2019 OCP Global Summit

We present a method for performing accurate and efficient telemetry based on dataplane events. Instead of capturing large volumes of statistical data (sflow or INT), we target specific network events such as packet drops, physical layer drop, buffer congestion and latency thresholds. An event record is generated with the truncated packet along with a cause or reason. An optional local aggregation mechanism is proposed that will reduce the network load by filtering any redundant data. For example, in a buffer tail drop, only a single representative packet per flow need be sent with relevant counters. A mechanism for integration in SONiC will be presented.