Dynamic Load Balancing (DLB)
Dynamic Load Balancing redistributes eligible running workloads across AtomOS compute nodes to improve utilization. It is a placement mechanism — not high availability or automatic recovery from host failure.
Status: Preview — see Feature availability.
Purpose vs HA
| Concern | DLB | HA / recovery |
|---|---|---|
| Goal | Balance load among healthy nodes | Restore service after failure |
| Trigger | Utilization / policy / maintenance | Node loss, storage or network fault |
| Storage | Reachable storage for moves | Redundant design and runbooks |
| Outcome | Guest may live-migrate per policy | Guest restarted or re-pledged elsewhere |
Failure domains and live migration limits: Availability & recovery.
Eligibility
Guests are typically eligible when they:
- Run on compute cluster members with working peer connectivity
- Use storage and networks reachable from candidate destinations
- Do not use blockers such as GPU / PCI passthrough or incompatible CPU features
Exclude passthrough guests from automatic balancing until you validate otherwise.
Storage and migration prerequisites
DLB inherits live-migration constraints:
- Volume data accessible from source and destination
- No exclusive local-only disks unless you have a validated storage-motion procedure
- Working bridges and firewall paths between hosts
Validate vault layout in Storage before enabling automated moves.
Operational behavior
- Failed migrations surface as failed balancing actions — inspect CPU compatibility, networking, and compute journals.
- Put hosts into maintenance / drain (or disable the Electros target) before patching so new placements and moves avoid that node.
- Scoring inputs and thresholds are release-specific. Public docs do not publish CPU percentages or API fields; obtain parameters from release notes or staging observation.
Visibility
bash
sudo atomos-cli --status
sudo atomos-cli --healthcheck
sudo atomos-cli --journal compute 1hWatch Electros VM and target state during moves. Balancing increases network and storage load.