Native transport¶

Fjord carries numeric tensors between OpenFOAM ranks and ModelHost. Rune defines the binary frame, while Harbor applies atomic exchange semantics over a selected channel.

sequenceDiagram
    participant S as OpenFOAM rank
    participant H as Harbor
    participant M as ModelHost
    S->>H: Announce identity, tensor count and bytes
    H->>M: Reserve complete request
    S->>M: Publish tensor frames
    S->>M: Complete request
    M->>S: Publish validated atomic response

Exchange contract¶

Identity field	Purpose
Exchange index	Orders repeated calls from one rank
Solver time index	Associates the request with solver progression
Physical time	Preserves the OpenFOAM time value
Rank	Identifies the rank-local field partition
Topology generation	Rejects responses produced for stale mesh storage
Tensor count and bytes	Supports admission before body allocation

Before tensor bodies, the solver announces exchange identity, tensor count, and total payload bytes. ModelHost reserves the complete request, then accepts the bodies. Completion commits only when count, size, shape, and identity match.

Outputs use the same atomic boundary. Every tensor is validated, floating outputs must be finite, and publication credit remains held until the receiver accepts the complete result.

Rune carries the solver-owned topology generation in request announcements, tensors and completion markers. Harbor requires one generation throughout an atomic batch, and the OpenFOAM adapter rejects stale responses before writing any solver-owned destination.

Channels¶

Channel	Intended use
Unix socket	local control and node-local fallback
Shared memory	normal same-node bulk payload path
TCP	explicit portable cross-node path
UCX	optional high-performance native path

Capability negotiation selects only a channel supported by both peers. An explicit request fails if unavailable; it does not silently change transport.

Memory ownership¶

Memory	Owner until
Receive credit	Complete request is rejected or consumed
Input frames	ModelHost accepts and validates the request
Execution workspace	Evaluation and output validation finish
Output publication	Solver accepts the complete response
Retained workspace	Capacity is needed by an admissible request or the host exits

Receive buffers, active execution storage, retained workspace capacity, and output publication each have an explicit owner and bound. Retained storage is a cache and is released when it would block new work. Cancellation wakes admission waiters and returns credits.

Each solver rank has an independent exchange. A node-local ModelHost serves only ranks on its node, keeping shared memory and model state node-local.