Predicting the online performance of video service providers on the internet

Video services on the Internet are not able to offer consistent and assured performance to users or third-party applications. Measuring levels of performance over time is difficult, and obtaining accurate measures in real time is problematic; thus, reactive measures to address loss of performance are also problematic. The ability to predict service performance can be viewed as an important added-value, one that can help users or third-part applications select the proper online service provider. With this aim in view, we have designed a measurement system and deployed it in eleven provinces and cities in China to monitor two popular websites, Youku and Tudou. The analysis indicates that the performance trend of these two service providers followed daily changing patterns, such as rush hour traffic and lower service workloads at midnight; this is consistent with user behaviors. It was also confirmed that the future performance was related to the historical records. Based on these findings, we have decided to investigate the use of modified time series models to forecast the performance of such video services. Meanwhile, some machine learning models are implemented and compared as baseline models, such as Artificial Neural Network, Support Vector Machine, and Decision Tree. In addition, a hybrid model, which is generated by combining different machine learning models, is also studied as the baseline. An investigation shows that time series models are much more suitable to this prediction problem than baseline models in most situations. To alleviate the data sparseness problem in training the predictor, a new predictor that combines different information sources is proposed, thus improving prediction precision. Furthermore, the predictor is quite stable, and we have discovered that the average performance estimation is more accurate if the model is updated within 2–3 days, which is useful in some applications, e.g., video source analysis and recommendation systems.