RDMA and MPI for High-Performance Computing

Make MPI fast on RDMA fabrics: libfabric, collectives, RMA, memory registration and GPU-aware communication.

Advanced30 lessons written lessons≈ 15 h

High-Performance Networking & Cloud-Native · 30 illustrated lessons (text/HTML)

Outcomes

Every lesson in this curriculum, in the order we teach it.

1Course introduction
2HPC fundamentals
3RDMA technical principles
4InfiniBand architecture explained
5RoCE and iWARP explained
6Overview of the MPI standard
7MPI implementations and the layers beneath them
8Architecture explained
9Open MPI OOB and BTL
10MPICH CH4 and netmod explained
11The OFI (libfabric) interface: architecture and core concepts
12The UCX communication framework
13RDMA communication primitives
14RDMA optimisation of MPI point-to-point communication
15RDMA optimisation of MPI collective communication
16MPI RMA and RDMA integration explained
17Memory registration and pinning: RDMA registration principles, pinning overhead and registration caching
18Queue pair (QP) management
19Completion queues (CQ) and event handling
20Overlapping communication and computation
21Performance tuning basics
22Tuning MPI runtime parameters
23Network hardware tuning: key settings for HPC
24Optimising large-scale job startup
25Fault tolerance and resilience: RDMA link-failure detection, MPI error handling, checkpoint/restart with RDMA
26Heterogeneous computing and RDMA: GPUDirect RDMA and MPI integration
27Advanced communication patterns: RDMA support for persistent, partitioned and neighbourhood-collective MPI
28Case study 1: cluster deployment and configuration with Open MPI and UCX
29Case study 2: performance analysis and tuning of an application with MPICH and OFI
30Outlook: smart NICs and the future of communication libraries

Every curriculum on this page, with the labs and lab sequences, for a year.