Balancing speedup and accuracy in smart city parallel applications

Smart city and Internet of Things applications can benefit from the use of distributed computing architectures, due to the large number and pronounced territorial dispersion of the involved users and devices. In this context, a natural method to parallelize the computation is to consider the territory as partitioned into regions, e.g., city neighborhoods, and associate a computing entity with each region. The application considered in this paper is the prediction of the amount of internet traffic generated within a given region, which requires to consider not only the devices located in the region but also the mobile devices that are expected to enter the local region in the future. When setting the number of neighbor regions included in the computation, it must be considered that this parameter has opposite effects on two important objectives: increasing the number of neighbors tends to improve the accuracy of the prediction but slows down the computation because more computing entities need to synchronize among each other. Similar considerations apply when setting the size and number of regions that partition the territory. This paper offers an insight onto these important tradeoff issues.