Session

This tutorial will introduce developers familiar with Linux networking to the Linux RDMA stack and how it can be used to implement high-performance communication for applications. An outline of the proposed tutorial is:

• Key features of RDMA for high performance communication o Asynchronous work / completion queues o Kernel bypass . Transport offload from application CPUs, including reliability / retransmission o One-sided (and two-sided) ops . Send-receive vs RDMA read/write . Memory pre-registration / pinning o Transport . IB and RoCE . RoCE vs iWARP/TCP vs proprietary (eg AWS EFA) . RC vs UD
• Simple example application (RC pingpong) o Key objects: QP, CQ, MR o Register memory (extensions for accelerator memory, etc) o Connection establishment with IP addressing & librdmacm o Post work requests to send and receive work queues o Poll CQ o Compare with UD version