Stefan Georgiev, Wolfgang Schreiner,
"Evaluation of Cluster Middleware in a Heterogeneous Computing Environment"
, Serie Internationaler Universitätslehrgang Informatics: Engineering & Management (ISI Hagenberg), 7-2009
Evaluation of Cluster Middleware in a Heterogeneous Computing Environment
Sprache des Titels:
Deploying a Beowulf-type high-performance cluster is a challenging task. Many problems, regarding the underlying hardware infrastructure and the used software components need to be solved in order to make a set of machines work as a single computer. The task becomes even more complicated when heterogeneous commodity hardware platforms are utilized. This research is provoked by the idea of using a company�s old desktop computers for achieving computing performance at a low cost. While the existing cluster middleware provides many solutions for utilizing distributed resources, it still fails to be suitable for some cluster setups. This thesis represents the results of an evaluation in the field of heterogeneous high-performance computing. A systematic analysis of available tools for high-performance clustering aims to reveal both their strong and their weak sides. In addition, a detailed evaluation is provided from the point of view of administrators and users. The thesis describes the process of building a fully-functioning parallel environment using three different tools for cluster deployment - OSCAR v6.0.2, ROCKS v5.1 and CAOS-NSA v 1.0. These tools are compared in detail, so that the most suitable one can be determined. As a result two test environments are built. Numerous examples of tests runs, executed on the two environments, are described. Usability, with regards to different tools like resource managers, schedulers, and MPI, is assessed taking into consideration the underlying hardware platforms. The capabilities of the clusters are evaluated against a parallel application for route optimization.