The performance metrics we use in our experimental evaluation are the Hit Ratio, and Traffic Increase. The Hit Ratio is the ratio of the requests that are serviced from prefetched documents, to the total number of requests. It represents the ``guessing ability'' of our algorithm. The higher this ratio is, the lower the client latency and the server load.
The Traffic Increase is the increase in traffic due to unsuccessfully prefetched documents. Since no program can predict the future, some of the prefetched documents will not be actually requested, and thus, they should not have been prefetched in the first place.
The design of a prefetching policy, is the art of balancing the conflicting factors of Hit Ratio and Traffic Increase, while trying to guess future requests. At one extreme, if an algorithm never prefetches any documents, it will not suffer any traffic increase, but it will also have zero hit ratio. At the other extreme, a very aggressive prefetching algorithm may prefetch the entire Web (assuming enough resources), but this would saturate the network with prefetched documents that will never be requested.
In all our experiments we assume that a prefetch document stays in the cache for the entire duration of a time interval, so that a small cache will not distort our results. This assumption is not unrealistic, however, since the cost of magnetic disks is low, usually lower than network bandwidth cost.