The normal way computers talk
Memory (RAM) is the short-term workspace of a computer. When two computers share data over a network, the data usually takes a long trip. It goes from the program to the operating system, through the network card, across the wire, and back up through the other computer's operating system.
Each stop uses the CPU (the computer's brain) and adds delay. Delay is also called latency. It is the waiting time before data arrives.
What RDMA changes
RDMA skips many of those stops. The network card of one computer places data straight into the memory of the other computer. The CPUs and operating systems stay out of the middle.
Think of two offices. The normal way is passing a note through several desks. RDMA is a pneumatic tube that drops the note right onto the other person's desk.
Why it is useful
- Low latency: data arrives in a very short time.
- High speed: it can fill very fast network links.
- Low CPU use: the CPU is free to do real work.
Where it is used
- Large data centers and storage systems.
- High-performance computing, such as science and AI work.
- Fast database clusters.
Common types
- InfiniBand: a special network type built for RDMA.
- RoCE: RDMA over Converged Ethernet. It runs RDMA on special Ethernet gear.
- iWARP: RDMA on top of the normal TCP network protocol.
What you need
Both computers need network cards that support RDMA. The switches between them must also be set up for it. Normal home or small business hosting does not need RDMA. Most websites work well without it.
Quick recap
- RDMA moves data directly between the memory of two computers.
- It cuts delay and saves CPU power.
- It needs special network cards and setup.
- It is used mainly in big data centers, not small websites.