Parallel hierarchical hybrid linear solvers for emerging computing platforms
Emmanuel Agullo, Luc Giraud, Abdou Guermouche, Jean Roman · Comptes Rendus Mécanique · 2011
The design of the extreme-scale platforms that are expected to become available in the forthcoming decade will represent a convergence of technological trends and the boundary conditions imposed by over half a century of algorithm and application software development. These platforms will be hierarchical because they provide coarse grain parallelism between nodes and fine grain parallelism within each node. They are also expected to be very heterogeneous since multi-core chips and accelerators have completely different architectures and potentials. It is clear that such a degree of complexity will embody radical changes that will render obsolete the current software infrastructure for large-scale scientific applications. In this paper, we illustrate a hierarchical algorithmic approach for the implementation of an efficient parallel sparse linear solver that combines direct and iterative methods. Such a hybrid approach exploits the advantages of both numerical techniques and enables the use of several levels and grains of parallelism. This combination express different levels of parallelism and permits an optimal trade-off between numerical and parallel efficiency. Consequently, such a numerical technique appears as a promising candidate for intensive simulations on future many-core parallel platforms.