Hi, A real workload has since run into the stall this series is about.
Mooncake, the KV cache store used with vLLM and SGLang, backs each rank's host segment with MADV_HUGEPAGE memory and registers it with the EFA NIC, which pins it long-term. Its commit message [1] says: "under the default local NUMA policy every THP fault on a full node runs direct compaction before falling back. With 8 ranks x 190GB this made registration take ~38 minutes." They worked around it by binding the segment MPOL_INTERLEAVE across all allowed nodes, so no node fills before the others. That gives up node-local placement, and a task that sets its own policy still gets the stall. [1] https://github.com/kvcache-ai/Mooncake/commit/c3fa13ecaf1038df933dcb784ff8ffe026381100 Thanks, Qiliang
