2019 OCP Global Summit

Modern data-centers use high-density L2 and L3 topologies to accommodate heterogeneous mission critical applications. In such an environment visibility is essential to reliably operate and manage the networking elements. We present an open framework called “Inband Network Telemetry” (INT). In INT each network element (e.g. a SmartNIC or a Barefoot Network switch) can be programmed to insert metadata about the network state in the packet as it traverses the network. This enables visibility on all traffic (e.g. Compute, Storage/RDMA) carried over ethernet for transient issues due to bottlenecks, failures or configuration errors. In such a solution machine learning techniques can be used to Baseline Network Performance (Path and Latency), match patterns of behavior and correlate phenomena. Such techniques can be used to promptly react to network issues or predict and anticipate problems.