This dissertation is focused on the study of the prefetching technique applied to the World Wide Web. This technique lies in processing (e.g., downloading) a Web request before the user actually makes it. By doing so, the waiting time perceived by the user can be reduced, which is the main goal of the Web prefetching techniques.

The study of the state of the art about Web prefetching showed the heterogeneity that exists in its performance evaluation.
This heterogeneity is mainly focused on four issues: 
i) there was no open framework to simulate and evaluate the already proposed prefetching techniques;
ii) no uniform selection of the performance indexes to be maximized, or even their definition;
iii) no comparative studies of prediction algorithms taking into account the costs and benefits of web prefetching at the same time;
and iv) the evaluation of techniques under very different or few significant workloads.

During the research work, we have contributed to homogenizing the evaluation of prefetching performance by developing an open simulation framework that reproduces in detail all the aspects that impact on prefetching performance. In addition, prefetching performance metrics have been analyzed in order to clarify their definition and detect the most meaningful from the user's point of view.
We also proposed an evaluation methodology to consider the cost and the benefit of prefetching at the same time.
Finally, the importance of using current workloads to evaluate prefetching techniques has been highlighted; otherwise wrong conclusions could be achieved.

The potential benefits of each web prefetching architecture were analyzed, finding that collaborative predictors could reduce almost all the latency perceived by users.
The first step to develop a collaborative predictor is to make predictions at the server, so this thesis is focused on an architecture with a server-located predictor.

The environment conditions that can be found in the web are also heterogeneous, so the optimal way of conducting web prefetching may be different depending on these conditions.
We studied under which environments it is better to maximize either the amount of prefetched objects or the size of these objects.
The results showed that when the user's available bandwidth increases prefetching techniques should prefetch a high amount of objects; whereas when the server processing time decreases, the amount of prefetched bytes should be maximized.

By analyzing current prefetching techniques we found that, due to the evolution of the Web, some prefetching techniques that worked well some years ago have their performance dropped on the current Web.
This fact is mainly caused by the high amount of images or similar objects that are currently included in each page. These images notably interfere on the accesses prediction.
Taking into account this unwanted effect, a novel algorithm (DDG) has been proposed and tested dealing with the characteristics of the current web.
The DDG algorithm distinguishes between container objects (HTMLs) and embedded objects (e.g., images) to create the prediction model and to make the predictions.
Results showed that, given an amount of extra requests, DDG always reduces the latency more than the other existing algorithms.
In addition, these results were achieved without increasing the order of complexity of the existing algorithms.