Zero copy
RDMA allows NICs to directly read from or write to memory, enabling data to be transferred between memory spaces without the need for copying between operating system buffers. This enables direct communication between distributed block storage services across nodes, significantly reducing latency.
Kernel bypass
RDMA avoids data copying between application memory and kernel in the traditional transmission process. Without any kernel involvement, data can be sent from application memory to local NICs and transmitted over the network to remote NICs. The kernel bypass capability of RDMA makes it highly effective in large-scale clusters and distributed systems, enabling the system to handle data transfer across numerous nodes more efficiently.
CPU offload
RDMA technology enables data transfer and processing directly on the network adapter, significantly reducing CPU usage even under high bandwidth demands. This allows the CPU to focus on computation rather than data transfer management, enhancing overall system throughput and efficiency. This mechanism is suitable for high-concurrency and large-scale data processing environments, which effectively reduces system bottlenecks.
