baracuda-cufile

Safe Rust wrappers for NVIDIA cuFile — the GPUDirect Storage user-space library. Direct DMA between NVMe / network storage and GPU memory, no host-RAM bounce buffer.

Coverage

Comprehensive across the cuFile API:

• Driver lifecycle: Driver::open, Driver::close, Driver::properties, Driver::set_poll_mode.
• File-handle register: typed RAII FileHandle that registers with cuFile on construction and unregisters on drop.
• Buffer register: typed BufferRegistration for pre-registering device or pinned host buffers (avoids per-IO registration overhead).
• Sync read / write: read, write.
• Async read / write on a stream: read_async, write_async.
• BatchIO: BatchHandle::setup / submit / poll / cancel / destroy. Lets you submit dozens of IOs in one syscall, poll for completion, and overlap with kernel execution.
• Limits: configurable direct-I/O size, cache size, pinned-memory size — all via Driver::set_*.
• Error reporting: typed OpStatus with error_string helper.

use baracuda_cufile::{Driver, FileHandle, OpenFlags};
use baracuda_driver::{Context, Device, DeviceBuffer};

# fn demo() -> Result<(), Box<dyn std::error::Error>> {
let _driver = Driver::open()?; // RAII: closes on drop
let ctx = Context::new(&Device::get(0)?)?;
let mut buf: DeviceBuffer<u8> = DeviceBuffer::zeros(&ctx, 4 << 20)?;

let fh = FileHandle::open("dataset.bin", OpenFlags::ReadOnly)?;
let bytes_read = fh.read(&mut buf, 0)?;
# Ok(()) }

Platform support

cuFile / GPUDirect Storage is Linux-only. The loader fails fast with a clear error on Windows / macOS.

Pairs with baracuda-cufile-sys for the raw FFI surface.

Part of the baracuda workspace.

License

Dual MIT / Apache-2.0.