Skip to main content

Health Checks Documentation

Comprehensive documentation for all GCM health checks. Health checks validate system health, hardware functionality, and configuration correctness across compute nodes.

Overview​

Health checks are organized into the following categories:

Quick Reference​

CheckCategoryPurpose
check-airstoreStorageValidates Flash Array credential configuration
check-authenticationSecurityVerifies password status and file access permissions
check-blockdevStorageMonitors NVMe device health via SMART data
check-dcgmiGPUNVIDIA DCGM diagnostics and NVLink validation
check-ethlinkNetworkValidates Ethernet interface configuration and state
check-hcaNetworkVerifies InfiniBand HCA device count
check-ibNetworkInfiniBand link health and performance validation
check-ipmitoolHardwareSystem Event Log (SEL) analysis for hardware errors
check-ncclGPUNCCL collective operation performance testing
check-nodeSystemNode uptime, kernel modules, and package repositories
check-nvidia-smiGPUComprehensive GPU health validation via NVML (including clock policy drift checks)
check-pciHardwarePCI device presence and PCIe link validation
check-processProcessProcess existence and state validation
check-processorCPUCPU/processor configuration validation
check-sensorsHardwareFan and PSU sensor monitoring via IPMI
check-serviceSystemService status and package version validation
check-ssh-certsSecuritySSH certificate validation against IPA
check-storageStorageDisk usage, mounts, and file system validation
check-syslogsSystemSystem log analysis for hardware and network errors
check-telemetryUtilityTelemetry publishing test utility
memtestGPUGPU memory integrity testing

Exit Codes​

All health checks follow a standard exit code convention:

| Exit Code | Description | |-----------|--------|-------------| | OK (0) | Check passed successfully | | WARN (1) | Non-critical issues detected | | CRITICAL (2) | Critical issues detected | | UNKNOWN (3) | Check could not complete |

Command-Line Options​

Common Options:

OptionTypeDefaultDescription
--timeoutInteger300Command execution timeout in seconds
--sinkStringdo_nothingTelemetry sink destination
--sink-optsMultiple-Sink-specific configuration
--verbose-outFlagFalseDisplay detailed output
--log-levelChoiceINFODEBUG, INFO, WARNING, ERROR, CRITICAL
--log-folderString/var/log/fb-monitoringLog directory
--heterogeneous-cluster-v1FlagFalseEnable heterogeneous cluster support

Check-specific options are documented in each health check's page.

Feature Flag Disabled​

If a check is disabled via feature flag, it will return OK immediately without performing validation. Check feature flag configuration if unexpected.

Contributing​

When adding new health checks:

  1. Implement in gcm/gcm/health_checks/checks/
  2. Create documentation following the format of existing checks
  3. Add entry to this README in appropriate category
  4. Include usage examples
  5. Document all command-line options and exit conditions