Transactional filesystems

Dissertação de Mestrado em Engenharia Informática

The task of implementing correct software is not trivial; mainly when facing the need for supporting concurrency. To overcome this difficulty, several researchers proposed the technique of providing the well known database transactional models as an abstraction for existing programming languages, allowing a software programmer to define groups of computations as transactions and benefit from the expectable semantics of the underlying transactional model. Prototypes for this programming model are nowadays made available by many research teams but are still far from perfection due to a considerable number of operational restrictions. Mostly, these restrictions derive from the limitations on the use of input-output functions inside a transaction. These functions are frequently irreversible which disables their compatibility with a transactional engine due to its impossibility to undo their effects in the event of aborting a transaction. However, there is a group of input-output operations that are potentially reversible and that can produce a valuable tool when provided within the transactional programming model explained above: the file system operations. A programming model that would involve in a transaction not only a set of memory operations but also a set of file operations, would allow the software programmer to define algorithms in a much flexible and simple way, reaching greater stability and consistency in each application. In this document we purpose to specify and allow the use of this type of operations inside a transactional programming model, as well as studying the advantages and disadvantages of this approach.