cobre-comm

Pluggable communication backend abstraction for the Cobre distributed solver ecosystem.

Defines the Communicator and SharedMemoryProvider traits that decouple distributed computations from specific communication technologies. Two backends are implemented: LocalBackend, a single-process no-op that is always available and carries zero overhead, and FerrompiBackend, an MPI 4.x implementation built on the ferrompi crate and enabled via the mpi Cargo feature. The backend is selected at build time through Cargo feature flags with an optional runtime override via the COBRE_COMM_BACKEND environment variable. All dispatch over the Communicator trait is static, so there is no dynamic dispatch overhead on the hot path.

When to Use

Depend on cobre-comm directly when you are writing a distributed algorithm that needs collective communication (broadcast, reduce, scatter/gather) and you want to test it locally without an MPI installation. Algorithm crates such as cobre-sddp depend on this crate and accept a generic Communicator parameter; you only need to depend here when adding a new algorithm crate or a new backend.

Key Types

• Communicator — the core trait for collective communication operations (broadcast, reduce, barrier) implemented by every backend
• LocalBackend — single-process no-op backend; always available, no external dependencies
• FerrompiBackend — MPI 4.x backend built on ferrompi; enabled with the mpi feature flag
• BackendKind — enum listing the available backends detected at compile time, used to select a backend via create_communicator
• SharedMemoryProvider — trait for intra-node memory-region allocation used by collocated processes

Links

Status

Alpha — API is functional but not yet stable.

License

Apache-2.0