DUNE: Improving Accuracy for Sketch-INT Network Measurement Systems

In-Band Network Telemetry (INT) and sketch algorithms are two representative methodologies for measuring network traffics in real time. To combine sketch with INT and to keep their advantages, the "reconstructing sketch at endhost" approach, which uses INT to send small pieces of switch sketch (i.e., sketchlet) to end-host for reconstructing an identical sketch, is a promising direction. However, we reveal that the naive column sketchlet is not efficient, and inaccuracies arise because of the invalid and stale measurement data in the endhost reconstructed sketch. In this paper, we present an innovative sketch-INT measurement system named DUNE. DUNE is lightweight by following the "reconstructing the sketch at endhost" approach, and to improve the measurement accuracy, we make two innovations: First, we design a novel sketchlet named scatter sketchlet that is more efficient in transferring measurement data by allowing a switch to select individual sketch buckets to add to sketchlet; Second, we develop data structures for tracing "freshness" of the sketch buckets, and present algorithms for smartly selecting buckets that contain valuable measurement data to send to end-host. We theoretically prove that our proposed methods are superior comparing with the existing solution, and implement a prototype on the commodity Barefoot Tofino switches. We conduct extensive experiments on DUNE, and the evaluation results show that the system considerably improves measurement accuracies at negligible costs. In particular, with less than 0.36% loss of the packet forwarding rate, DUNE avoids up to 60% errors in the end-host reconstructed sketch.