This is gone over in the talk. We allocate the page locally to where the data is used. The idea is that we'd prefer the NVME drive to eat any latency for the

This is gone over in the talk. We allocate the page locally to where the data is used. The idea is that we'd prefer the NVME drive to eat any latency for the NUMA bus transfer, and not have the CPU (SW TLS) or NIC (inline HW TLS) stall waiting for a transfer.