Internet Performance Prediction Framework Based on PingER Dataset

The Internet performance directly affects the scalability, reliability and availability of the online applications. Delay of a few millisecond may cause companies lose millions of dollars. Therefore, Internet measurements are carried out to capture the performance of the Internet links worldwide. Most of the Internet performance monitoring frameworks are active in nature i.e., they can only capture the real-time performance of the Internet links. Thus, these monitoring frameworks are unable to forecast the near future performance of the Internet links in a region. Such estimates are quite critical for the network administrators to carry out bandwidth extensive experiments between different sites, policy makers to suggest future upgrades to the Internet infrastructures or streaming service providers to enhance the quality of service to their customers. Therefore, we analyze different machine learning algorithms including Multiple Linear regression, Random Forest algorithm, Gradient Boosting, and eXtreme Gradient Boosting to predict the performance of the Internet links using PingER (Ping End-to-End Reporting) dataset for the countries like China, India and Japan. Our experimental results show that the Multiple Linear regression has improved Internet performance prediction accuracy compared with the other methods. Our work can be utilized by the Internet service providers, streaming service providers or policymakers for the design, deployment, and evaluation of next-generation Internet infrastructure.