RDMA: The Reason Remote Storage Feels Like It's Plugged Into Your Motherboard
Remote SSDs don't have to be slow. Learn how RDMA, RoCE, and NVMe-oF cut network overhead and CPU cost, and where they help AI training and inference.
Remote Direct Memory Access (RDMA) revolutionized remote storage by eliminating network latency. Traditional remote storage required data to traverse the kernel network stack, which added 150 microseconds to what was already 80 microseconds for SSD reads. RDMA bypasses the network stack entirely, allowing direct memory access between computers.
This is achieved through three key techniques: zero-copy, kernel bypass, and one-sided operations. Zero-copy moves data directly between application buffers without kernel involvement, kernel bypass uses the NIC to handle transport in hardware, and one-sided operations allow remote memory access without the remote CPU participating.
RDMA comes in various forms, such as InfiniBand, RoCE (RDMA over Converged Ethernet), and iWARP (RDMA over TCP/IP). The most common use case is NVMe over Fabrics (NVMe-oF), which enables remote SSD access with minimal latency. Compared to other remote storage protocols, RDMA significantly reduces fabric-added latency and CPU utilization at high IOPS levels.
Major tech companies and storage vendors have adopted RDMA in their products, demonstrating its widespread adoption and impact on enterprise storage architectures.
Written by urgent.news from HackerNoon's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
