InfiniBand Fabric Engineering: Architecture, RDMA, and Performance at Data-Center Scale – Kobo eBook

InfiniBand Fabric Engineering is a practical guide to designing, programming, and operating InfiniBand networks for HPC and AI clusters.

The book starts with the architecture: queue pairs, transport types, addressing (LIDs, GIDs, P_Keys), packet formats, credit-based flow control, and the management plane. It then covers the physical and link layers, including link generations from SDR to XDR, signaling, FEC, cabling and optics, and physical-layer health monitoring. The RDMA chapters explain the verbs programming model, reliability and ordering semantics, memory registration, and the mlx5 direct-verbs interface and hardware offloads.

On the fabric side, you will find topology choices, routing, adaptive routing, and congestion control, followed by subnet management with OpenSM, partitions, QoS, high availability, and automation. Further chapters cover integration with MPI, storage, GPUs, virtualization, and containers, then benchmarking, telemetry, troubleshooting, resilience, and the roadmap for hybrid InfiniBand and Ethernet environments.

Code listings and command examples illustrate the concepts, and sample outputs are labeled where they are illustrative. The book assumes working knowledge of Linux and networking. It is written for fabric architects, HPC and AI infrastructure engineers, and systems developers who build on RDMA.