Traditional TCP/IP networks offer adequate bandwidth for many workloads, although it comes with some limitations, including CPU overhead and higher than wanted round-trip latency. The RDMA protocol reduces CPU overhead and latency by skipping through layers of the stack, resulting in rapid data transfers.
Two network implementations of RDMA, the Internet Wide Area RDMA protocol (iWARP) and RDMA over Converged Ethernet (RoCE), are embedded in the network adapter hardware, offloading nearly all the work from the CPU to offer faster networking:
- iWARP implements RDMA over IP networks using TCP, making it ideal for organizations that want to use RDMA over their existing IP network infrastructure without any specialized hardware. iWARP requires no additional configuration at the ToR switches for its implementation.
- RoCEv2 uses UDP and requires QoS (Quality of Service) to ensure packet delivery. It relies on an Ethernet network configured to use Layer 2 Priority Flow Control (PFC) or Layer 3 DSCP PFC to minimize congestive packet loss.
Figure 39. iWARP vs RoCE
See Reference Guide: Switch Configurations – RoCE Only (Mellanox Cards) for details about the switch configurations for all RoCE-based deployments for Microsoft Azure Stack HCI Solutions from Dell Technologies.
See Reference Guide: Switch Configurations - iWARP Only (Qlogic Cards) for details about the switch configurations for all iWARP-based deployments for Microsoft Azure Stack HCI Solutions from Dell Technologies.