| Tartalmi leírás: |
Server farm telemetry plays a crucial role in overseeing and ensuring the health, performance, and efficiency of interconnected servers, which deliver computing resources for applications and services. An essential step in the telemetry process is the analysis of data collected from servers. Anomaly detection is a significant task in this step: by automatically detecting signs of abnormal behaviour, operators can prevent issues from escalating into major operational setbacks. This paper presents a new pre-processing procedure to improve our previous anomaly detection algorithms and state-of-the-art detectors. Motivated by the finding that periodic datasets often pose challenges to anomaly detection, our method transforms a dataset by removing modes containing regular, periodic behaviour while preserving signs of anomalies. Our proposed pre-processing procedure improved the performance of all anomaly detectors we tested, while our latest detector achieved three times better performance on periodic data than originally. We also present an extensive numerical analysis of our pre-processing parameters and state-of-the-art anomaly detection algorithms regarding performance via the F-score metric.
|
