Expand description
Offset-addressed RDMA arena for the SSM-snapshot spill tier.
A minimal, synchronous transport over the CX7 verbs + RW-blade paging
peer protocol the KV overflow tier uses, but addressed by a flat byte
offset (snapshots are keyed by an opaque id → arena slot) rather than the
KV GroupKey/group_stride layout — reusing that layout would corrupt live
KV (its write_from_host asserts src.len()==group_bytes).
The paging peer is layout-agnostic (client sends total_bytes, the peer
registers ONE RW MR and serves base+offset), so a second peer instance on
its own port serves the snapshot arena with zero peer-side change. Each op is
drained to completion before returning (one blob ~64 MB, ~5–7 ms — the
spill/fault path is latency-, not throughput-critical), so the caller’s
SnapshotBlobStore::{put,get} contract (durable on return) holds.
Gathering the scattered per-layer SSM state into the contiguous blob and all
device-stream ordering already happen in SsmSnapshotPool::{spill_slot, fault_in_slot}; this transport only moves host bytes.
Structs§
- Rdma
Snapshot Arena - Placeholder when RDMA verbs / CUDA aren’t built.
connectalways errors, socratedependents degrade to the host-RAM tier;write/readare unreachable (a stub arena is never successfully constructed).