Remote Direct Memory Access Transport (RDMA)
On computers that support remote direct memory access (RDMA), network adapters can move data directly from memory in one computer to memory in another computer across a network. This strategy can reduce or eliminate several sources of latency resulting from overhead, including network protocol overhead, operating system overhead, and context switching overhead.
An RDMA transport can run peer-to-peer, or be server-based for client transport in a persistence cluster.
On each host computer, the transport allocates buffers for each connection on an RDMA transport.
RDMA is a connection-oriented protocol. Its transport definitions are inherently fragmentary. You must define at least two complementary transport definitions to establish a bus. For more information, see Pair Connections.
Although an individual RDMA connection links exactly two host computers, you can combine several connections into a bus with a more complex topology (see Assembling Larger Topologies from Pair Connections).
- RDMA Transport: Parameters Reference
The following tables describe the parameters specific to RDMA transports in the configuration interface. - RDMA Hardware Reference
RDMA transports require hardware that is compliant with implementation standards in the following table. - RDMA Implementation Library Reference
In addition to RDMA hardware, RDMA transports require the verbs and RDMACM libraries from OFED. - Configuring Locked Buffer Memory for RDMA
The base library reserves buffers for RDMA transports. These buffers are locked, that is, the operating system cannot swap them out to virtual memory. On Linux platforms, you must configure sufficient memory resources for these buffers.