PNT Sketch: A Generic Sketch Algorithm for Periodic Network Telemetry

Network telemetry provides operators with a fine-grained internal view of network behavior. Recently, various lightweight and high-accuracy sketch algorithms have been proposed for real-time full flow record collection. However, they still have at least two deficiencies: (1) they are mostly designed for counting and cannot accurately support other kinds of aggregation functions, such as max and mean; and (2) the vast majority of them are focused on improving the measurement accuracy of each individual period, overlooking the fact that many elephant flows last for multiple periods. In this paper, we propose PNT sketch, a generic algorithm with well-designed data structures and strategies. The PNT sketch implements multiple aggregation functions to record flow attributes for elephant flows and improves the accuracy of a new period through the full use of historical flow records collected in the previous period with our history-aware reset strategy. We conduct a theoretical analysis of our algorithm, and the results show that it has a more rigorous error bound than the Elastic sketch. We implement our algorithm on both the x86 and P4 platforms and evaluate its performance on seven different tasks. Compared to the state-of-the-art, the PNT sketch achieves 1.4 similar to 437.2 smaller error rates.