RDMA vs TCP/IP: Protocol Comparison and Overview of RDMA Types for Distributed Storage Networks
This article explains the differences between RDMA (including RoCE, InfiniBand, and iWARP) and traditional TCP/IP, describes the three main RDMA network types, compares their characteristics, and outlines their typical usage in distributed storage environments.
RDMA and TCP/IP
High‑performance computing and big‑data analytics demand I/O‑intensive, low‑latency communication that traditional TCP/IP stacks cannot satisfy because they rely on kernel‑mediated message passing, incurring significant data movement and copy overhead.
Remote Direct Memory Access (RDMA) allows network interfaces to read/write remote memory directly without OS kernel involvement, enabling high throughput and ultra‑low latency, especially suitable for large parallel clusters.
Figure 1‑1 – Comparison of RDMA and traditional TCP/IP
Types of RDMA
There are three main RDMA networks: InfiniBand, RoCE (RDMA over Converged Ethernet), and iWARP.
InfiniBand is a purpose‑built RDMA network offering high bandwidth and low latency but at a high cost. RoCE and iWARP are Ethernet‑based RDMA solutions that bring RDMA’s speed and low CPU usage to the ubiquitous Ethernet infrastructure.
RoCE has two versions: RoCEv1 operates at the Ethernet link layer and requires switches that support flow‑control (PFC); RoCEv2 runs over UDP/IP, adding IP‑level scalability.
Figure 1‑2 – RDMA network types
Table 1‑1 – Comparison of RoCE and InfiniBand
InfiniBand: Designed for RDMA from the hardware level, provides higher bandwidth and lower latency, but requires expensive IB NICs and switches.
RoCE: Ethernet‑based RDMA, consumes fewer resources than iWARP and supports more features; works with standard Ethernet switches if the NIC supports RoCE.
iWARP: TCP‑based RDMA that leverages TCP’s reliability; in large deployments it can create many TCP connections, increasing memory usage and requiring higher system specifications.
Common Network Protocols in Distributed Storage
IB – typically used for front‑end storage networks in DPC scenarios.
RoCE – commonly deployed in back‑end storage networks.
TCP/IP – used for general business networks.
For further reading on RoCE technology, see the linked articles:
Three Implementations and Applications of RoCE
High‑Performance Computing: RoCE Analysis and Applications
Choosing Between RoCE v2 and InfiniBand for HPC
Lossless Data‑Center Networking (IP, RDMA, IB, RoCE, AI Fabric)
NVMe over RoCEv2: Requirements and Test Specifications
Disclaimer: Thanks to the original author for the hard work. All reproduced articles are properly credited; please contact us for any copyright concerns.
Architects' Tech Alliance
Sharing project experiences, insights into cutting-edge architectures, focusing on cloud computing, microservices, big data, hyper-convergence, storage, data protection, artificial intelligence, industry practices and solutions.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.