The problem

When two CPUs are available, which one is actually best for the task?

What current research shows

Linux's capacity-aware scheduling documentation describes placement across CPUs with different capacities and uses utilization and capacity signals to avoid obviously unsuitable placements. NUMA adds another dimension: a CPU can be available while its nearby memory and cache relationships are unfavorable for the task.

Where the evidence stops

Placement quality depends on hardware topology, memory policy, workload behavior, and competing tasks. A placement rule that helps one machine can hurt another.

What Valen Systems is testing

Build topology-aware receipts that record CPU, cache domain, NUMA node, cooperating threads, and measured completion cost for the same work.

Sources

Linux scheduler: capacity-aware schedulingLinux scheduler debugfs: NUMA balancing