Experiences with scientific applications on an SCI-based Linux cluster

Today clusters of PCs are powerful computing platforms offering an inexpensive alternative to traditional supercomputers. Besides the low price, the major reason for the emergence of PC clusters is the availability of a wealth of public domain as well as important commercial software on top of the Linux operating system. Furthermore, a PC cluster can be built in a modular fashion enabling a potential customer to perfectly meet his requirements. For instance, the number of nodes, the node types, and the network can be chosen. In this note, experiences with an SCI-based Linux cluster installed at the Computing Center of Aachen University of Technology are reported. Timing results for two different applications, a two-dimensional fast Fourier transform and a molecular dynamics simulation, are presented. Our experiences demonstrate the importance of a fast network when dealing with large-scale scientific applications on a Linux cluster.