RDMA (Remote Direct Memory Access)

Remote Direct Memory Access (RDMA) is a technology that lets one computer read from or write to another computer's memory directly over the network, without involving the operating system on either machine. The network interface cards (NICs) set aside regions of memory and move data between them on their own, so a transfer skips the usual layers of OS networking code. The result is very low latency and high throughput, which is why RDMA is one of the most important technologies underlying high-performance computing (HPC).

What is RDMA?

In conventional networking, data passes through the operating system's networking stack on both the sending and receiving machines, which copies buffers and consumes CPU cycles at every step. RDMA removes the OS from the data path. A portion of RAM is registered with the NIC of each machine, and certain RDMA operations then move data between those regions directly. Because the CPU and kernel are largely bypassed, the transfer completes with minimal overhead and leaves processor cycles free for computation.

RDMA is provided by high-speed interconnects such as InfiniBand, as well as by RoCE (RDMA over Converged Ethernet) and iWARP on Ethernet networks.

Why RDMA matters

Speed alone is not enough to reach HPC-level performance. Even with a fast interconnect, the operating system on each compute node will bog down while processing a flood of network requests, and the link's full bandwidth is never realized. RDMA is often the component that actually enables HPC speeds, because it lets nodes exchange data without running large amounts of OS-level networking code.

This matters most for tightly coupled work. A large MPI-based parallel computing job passes data between its nodes constantly, and doing so over RDMA rather than through each node's OS dramatically speeds up the application. Parallel file systems and storage networks use RDMA for the same reason: to keep data flowing to processors without a software bottleneck.

How RDMA works

• Memory registration pins regions of application memory and makes them known to the NIC.
• Kernel bypass lets the application post transfer requests directly to the NIC rather than through system calls.
• Zero-copy transfer moves data straight from one machine's registered memory to the other's, avoiding intermediate buffer copies.
• Hardware offload has the NICs, not the CPUs, carry out the data movement.

Together these steps cut both latency and CPU usage, so more of each node's processing power goes toward the actual computation instead of moving data. That efficiency is what allows parallel applications to scale across large numbers of nodes while keeping communication overhead low.