Scalable execution control of grid-based scientific applications on parallel systems

Computational sciences require more computer power than current hardware technology is able to deliver at acceptable costs. Exploiting parallelism is the most promising way to obtain substantial performance increases. However, the absence of suitable parallel software delays a wider use of parallel systems. Automatic parallelization can help to overcome this problem. General automatic parallelization is a very difficult or even unsolvable problem. This paper shows that automatic parallelization becomes feasible when restrictions are imposed on the class of applications to be handled. The system presented in this paper is specialized for grid-based numerically intensive computations. An important feature of systems for parallel processing is scalability. Mechanisms to make the execution control of the automatic parallelization system scalable are presented. The validity of the concepts is shown by performance data.<
>