RDMAcapable

RDMAcapable describes hardware or software that can perform Remote Direct Memory Access (RDMA) operations. It denotes the ability to participate in RDMA communications, where memory-to-memory data transfers occur directly between systems with minimal CPU involvement.

RDMA enables low-latency, high-throughput data movement by allowing devices to read and write remote memory without CPU intervention. The main RDMA transport technologies are InfiniBand (IB), RDMA over Converged Ethernet (RoCE), and RDMA over Internet Wide Area Protocol (iWARP). RDMA-capable devices implement an RDMA engine and expose a verbs-based API for applications to post work requests, register memory, and receive completion notifications. On Linux and other systems, user-space libraries and the kernel stack in the rdma-core (OFED) provide the interface between applications and the RDMA hardware.

A system is RDMAcapable when the host network interface card (NIC) supports RDMA, the underlying network fabric supports the chosen transport (IB, RoCE, or iWARP), and the software stack is installed and configured. Common components include RDMA-capable NICs (for example, Mellanox/NVIDIA ConnectX, Intel Ethernet adapters, Chelsio), InfiniBand or RoCE switches, and the corresponding kernel and user-space RDMA stacks.

Benefits of RDMAcapable systems include reduced CPU overhead for data movement, lower latency, and higher throughput, making them suitable for high-performance computing, data centers, storage systems, and latency-sensitive databases. Limitations involve the need for compatible hardware, properly configured network fabrics (often lossless Ethernet for RoCE), and a supporting software stack.

Notes: RDMAcapable is a descriptive term reflecting hardware, firmware, drivers, and software support; verification typically involves checking hardware specifications and conducting RDMA performance tests.