Knowledge about future traffic in backbone optical networks may greatly improve a range of tasks that Communications Service Providers (CSPs) have to face. This work proposes a procedure for long-term traffic forecasting in optical networks. We formulate a long-terT traffic forecasting problem as an ordinal classification task. Due to the optical networks’ (and other network technologies’) characteristics, traffic forecasting has been realized by predicting future traffic levels rather than the exact traffic volume. We examine different machine learning (ML) algorithms and compare them with time series algorithms methods. To evaluate the developed ML models, we use a quality metric, which considers the network resource usage. Datasets used during research are based on real traffic patterns presented by Internet Exchange Point in Seattle. Our study shows that ML algorithms employed for long-term traffic forecasting problem obtain high values of quality metrics. Additionally, the final choice of the ML algorithm for the forecasting task should depend on CSPs expectations.

The problem that this paper investigates, namely, optimization of overlay computing systems, follows naturally from growing need for effective processing and consequently, fast development of various distributed systems. We consider an overlay-based computing system, i.e., a virtual computing system is deployed on the top of an existing physical network (e.g., Internet) providing connectivity between computing nodes. The main motivation behind the overlay concept is simple provision of network functionalities (e.g., diversity, flexibility, manageability) in a relatively cost-effective way as well as regardless of physical and logical structure of underlying networks. The workflow of tasks processed in the computing system assumes that there are many sources of input data and many destinations of output data, i.e., many-to-many transmissions are used in the system. The addressed optimization problem is formulatedin the form of an ILP (Integer Linear Programing) model. Since the model is computationally demanding and NP-complete, besides the branch-and-bound algorithm included in the CPLEX solver, we propose additional cut inequalities. Moreover, we present and test two effective heuristic algorithms: tabu search and greedy. Both methods yield satisfactory results close to optimal.