Distributed load balancing for heterogeneous fog computing infrastructures in smart cities

Smart cities represent an archetypal example of infrastructures where the fog computing paradigm can express its potential: we have a large set of sensors deployed over a large geographic area where data should be pre-processed (e.g., to extract relevant information or to filter and aggregate data) before sending the result to a collector that may be a cloud data center, where relevant data are further processed and stored. However, during its lifetime the infrastructure may change, e.g., due to the additional sensors or fog nodes deploy, while the load can grow, e.g., for additional services based on the collected data. Since nodes are typically deployed in multiple time stages, they may have different computation capacity due to technology improvements. In addition, an uneven distribution of the workload intensity can arise, e.g., due to hot spot for occasional public events or to rush hours and users’ behavior. In simple words, resources and load can vary over time and space. Under the resource management point of view, this scenario is clearly challenging. Due to the large scale and variable nature of the resources, classical centralized solutions should in fact be avoided, since they do not scale well and require to transfer all data from sensors to a central hub, distorting the very nature of in-situ data processing. In this paper, we address the problem of resources management by proposing two distributed load balancing algorithms, tailored to deal with heterogeneity. We evaluate the performance of such algorithms using both a simplified environment where we perform several sensitivity analysis with respect to the factors responsible for the infrastructure heterogeneity and exploiting a realistic scenario of a smart city. Furthermore, in our study we combine theoretical models and simulation. Our experiments demonstrate the effectiveness of the algorithms under a wide range of heterogeneity, overall providing a remarkable improvement compared to the case of not cooperating nodes.