Efficient and scalable execution of smart city parallel applications

Recent research efforts in the field of urban computing aim to develop innovative services for citizens through the application of ubiquitous and pervasive computing paradigms in urban spaces. Smart city applications need to cope with a large number of involved users and devices. Since data and objects are strictly related to the territory on which they are defined and used, it is preferable, when possible, to perform computation locally through the adoption of dispersed computing nodes such as CPU-equipped sensors. In this context, the computation related to smart city applications can be profitably and efficiently parallelized by partitioning the territory into regions and assigning the computation related to each single region to a local node. Nevertheless, the adoption of parallel computing models poses several communication and synchronization issues, especially when the number of nodes is large and the time constraints of applications are compelling. This paper presents and analyzes a parallel computing model for smart city applications in which each node needs to exchange information only with a subset of neighbor nodes, allowing the synchronization overhead to be significantly reduced. As sample application, we consider the analysis and prediction of internet traffic generated by vehicle and pedestrian devices moving on a smart avenue equipped with distributed computing nodes. This work offers a detailed performance evaluation in a number of scenarios, including uniform and nonuniform user distribution and different types of user mobility behavior. The results show that the presented computation model offers notable advantages in terms of computation efficiency and speedup, with respect to a classical all-to-all synchronization paradigm, in which the nodes need to coordinate with a central entity.