CLI reference

Synopsis

kuberina <COMMAND>

Declared name: kuberina. Description: Maritime stowage-inspired K8s scheduling optimizer.

Commands

`plan`

Generate an optimised scheduling blueprint.

kuberina plan --infra <PATH> --workloads <PATH> [--pareto <PERCENT>]

Flag	Type	Required	Description
`--infra`	path	yes	Cluster topology YAML — node capacity, taints, labels
`--workloads`	path	yes	Workload manifests YAML — pods and their resource requirements
`--pareto`	float	no	Scale node capacity by this percentage while solving, e.g. `80`

plan is currently the only subcommand.

Output

The blueprint is printed to the console and written to kuberina_solution.yaml in the working directory.

When --pareto is supplied, solving happens against the scaled capacities, but the printed blueprint reports against the true net capacities — so the numbers you read are real, not the derated ones.

Invoking it through cargo

The Make targets wrap cargo run, which needs -- to separate cargo’s own arguments from the program’s:

cd solver
cargo run --release -- plan --infra testdata/homelab_infra.yaml \
                            --workloads testdata/homelab_workloads.yaml

Solver behaviour

Automatic GA sizing

The genetic algorithm’s parameters are selected from the problem size rather than configured by flag. The tiers are:

Pods	Population	Max generations	Notes
Under 100	128	—	Quick convergence
100–500	256	500	Standard GA, early stop after 100
Over 500	1,024	1,000	Datacenter-scale; early stop after 200

At datacenter scale the solver announces the switch on stderr:

Datacenter-scale detected (2714 pods) — cranking GA to maximum

Large-tier tuning also sets tournament size 5, mutation rate 0.03, crossover rate 0.85, and a fixed random seed of 42 — so runs are reproducible.

Pipeline phases

A plan invocation runs the three-phase hybrid pipeline described in the paper, preceded by a daemonset pre-deduction pass:

Phase	Step	Purpose
0	Daemonset pre-deduction	Subtract per-node daemonset overhead from capacity before packing
1	Vector Packing FFD	First-Fit Decreasing warm start, giving the GA a sane initial population
2	Genetic Algorithm	Evolutionary optimisation with gang-aware repair
3	CSP forward checking	Constraint enforcement, guaranteeing a feasible result