The problem

When should data move toward a task, and when should the task move toward data?

What current research shows

Linux documents that remote memory access has a performance penalty and that NUMA policy and automatic balancing use page-fault-based observation to influence placement. Shared memory can be accessed by tasks on different nodes, which complicates a single local answer.

Where the evidence stops

Migration itself costs bandwidth and time, and shared data may not have one ideal node.

What Valen Systems is testing

Compare task migration, page migration, remote access, fault overhead, and completion cost for the same data-sharing graph.

Sources

Linux scheduler debugfs: NUMA balancingLinux scheduler: capacity-aware scheduling