SmartEdge: Towards Configuring Complex Applications in Mobile Edge Computing

The integration of Internet of Things (IoT) devices and services in the access network enables applications that could not be executed on the resource-poor IoT devices. At the same time, the services that are used by these applications can be provided at very low latency. In mobile edge computing, the computing resources enabling such applications are provided by nearby devices, servers located at the edge, or the cloud. Thus, so-called "complex applications" can be configured by binding a set of required components to distributed services provided by devices in the edge environment. One of the major challenges of mobile edge computing is its inherent dynamism. Mobility of users and devices lead to ever changing environments. For complex applications, this means that a reconfiguration is necessary whenever a service provided by a mobile resource becomes unavailable. In this work, we propose SmartEdge, an approach that leverages the benefits of both classical constraint-based optimization (CCO) and Graph Neural Networks (GNNs) to find a configuration. To overcome the cold start problem of learning based approach, we use CCO to find initial configurations. The output from the CCO is then used to train a reinforcement learning algorithm consisting of two GNNs (representing the services offered by the edge environment and the application model, respectively). The trained model is then used to reconfigure the application if changes in the dynamic mobile edge computing environment negatively affect the provided quality of service.