Better client OFF time prediction to improve performance in web information systems

• Authors: Alan Berfield;Bill Simons;Panos K. Chrysanthis;Kirk Pruhs
• Venue: Proceedings of the 3rd international workshop on Web information and data management
• Year: 2001

Prefetching is a potential technique for reducing latency in Web information Systems. However, it has been shown that the burstiness of standard prefetching can drastically increase network congestion, and can even increase, rather than decrease, average user perceived latency. Accurate OFF time, the idle periods between user requests, prediction potentially allows the document to be downloaded at an even rate over the OFF time, which can ameliorate the burstiness, and significantly improve both network congestion and average user perceived latency. Yet accurate prediction of such OFF times has been difficult to achieve. This paper examines the use of two machine-learning techniques, namely, neural networks and genetic algorithms, for OFF time prediction. Our performance evaluation results show that these techniques provide better accuracy than those previously reported, with an average increase of twice the correlation. Our results also show that document type is the best predictor of OFF time. Further, our functions can be tailored to favor underpredictions, which would have less negative effects on the overall network than overpredictions.