We have shown that accurate Web prefetching is possible based on following HREFs in recently fetched pages. Letting the client control prefetching and aging of usage statistics has many advantages and may be the only practical approach in a world where proxies are commonplace. However, placing control at the client is also problematic because many pages contain large numbers of hyperlinks, and simply prefetching them all is worse than nothing. Also the client cannot be expected to have a good understanding of HREF reference patterns unless the page is one read frequently and/or the page rarely changes.
Our approach to this problem is to have clients pass record of their references up to the relevant server, which then distributes them to all clients. It is hypothesized that HREFs within a page are strongly skewed to ``hot'' and ``cold,'' so that one client can learn from the usage patterns of others. The results in Section 5.1 bear this out.
The information exchange between clients and servers complicates deployment; however, there are other examples of recent work that depend on a similar flow of information [23, 7], suggesting that the idea may be useful more generally.