Clustering en Linux

Abstract

High performance computing is essential and indispensable for present research and performed by supercomputers. The current supercomputers are based on a set of computers not so different to those who might have at home but connected by a high-performance network constituting a cluster. This work is a study of several existing clustering solutions for HPC performance. Starting with Beowulf Clusters, we introduce Parallel Virtual Machine as prior technology (though it still can be used by itself) but the study focuses on MPI. This kind of solutions (MPI and PVM) require rewriting of previous programs using parallelization or message passing libraries. As parallelization can be a complex task, also we show another solution available for HPC: MOSIX, that adds clustering capabilities to the operating system kernel in a transparent way to the user and does not require the parallelization of programs nor any special library. It is suitable for the implementation of intensive computing tasks with low/moderate use of input/output. Finally we review the supercomputer Mare Nostrum III in the Barcelona Supercomputing Center. Reading the MN3 User's Guide shows that MPI and multithreading are essential in the development of programs running on current supercomputers.