ADR-0004: Scheduler Foundation

Status

Accepted

Context

The system targets multicore performance from the start but cannot afford scheduler complexity that blocks bring-up.

Decision

The first scheduler implementation uses:

• preemptive scheduling
• priority bands
• per-CPU run queues
• round-robin within priority bands
• SMP load balancing with affinity awareness

The design must preserve an upgrade path to weighted fairness, topology awareness, and later real-time classes.

Consequences

Positive:

• implementation remains tractable for early milestones
• contention is reduced relative to a single global queue
• wake-up and balancing logic can evolve incrementally

Negative:

• fairness is approximate in early versions
• topology and energy awareness are deferred

Implementation notes

• starvation prevention via aging is required even in the first version
• scheduler tracing hooks are mandatory so later tuning is evidence-based
• preemption and timer behavior must be testable in QEMU