P4 Telemetry collector

Abstract

As the complexity of computer networks increases to accommodate the demand for massive connectivity and cloud services, so does the probability of fault occurrence and the surface of attacks. Hence the need for constant monitoring of network devices and accurate analysis of traffic patterns to ensure the highest performance and maximum security. This requires collecting and processing telemetry data from many sources in the network which leads to extra bandwidth usage and strains the CPU at the monitoring system resulting in scalability issues as the network grows. In this paper, we propose a two-stage postcard telemetry collector based on data plane programmability using the P4 language to address the scalability issues. We show a decrease in the CPU load of the telemetry server by over 70% while lowering the bandwidth to less than 7% in the most extreme scenario, at the cost of variable delay introduced in the collection of the postcards.