One-to-One Matching of RTT and Path Changes

TL;DR

This paper evaluates change detection methods for RTT measurements, introduces a new dataset and scoring approach, and analyzes the relationship between RTT and path changes in inter-domain traffic engineering.

Contribution

It provides a comprehensive evaluation framework for RTT change detection, including a labeled dataset, tailored scoring, and a data transformation, addressing previous shortcomings in the field.

Findings

Identified frequent AS path changes with minimal RTT impact

Proposed a new data transformation improving detection performance

Analyzed correlation between RTT and path changes

Abstract

Route selection based on performance measurements is an essential task in inter-domain Traffic Engineering. It can benefit from the detection of significant changes in RTT measurements and the understanding on potential causes of change. Among the extensive works on change detection methods and their applications in various domains, few focus on RTT measurements. It is thus unclear which approach works the best on such data. In this paper, we present an evaluation framework for change detection on RTT times series, consisting of: 1) a carefully labelled 34,008-hour RTT dataset as ground truth; 2) a scoring method specifically tailored for RTT measurements. Furthermore, we proposed a data transformation that improves the detection performance of existing methods. Path changes are as well attended to. We fix shortcomings of previous works by distinguishing path changes due to routing protocols (IGP and BGP) from those caused by load balancing. Finally, we apply our change detection methods to a large set of measurements from RIPE Atlas. The characteristics of both RTT and path changes are analyzed; the correlation between the two are also illustrated. We identify extremely frequent AS path changes yet with few consequences on RTT, which has not been reported before.