Distributed computing

In 1998, inside the research project ALGEN, funded by the National Institute for the Nuclear Physics (INFN, department of Catania), Giuseppe Patanè began the construction and the administration of a powerful system for distributed computing. Till now, the funds obtained for its growth and management have been given by several sponsors (as well as the INFN, of course): the National Institute for the Physics of Matter (INFM, department of Messina), the Computing Center of the University of Messina and the University of Messina itself. At present, the distributed system (called ``the MULTISOFT machine'') is composed of 32 PCs and 2 servers connected by a fast-ethernet LAN. The operating system is LINUX and the inter-process communication is realized by a message passing approach using PVM. A detailed description of the system and of the particular techniques for its management can be found in [MyParClustering]. At the moment, the system is being upgraded by equipping each host with 4 fast-ethernet network interface cards connected by 24-port switches.

In the scientific world, several types of applications requires high computational power. In those cases, a parallel computing system can be a valuable instrument for reaching high performances. Some techniques, such as clustering ones, need also large communication bandwidths; others, such as evolutionary algorithms, require less communication resources. There are also other family of algorithms, that, although heavily use the communication system, can be optimized by adjusting the topology of the communication network. In that case, they can reach high performances also when using low-speed links. A valuable instrument for studying such algorithms has been identified just in cluster of PCs connected by easily reconfigurable multiple fast-ethernet networks (i.e. each host has multiple network cards). A set of switches seems to be an easy way for testing several and very different logical interconnection topologies such as linear, planar, tridimensional, ring-shaped, etc. In fact, many of such topologies can be tested by simply exploiting the ability of switches in allowing high-rate data transfer between its ports.

Distributed clustering

Several clustering algorithms have been designed and implemented. For a description of the most important results obtained, the reader can look at the related section.