This feels racy, and I'm not sure it will work on multi-node NUMA architectures. It's one thing to delay and hope that your write is visible in the local proces

This feels racy, and I'm not sure it will work on multi-node NUMA architectures. It's one thing to delay and hope that your write is visible in the local processor cache; it's another to ensure that write is visible to other processor nodes, when there's no explicit cache coherence rule that says it will be.