A low-latency internode messaging scheme bypasses the nodes' I/O stacks to use fabrics or links that support memory process logic (e.g., SMI3) or electrical process logic (e.g., PCIe) on the “node side” between the nodes and a pooled memory controller (or pooled storage controller), and on the “pooled side” between that controller and its pooled memory or storage. The controller may translate and redirect messages and look up addresses. The approaches accommodate 2-level memory (locally attached node memory and accessible pooled memory) with either or both levels private, globally shared, allocated to a subset of the nodes, or any combination. Compatible interrupt schema use the messaging links and components.