Traffic Prediction on the Internet

Large swings in the demand for content are commonplace within the Internet. To service this demand adequately without over-provisioning requires that the number of servers processing requests be adjusted dynamically with the demand. Since such adjustments require negotiation and the transfer of data across a distributed network, there is a time lag between the decision to adjust and the availability of increased capacity. Failure to perform such adjustments sufficiently in advance can lead to service degradation or even denial of service for a large number of requests. Hence, with enough advance notice, knowledge of future traffic levels can lead to significant improvements in service quality. Performing accurate predictions appears to be a daunting challenge at first glance, but this paper shows that, when applied to Internet traffic, even elementary prediction techniques can have a surprising forecasting power. Our analysis of several different Internet traces reveals that server overloads can often be seen well in advance. This allows steps to be taken to reduce substantially the degradation of service quality. We discuss various systems, such as server and peer-to-peer cooperatives, where traffic predictors (even imperfect ones) can be used to great advantage in reducing latencies and improving resilience to hotspots, whether they be caused by flash crowds, failures in routers, links, etc., and whether they be innocent in origin or produced by vandals or terrorists.