RDMA Architecture & Mechanism
RDMA operates through specialized network interface cards (RNICs) that handle transport protocol processing and direct memory operations in hardware. Key features include:- Zero-Copy Data Transfers: Applications can read/write remote memory without CPU intervention.
- Bypassing Kernel Overhead: Eliminates traditional network stack processing, reducing latency to sub-microsecond levels.
- Explicit Memory Registration: Memory regions must be registered with the RNIC, which maintains virtual-to-physical address translation for secure access.
- Reliable Connected (RC): Guarantees in-order, reliable delivery for critical workloads.
- Unreliable Datagram (UD): Optimized for low-latency applications where minor packet loss is acceptable.

InfiniBand: High-Performance RDMA Networking
InfiniBand is a widely adopted RDMA-based networking architecture designed for high-performance computing (HPC) and data-intensive applications. It provides:- Ultra-Low Latency: Achieves as low as 600 nanoseconds end-to-end.
- High Bandwidth: Supports data rates from 10 Gb/s (SDR) to 100 Gb/s per port over copper or optical fiber.
- Host Channel Adapters (HCAs): Offload protocol processing and memory translation to hardware, reducing CPU load.

RDMA & InfiniBand in Modern Computing
These technologies are critical for:- Distributed Machine Learning: Efficient communication for large-scale AI training workloads.
- HPC & Scientific Computing: Accelerated data sharing across compute clusters.
- Cloud Data Centers: Optimized storage access (NVMe-over-Fabrics) and disaggregated computing.