DARTBOARD: Designing Multi-Pipeline Algorithms for Round-Trip Time Monitoring

Abstract

Network telemetry research is critical to improving security and performance of the world's largest networks. Today's high-speed network switches use multiple input and output pipelines in parallel to process packets, a stark contrast from the research which assumes a one-input, one-output case. Expanding off of the round-trip time algorithms presented in DART, we design and evaluate strategies for multi-pipeline telemetry. We clearly outline what parts of DARTBOARD are new work and what existing work is taken from DART, highlighting the new modules of the Multi-Pipe Switch, a Mapping Function, and Result Aggregation. We demonstrate that our designs are resilient to additional constraints on per-pipe memory or flow-based sampling. Using flow-based sampling, we show that we only need to use a fraction of the single-pipeline memory (1 in 10 flows) to obtain average RTT times within a relative error threshold of 20%. Optimizing for minimal recirculations, our designs require an effective recirculation rate of less than 5%, when we recirculate only the necessary packet header fields. Finally, by limiting recirculations to acknowledgements only, we reduce total recirculations by more than 50%. Our research improves multi-pipeline's ability to perform accurately and efficiently in memory, bridging the gap between research advancements in network telemetry and the architecture constraints of the modern network.