GPUs detected on the node does not match the expected GPU count.

The system monitoring tool compares the expected GPU count with the GPUs currently detected by the node.


Summary

Incorrect number of GPUs detected on the node.

Possible causes include:

  • GPU not detected by the OS

  • PCIe device detection issue

  • GPU driver issue

  • GPU allocation problem in Kubernetes (k3s)

  • Hardware failure


Severity

P3 – Medium Impact

The node may not fully utilize GPU resources, which can impact workloads and cluster scheduling.


L1 Engineer Actions

Follow these steps when the alert is received.

Step 1 – Log the Alert

  • Record the alert in the Alert Summary sheet


Step 2 – Create Ticket

  • Open an incident ticket for tracking


Step 3 – Escalation (If Required)

If the issue appears significant:

  • Create an internal ticket

  • Inform L2 Engineer


Step 4 – Documentation

After confirmation with L2:

  • Update the Issues Master Sheet


L2 Engineer Actions

L2 engineers should perform detailed checks on the node.

Step 1 – Check GPU Detection

Run:

nvidia-smi -L

Verify the number of GPUs detected by the system.

Step 2 – Check GPU Allocation from HGX side and from K3s Perspective from login node / K3s. 

Verify whether GPUs are properly allocated or restricted in k3s.

Check:

  • GPU allocation

  • GPU resource availability

  • Running workloads using GPUs

Example Command: for n in $(kubectl get nodes | awk '{print $1}'); do echo $n; kubectl describe node $n | grep Allocatable -A 10 | grep  nvidia.com/gpu | grep -v 8; done

Step 3 – Verify GPU Detection at PCI Level

Run:

lspci | grep -i nvidia

This confirms whether the operating system detects the GPUs at the PCIe hardware level.

Step 4 – Check GPU Driver Errors

Run:

dmesg | grep -i xid

Look for NVIDIA Xid errors, which may indicate:

  • GPU hardware failure

  • Driver crash

  • GPU reset events


Escalation

Escalate to platform/OEM support if:

  • GPUs are not visible in lspci

  • Persistent Xid errors occur

  • GPUs disappear after reboot

  • Hardware failure is suspected

    • Related Articles

    • GPU Memory Mismatch – Troubleshooting Guide

      Alert Information Alert Name: GPU Memory Mismatch Severity: P3 – Medium Impact Alert Description: This alert is generated when the detected GPU memory size on a node does not match the expected configuration. Alert Summary The monitoring system has ...

    • GPU Memory Temperature Critical - Troubleshooting Guide

      Purpose This document provides a standardized approach for handling and troubleshooting the GPU Memory Temperature Critical alert, including L1 and L2 actions, escalation procedures, and resolution criteria. Alert Name GPU Memory Temperature Critical ...

    • Network Down IB Alert - Troubleshooting Guide

      Purpose This document provides a standardized approach for handling and troubleshooting the Network Down – IB alert, including L1 and L2 actions, escalation process, and resolution criteria. Alert Name Network Down – IB Alert Description This alert ...

    • Nvidia-smi Tool Alert - Troubleshooting Guide

      Purpose This document provides a standardized approach for handling and troubleshooting the Nvidia-smi Tool alert, including L1 and L2 actions, escalation guidelines, and resolution criteria. Alert Name Nvidia-smi Tool Alert Description This alert is ...

    • ZFS Pool Capacity Warning Alert - Troubleshooting Guide

      Purpose This document provides a standardized approach for handling and troubleshooting the ZFS Pool Capacity Warning alert, helping L1 and L2 engineers prevent storage exhaustion and maintain system stability. Alert Name ZFS Pool Capacity Warning ...