Supported Slurm APIs¶
What ORB Uses¶
Primary: Power Save Hooks¶
| Interface | Direction | Purpose |
|---|---|---|
ResumeProgram |
Slurm → ORB | The Slurm Workload Manager by SchedMD calls this to power up nodes. ORB provisions cloud instances. |
SuspendProgram |
Slurm → ORB | Slurm calls this to power down nodes. ORB terminates instances. |
These are the only required Slurm integration points. Everything else is optional.
Optional: slurmrestd REST API (Monitoring Only)¶
Used for health checks and status reporting. ORB connects to slurmrestd as a read-only client.
| Endpoint | Method | Purpose |
|---|---|---|
/slurm/v0.0.44/diag |
GET | Health check — verify cluster is responsive |
/slurm/v0.0.44/nodes |
GET | List all nodes with current states |
/slurm/v0.0.44/node/{name} |
GET | Get details for a specific node |
/slurm/v0.0.44/partitions |
GET | List all partitions |
/slurm/v0.0.44/partition/{name} |
GET | Get details for a specific partition |
Authentication: JWT token via X-SLURM-USER-TOKEN header (if cluster requires it).
Optional: CLI Fallback¶
When slurmrestd is not available, ORB falls back to CLI commands for the same information:
| CLI Command | Equivalent REST Endpoint | Purpose |
|---|---|---|
scontrol ping |
/diag |
Health check |
sinfo -N -h -o "%N %T %P %c %m" |
/nodes |
List nodes with states |
scontrol show node {name} |
/node/{name} |
Node details |
sinfo -h -o "%P %a %l %D %C" |
/partitions |
List partitions |
scontrol show partition {name} |
/partition/{name} |
Partition details |
What ORB Does NOT Use¶
ORB is a resource provider, not a job scheduler. The following Slurm interfaces are intentionally NOT integrated:
| Interface | Reason for Exclusion |
|---|---|
sbatch |
Job submission — not ORB's responsibility |
srun |
Interactive job launch — not ORB's responsibility |
squeue |
Job queue inspection — ORB doesn't manage jobs |
scancel |
Job cancellation — ORB doesn't manage jobs |
sacctmgr |
Account management — outside ORB's scope |
salloc |
Resource allocation — Slurm handles this |
sprio |
Job priority — Slurm handles this |
| Job completion callbacks | ORB doesn't track individual jobs |
| Prolog/Epilog scripts | Node setup/teardown managed by Slurm |
Architecture Boundary¶
┌─────────────────────────────────────────────────────┐
│ SLURM Domain │
│ │
│ Jobs: sbatch, srun, squeue, scancel (NOT ORB) │
│ Scheduling: priority, fairshare, limits (NOT ORB) │
│ Accounting: sacctmgr, sreport (NOT ORB) │
│ │
│ ┌──────────────────────────────────────────────┐ │
│ │ Power Management Interface │ │
│ │ ResumeProgram / SuspendProgram │ │
│ └──────────────────────┬───────────────────────┘ │
└─────────────────────────┼────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────┐
│ ORB Domain │
│ │
│ Resource Provisioning: EC2Fleet, ASG, RunInstances │
│ Machine Lifecycle: launch, monitor, terminate │
│ Template Management: instance types, networking │
│ Health Monitoring: slurmrestd/CLI (read-only) │
│ │
└─────────────────────────────────────────────────────┘