Crate nexar_rdma 

RDMA and GPUDirect transport extensions for nexar.

This crate provides InfiniBand/RoCE kernel-bypass transport as an extension to nexar’s QUIC-based communication. It uses the extension slot on PeerConnection to attach RDMA state without requiring any #[cfg] flags in the core nexar crate.

Features

• default — RDMA transport via ibverbs (InfiniBand/RoCE)
• gpudirect — GPUDirect RDMA: NIC reads/writes GPU memory directly

Usage

use nexar_rdma::ext::PeerConnectionRdmaExt;

// After bootstrap, each rank calls establish_rdma_mesh:
nexar_rdma::bootstrap::establish_rdma_mesh(&client).await;

// Send via RDMA with QUIC fallback:
peer.send_raw_rdma(data).await?;

Re-exports

pub use ext::PeerConnectionRdmaExt;

pub use rdma::RdmaConnection;

pub use rdma::RdmaContext;

pub use rdma::RdmaEndpoint;

pub use rdma::RdmaMemoryPool;

pub use rdma::RdmaMr;

Modules

bootstrap

RDMA and GPUDirect mesh establishment.

client_ext

Extension traits for NexarClient that add GPU-direct send/recv.

ext

Extension traits for PeerConnection that add RDMA capabilities.

rdma