The problem

How can a decision that looks correct at one run queue create a bad outcome across the whole machine?

What current research shows

Linux scheduling combines local run-queue state with capacity, energy, utilization, affinity, topology, and user-space controls. Each signal describes part of the system, while the combined system also includes memory and device effects that may appear later.

Where the evidence stops

This isn’t evidence that heuristics are irrational; heuristics are necessary when the future is unknown and decisions must be made quickly.

What Valen Systems is testing

Build traceable decision receipts that connect placement, frequency, memory, device state, and final work completion so global effects can be studied rather than guessed.

Sources

Linux scheduler: capacity-aware schedulingLinux scheduler: Energy Aware SchedulingLinux scheduler: schedutil and PELT