baracuda-nvshmem

Safe Rust wrappers for the NVIDIA NVSHMEM host API — the OpenSHMEM symmetric-heap, one-sided GPU communication model.

NVSHMEM is the fine-grained, one-sided complement to baracuda-nccl's collectives: every PE (typically one GPU) allocates from a symmetric heap at a shared virtual address, and any PE can put / get directly into another PE's heap. The two crates coexist — collectives via NCCL, one-sided RDMA via NVSHMEM — and a single program may use both.

use baracuda_nvshmem::Context;

let ctx = Context::init()?;                 // environment-bootstrapped
let me = ctx.my_pe();
let world = ctx.n_pes();

let local  = ctx.malloc::<f32>(1024)?;      // collective symmetric alloc
let remote = ctx.malloc::<f32>(1024)?;

// one-sided: push our `local` into PE (me+1)'s `remote`
let peer = (me + 1) % world;
ctx.put(&remote, &local, 1024, peer)?;
ctx.barrier_all()?;                         // remote completion + sync
# Ok::<(), baracuda_nvshmem::Error>(())

Surface (Tier 1)

• Context — init / finalize, cached my_pe / n_pes, version, and the barrier_all / sync_all / quiet / fence ordering primitives (plus *_on_stream variants).
• Team — split_strided, my_pe, n_pes, translate_pe, destroy, with predefined Team::WORLD / Team::SHARED.
• SymmetricBuffer<T> — typed symmetric-heap allocation (RAII free).
• Host-initiated RMA — put / get and their stream-ordered forms.
• UniqueId — the unique-id bootstrap seed (see the crate docs for the version-specific init wiring).

Not covered

The device-side API (nvshmem_int_p / nvshmem_putmem_nbi etc. called from inside a CUDA kernel) requires linking libnvshmem_device.a into the consumer's kernel binary and cannot be a lazily-loaded host symbol. Consumers that need it write their own .cu against the NVSHMEM headers.

Availability

NVSHMEM is Linux-only in practice and requires compute capability sm_70+ (every baracuda-supported GPU qualifies). On hosts without the runtime, Context::init returns LoaderError::LibraryNotFound so callers can fall back to single-process execution. NVSHMEM is not bundled — it is loaded at runtime from the user's installation.

Part of the baracuda workspace.

License

Dual MIT / Apache-2.0.