Um serviço de cache de disco para grades computacionais entre-pares.

Abstract

Computational peer-to-peer (P2P) grids are a cheap and efficient platform to run applications such as "bag-of-tasks", consisting of a large number of tasks that do not need to communicate amongst themselves and can be executed independently. When these applications handle a large amount of data, it is crucial that the overhead due to data transfer does not impact application performance. Fortunately, many of these applications have a high rate of data reuse, in the same or in successive runs of the application, which allows a caching strategy to be used, so to minimize the loss of performance related to data movement. However, implementing a caching system atop non-dedicated resources and subject to usage policies that are not necessarily homogeneous is not a trivial task. In this work we propose a mechanism for disk cache management that gives support to this type of application by reducing the number of transfers amongst administrative domains. The mechanism promotes collaboration amongst the peers forming the P2P grid and results in a fair and efficient allocation of the disk space provided by the peers. Through simulation experiments, we demonstrated that the mechanism has reduced the amount of data transfers amongst administrative domains and during resource contention free-riders are marginalized in favor of collaborators, regardless of the number of nodes that comprise the system. Finally, a case study executed over a controlled P2P computational grid showed that the mechanism has reduced the amount of transfers, which consequently reduced the execution times of the applications when compared to the execution of the same applications without using the disk cache management mechanism proposed.