Slurm Adapter¶

Reads HPC cluster state from Slurm via CLI commands, REST API, or mock data.

Configuration¶

CLI ModeREST ModeMock Mode

slurm:
  enabled: true
  mode: cli
  partitions:
    - gpu
    - compute

slurm:
  enabled: true
  mode: rest
  slurmrestd_url: http://slurmrestd:6820
  partitions:
    - gpu

slurm:
  enabled: true
  mode: mock
  mock_data_path: examples/telemetry-replay/slurm-node-health-warning/inputs/telemetry.yaml

Modes¶

Mode	Source	Use Case
`cli`	`squeue`, `sinfo`, `sacct` shell commands	Production Slurm clusters
`rest`	Slurm REST API (`slurmrestd`)	REST-enabled clusters
`mock`	YAML file	Development and testing

Collected Entities¶

Slurm Command	Entity Type	Key Fields
`squeue`	`job`	job_state, assigned_gpu_count, user
`sinfo`	`node`	health_state, cpu_utilization, partitions
`sacct`	`job`	elapsed_time, exit_code, gpu_utilization

CLI Commands Used¶

squeue --json                    # Job queue state
sinfo --json                     # Node state
sacct --json -S now-1hour        # Recent job accounting

GCM Integration¶

Meta GCM Compatibility

OpenAICE's Slurm adapter is designed to complement Meta's GPU Cluster Monitoring (GCM). GCM provides the data collection layer, while OpenAICE adds the policy and recommendation layer on top.