Designing efficient parallel pde solvers in heterogeneous grid environments
Hatem Ltaief · 2007
In the last decade, the explosive interests and needs of the scientific community, in terms of computing power for solving large problems, have emerged. The resources on a single cluster are often very limited and shared among users. The idea of coupling distributed computers became paramount in helping scientists facing new challenges. But this direction has some drawbacks besides the grid constraints. One of them is that most of the applications running on a single machine are quite hardware/software dependent, and porting them to distributed heterogeneous computers is not straightforward. Moreover, while standard parallel applications based on master/slave architectures will overcome the network restrictions, scientific simulations solving Partial Differential Equations based on Domain Decomposition techniques will encounter serious difficulties in getting reasonable performances on the grid. Furthermore, many computational science programs are now being designed to run for days or even for months, and scalability and fault tolerance can create a bottleneck. This becomes more critical when dealing with applications such as weather nowcasting or biomedical problems where accuracy and real-time are highly solicited. The objective of this dissertation is to demonstrate how parallel Partial Differential Equation solvers can efficiently run on a computational grid. The main idea is to minimize the overhead due to the grid constraints using an approach based on parallel numerical algorithms and tools.