Module rdma_snapshot

metrale_storage

Module rdma_snapshot 

Source
Expand description

2026-09-25: RdmaSnapshotArena, the synchronous RDMA transport of the SSM snapshot spill tier, addressed by byte offset or, in paging mode, by key.

Owner: storage, SSM snapshot tier. Invariants:

  • Every transfer holds the arena’s lock throughout, and returns Ok only after its RDMA work has completed.
  • Without the cuda feature and metrale_rdma_verbs, both constructors fail.

Both modes send the v2 paging header with kind SSM. Raw mode (connect, which sends blob_bytes == 0) gets a private arena and the caller chooses offsets. Paging mode (connect_paging) lets the peer own residency, driven by ALLOC, COMMIT, GET and REMOVE on the TCP stream. This transport moves host bytes only; SsmSnapshotPool::spill_slot and fault_in_slot in metrale-model-engine gather the state and order the device streams.

Structs§

RdmaSnapshotArena
2026-09-25: Stand-in when RDMA verbs or CUDA are not built. Both constructors fail; the other methods panic.