The increased availability of parallel computing systems (like multi-core processors or networks of workstations) leads to the question how to minimize costs for execution time or other factors by spreading the computing load onto the available nodes of the network of processors. In the CompSpread project, we especially aim at heterogeneous target architectures like the Grid. The developed techniques should not be restricted to the Grid, however.
The goal of our project is to automate the process of finding a good distribution of computation at compile time and also to adapt that distribution at run time. For this purpose, we want to develope a prototype of a parallelizing compiler for distributed architectures. The generated Code will be using load-balancing techniques and thus dynamically adapting to the features of the network of computers that is used. We want to simplify the parallelization and use of high-performance computing programs in the Grid.
After experimenting with a series of first prototypes that implemented various approaches to load balancing techniques, we want to focus on developing techniques for other aspects of the Grid:
The project CompSpread is funded by the German Research Foundation (Deutsche Forschungsgemeinschaft--DFG); project start: late 2005; project duration: 2(+2) years).