mplementation of a Grid Environment for High Performance Computing

Abstract

High Performance Computing can be considered as the interconnection via WAN (Wide Area Network) of various Clusters defined, in turn, as either a homogeneous or a heterogeneous set of computers sharing common information and resources, and usually based on local high-speed interconnections. The Grid is, instead, a platform enabling the sharing, the selection and the aggregation of distributed resources through different administrative domains according to their availability and performance able to meet the different Quality of Service user-demands. On a Grid environment, typically, each node has its own resource manager and a particular allocation policy. Such resources are usually assigned in a cluster to a node acting as control workstation. Accordingly, two or more clusters managed in this way within a campus network can represent a Grid infrastructure (sometimes called simply "Grid") as well as are a Grid infrastructure individual machines or groups of machines managed in the same way and distributed across the world. It is easy to imagine the reasons that have fostered the diffusion of large-scale Grid approaches: the possibility of using resources managed by different organizations, regardless of their physical location, has opened new perspectives for the implementation of environments suitable for achieving high performance computing by running massive calculations. In the present thesis work, a prototype Grid architecture for running high performance and high throughput massive calculations, is presented. In particular, Condor and Globus toolkits have been used for implementing the University of Perugia (IT) campus Grid by adopting as a case study three different  (local) interconnected clusters. In addition, the developed Grid has been also validated by providing detailed results for different types of calculations essentially based on the MPICH-G2 parallel library.