PROTOTIPO DI UN'INFRASTRUTTURA VIRTUALE DI CALCOLO ESPANDIBILE DINAMICAMENTE PER L'ESPERIMENTO ALICE.

Although the LHC experiments at CERN - ALICE, ATLAS, CMS and LHCb - are widely known as a physical tool that will hopefully provide a better explanation of interactions of complex and dynamically evolving systems of finite size, thus extending the theory of elementary particles and their fundamental interactions (the Standard Model), what is less evident is the ancillary technology that has been developed in order to make this possible. Among all other technologies, computing plays a fundamental role in LHC experiments: simulating, storing and analyzing data requires both hardware and software infrastructures that are under constant and heavy development. LHC experiments cannot simply rely on existing "stable" and commercialized computing infrastructures, because they constitute a completely new use case in the computing world: this is the reason why physicists write most of the software they need on their own, and why these infrastructures are to be considered "leading-edge technologies", and may even constitute, in the near future, the bases for a change in the way computers are used by "ordinary" people, as it happened for the World Wide Web, born to satisfy the need to share documents within the HEP community. The main large-scale computing infrastructure introduced by the LHC experiments is the Grid, a distributed computing infrastructure designed in order to allow equal access to data and computing power by every Physicist in every part of the world. The only inconvenients from the technical point of view are that the Grid is not interactive, and the Physicist should wait a random time in order for its tasks to be executed, making it not usable for tasks that need, for instance, to be repeated several times with different parameters in order to choose the better ones: this is the reason why a solution like PROOF was developed, and particularly well integrated with the ALICE experiment's software. Interactive analysis is complementary to the Grid and does not represent an alternative to it: the two models do have to coexist and share resources because the costs for a dedicated interactive facility could seldom be afforded. This is the reason why efforts were made in order to efficiently share resources between batch and interactive tasks, mainly by running daemons as ordinary batch jobs that enable interactivity on demand on the Grid. This thesis is about the development, implementation and testing of a working prototype of a PROOF facility in Torino's Tier-2 centre that shares computing resources with the Grid dynamically (i.e. resources can be assigned or cut without any service interruption) through the use of virtual machines - a method that moves the problem of sharing resources from the Grid level to the machine level. We call our facility the Virtual Analysis Facility, and we'll discuss the advantages and disadvantages of this approach, by showing with the aid of benchmarks that virtualizing a high-performance and high-throughput computing infrastructure is nowadays feasible.