RDMA over Converged Ethernet (RoCE)

RDMA over Converged Ethernet is a network protocol that allows remote direct memory access (RDMA) over an Ethernet network. One of the benefits running RDMA over Ethernet is the visibility provided by standard sFlow instrumentation embedded in the commodity Ethernet switches used to build data center leaf and spine networks where RDMA is most prevalent.

The sFlow telemetry stream includes packet headers, sampled at line rate by the switch hardware. Hardware packet sampling allows the switch to monitor traffic at line rate on all ports, keeping up with the high speed data transfers associated with RoCE.

The diagram above shows the packet headers associated with RoCEv1 and RoCEv2 packets. Decoding the InfiniBand Global Routing Header (IB GRH) and InfiniBand Base Transport Header (IB BTH) allows an sFlow analyzer to report in detail on RoCE traffic.

The sFlow-RT real-time analytics engine recently added support for RoCE by decoding InfiniBand Global Routing and InfiniBand Base Transport fields. The screen capture of the sFlow-RT Flow-Trend application shows traffic associated with an RoCEv2 connection between two hosts, 10.10.2.22 and 10.10.2.52. The traffic consists of SEND and ACK messages exchanged as part of a reliable connection (RC).

The standard sFlow instrumentation provides comprehensive network wide visibility into RoCE and all other applications sharing the network resources. Real-time visibility is an essential part of automating networks, providing the feedback needed to ensure that resources are efficiently allocated and rapidly identifying overloaded resources so that remediation action can be taken before significant service degradation occurs.