NVMe over Fabrics extends the performance advantages of NVMe storage beyond individual servers to shared storage accessed over a network. By eliminating the protocol translation overhead of traditional SAN protocols like iSCSI and Fibre Channel, NVMe-oF delivers near-local storage performance across the data center fabric.
NVMe-oF Transport Options
NVMe over RDMA using RoCE v2 provides the lowest latency path, leveraging remote direct memory access to bypass the CPU for data transfers. This transport delivers sub-100-microsecond latency and multi-gigabyte-per-second throughput, making it suitable for latency-sensitive database and AI workloads that demand the highest storage performance.
NVMe over TCP removes the requirement for specialized RDMA-capable networking hardware, working over standard Ethernet infrastructure. While latency is modestly higher than RDMA transports, NVMe/TCP is significantly faster than iSCSI and dramatically simplifies deployment by leveraging existing network infrastructure without requiring fabric changes.
For data centers deploying NVMe-oF, the storage architecture shifts from monolithic storage arrays to disaggregated storage pools. Compute and storage scale independently, flash resources are shared efficiently across multiple servers, and storage capacity can be dynamically allocated to workloads based on their performance requirements.