Measuring TCP Round-Trip Time in the Data Plane

Abstract

We present a data-plane algorithm that passively and continuously monitors the Round-Trip Time of TCP traffic, by matching data packets with their associated acknowledgments and calculating a time difference. Compared with traditional measurement systems based on active probing or measuring only SYN/ACK packets, our algorithm passively produces many samples for long-running connections. This enables network operators to observe abnormal RTT increases, which signal possible security or performance issues in the network, in real-time. To satisfy the stringent memory size and access constraints of programmable switches, our algorithm uses a multi-stage hash table data structure to maintain records for in-flight packets; the records not receiving their acknowledgments are lazily expired and overwritten. We implement our algorithm on a Barefoot Tofino programmable switch. Evaluation using a real-world traffic trace from a 10 Gbps campus network link demonstrates that our solution can accurately capture 99% of available RTT samples, using only 4 MB of data-plane memory.