This dissertation studies the prefetching technique applied to the
World Wide Web from a realistic and practical point of view.  Web
prefetching is a technique to reduce he user perceived latency by
predicting and pre-processing next user accesses.

Until now the open research literature about web prefetching has
focused on theoretical questions and has not taken into account some
of the problems that arise when implementing the technique in real
conditions.  Furthermore, previous works have used simplified models
for evaluation that do not consider how practical issues really affect
the implementation of a prefetching technique.  Moreover, only few
works have considered performance indexes that are relevant to users
when evaluating the benefits that prefetching can achieve.

In order to overcome these three limitations, we developed Delfos, a
web prefetching framework that implements web prediction and
prefetching in a real environment, and can be integrated into the web
architecture without modifying the standard web protocols and in a
compatible way with the existing software products.  Delfos can also
be used to evaluate and compare existing prefetching techniques and
algorithms and to assist in the design of new ones because it provides
detailed statistics reports.  As an example, Delfos is used to
propose, test and evaluate a new technique (Predict At Prefetch, PAP)
able to considerably reduce the user perceived latency with no
additional cost compared to the basic prefetch mechanism.

The prediction algorithms proposed in the research literature that
achieve the highest precision involve a high computational cost, which
is an important drawback for including them in real systems.  To deal
with this disadvantage, a novel low-cost web prediction algorithm
(Referrer Graph, RG) is proposed in this PhD dissertation.  This
algorithm learns from users accesses and builds a Markov model that
can distinguish between dependencies in objects of the same page and
objects of different pages by using the object URI and the referrer in
each request.  RG includes a prune mechanism that controls the
computational resource consumption while sustaining the performance.

This dissertation also includes an empirical study to investigate the
maximum benefits that web users can expect from prefetching techniques
in the current web.  Unlike previous theoretical studies, this work
considers a realistic prefetching architecture using real and
representative traces.  In this way, the influence of real
implementation constraints are considered and analyzed.  The results
obtained show that web prefetching can improve page latency up to 52%
in the studied traces, which encourages researchers to focus future
works on this direction.