RDMAenabled

RDMAenabled describes hardware, software, or configurations that provide Remote Direct Memory Access functionality across nodes using RDMA protocols. It enables data transfer between memory regions with minimal CPU involvement, low latency, and high throughput.

RDMA uses specialized NICs and network protocols to transfer data directly between application buffers or memory regions without intermediate copies by the CPU. Applications register memory, allocate buffers, and use a verbs API to initiate transfers. The NIC handles data movement, often supporting zero-copy and scatter/gather.

RDMA protocols include InfiniBand, RoCE (RDMA over Converged Ethernet), and iWARP. InfiniBand provides a high-speed fabric commonly used in HPC. RoCE and RoCEv2 operate over Ethernet networks, with RoCE requiring lossless links or QoS, while iWARP runs RDMA over TCP/IP.

Hardware and software requirements include RDMA-capable NICs and switches, firmware, and drivers that expose the verbs API ( ibverbs, librdmacm ). Operating systems such as Linux or Windows must be configured for RDMA, including enabling modules and correct network settings (QoS, VLANs, MTU). Management tools and libraries are often used to configure connections, memory registration, and remote access policies.

Typical use cases include high-performance computing, distributed storage systems (for example, Ceph or Lustre), databases that require low-latency replication, and virtualized environments that benefit from fast memory-to-memory transfers.

Limitations and considerations include cost and the need for compatible software. Network design must accommodate lossless Ethernet or equivalent QoS for RoCE, and security considerations center on access control and proper isolation of memory regions to prevent unauthorized remote access.