Monitoring & Predicting QoS in IoT Services

Abstract

Internet of Things (IoT) applications can be built from a number of heterogeneous services provided by a range of devices, which are potentially resource constrained and/or mobile. These device characteristics can lead to the services deployed on them becoming unreliable as the device may lose power or move out of range. As these services and applications continue to be more widespread, a key research challenge is how to make them more reliable. The reliability of an application is influenced by the time to detection (TTD) of a failure and the time to recovery (TTR) of services in the application after the failure. TTD and TTR are affected by the accuracy of the prediction and by the time it takes to receive the prediction. This thesis focuses on reducing TTD by improving the prediction accuracy and reducing TTR by reducing the time it takes to receive the prediction.

Accurate short-term forecasts allow dynamic systems to adapt their behaviour when degradation is detected e.g., transportation forecasting supports alternative routing of traffic before gridlock and wind power forecasting enables the use of dispatchable energy such as hydroelectric power to reduce the difference between power consumption and power generation in the network. This rationale can be applied to service-oriented computing when creating and managing service applications, where such applications are composed of available collaborating services. The faster a problem with a service can be detected, the faster a suitable replacement service can be chosen. Previous approaches that have focused on QoS forecasting have used traditional time series methods, but these are not suitable as QoS does not exhibit traditional time series patterns (i.e., sudden peaks caused by network congestion or a device switching to a power saving mode). More modern recurrent neural network-based approaches such as GRUs and LSTMs have been proposed but the long training time means they take longer to incorporate recent QoS values. This can lead to a reduction in forecasting accuracy in dynamic IoT environments. This thesis proposes a noisy-echo state network approach that has been designed to be deployed at the edge of the network. The reduced training time allows the model to incorporate recent QoS values on devices at the edge. The results show increased forecasting accuracy compared to current state of the art approaches when tested on a combined dataset of IoT and web services, reducing TTD.

Once a problem has been detected with one of the services in a composition, the application needs to be recovered by using a functionally equivalent service with high QoS. Given that candidate services may not be currently executing, predictions based on a time series of current QoS values are not appropriate. Recommending possible replacement services requires a technique that avails of similar users' recent experience of those services, which is the most up-to-date information available about the services'