Publications & Reports - Document Abstract
Ond?ej Tománek, Lukáš Kencl
Multidimensional Cloud Latency Monitoring and Applications
CTU PhD Thesis
CTU Prague
This thesis discusses a novel concept called Multidimensional Cloud network latency monitoring and applications of collected measurements. Network latency has a direct impact on overall service performance, and, consequently, on the end-user experience. Unfortunately, latency is generally difficult to model, predict or control. Despite that, we show that a lot of information about Cloud service latency can be derived via thoughtfully designed active probing and, in turn, provide a foundation for improving various aspects of Cloud service performance. We initially propose a novel distributed monitoring methodology, based on measuring Cloud service at multiple dimensions it is composed of. This approach is cost effective, has the global reach, does not perturb the service and provides dependable measurements. We then offer a lightweight planet-scale implementation that serves as our data collection platform and confirms the feasibility of our approach. Later on, we leverage measurements, collected across two major public Cloud service providers, to improve Cloud services via three data-driven applications. Firstly, we use a set of descriptive-statistic methods and timeseries analyses to evaluate Cloud service, revealing both the known and unknown anomalous behaviors, as well as various insights and evolutionary trends. Secondly, we utilize preprocessed measurements for performance modeling of Cloud services, which, through a subsequent benchmarking process, are shown to exhibit certain unexpected and undesired spatio-temporal performance variations at different infrastructure levels and scales. Lastly, we show how these performance variations can be exploited by an informed gateway middleware to minimize adverse network effects and avoid many issues on the Cloud end. Applicability of the multidimensional monitoring is confirmed by, but not limited to the presented application areas. Nascent service models, computing paradigms and socio-economic trends present tremendous monitoring opportunities in the future.