The problem

When does the cost of discovering memory locality outweigh the locality it recovers?

What current research shows

Linux documents that automatic NUMA balancing scans address spaces, unmaps pages, observes hinting faults, and may migrate data toward a local node. The documentation explicitly says the overhead isn’t universally guaranteed to be offset by improved locality.

Where the evidence stops

The answer depends on page access patterns, task placement, memory policy, scan rates, and whether the workload is already explicitly placed.

What Valen Systems is testing

Report scan work, hinting faults, migration volume, remote accesses, completion time, and energy together rather than presenting locality as free.

Sources

Linux scheduler debugfs: NUMA balancing