2004 , Volume 9, № 6, p.34-39

Clusters have become a ubiquitous computing platform in numerical simulation. Their first and foremost advantage is seen to be the low entry price for smaller systems and the potential to expand the systems to thousands of processors. Furthermore, they provide standard technology which helps to keep portability and hence reduces the burden of the software developer. However, there is also the other side of the coin. The increased number of components also increases the level of complexity. And with this the stability of systems goes down while the total costs for running such systems go up.

Many scientists focus on only one aspect of this problem. Without a general and global view, however, progress is unlikely to be achieved. One obstacle for progress is the short sighted view of clusters as cheap systems. Typically only investment costs are considered and further costs are ignored because they do not become visible soon in a scientific setting. An analysis of the total cost of ownership shows that this approach is totally wrong. We find that in order to turn clusters into productive systems human costs have to be reduced. Consequently we have to take steps in a variety of fields if we want to succeed.

*68M10 Network design and communication

68U20 Simulation