Calculs pour les matrices denses : coût de communication et stabilité numérique. (Dense matrix computations : communication cost and numerical stability)

This dissertation focuses on a widely used linear algebra kernel to solve linear systems, that is the LU decomposition. Usually, to perform such a computation one uses the Gaussian elimination with partial pivoting (GEPP). The backward stability of GEPP depends on a quantity which is referred to as the growth factor, it is known that in general GEPP leads to modest element growth in practice. However its parallel version does not attain the communication lower bounds. Indeed the panel factorization represents a bottleneck in terms of communication. To overcome this communication bottleneck, Grigori et al [60] have developed a communication avoiding LU factorization (CALU), which is asymptotically optimal in terms of communication cost at the cost of some redundant computation. In theory, the upper bound of the growth factor is larger than that of Gaussian elimination with partial pivoting, however CALU is stable in practice. To improve the upper bound of the growth factor, we study a new pivoting strategy based on strong rank revealing QR factorization. Thus we develop a new block algorithm for the LU factorization. This algorithm has a smaller growth factor upper bound compared to Gaussian elimination with partial pivoting. The strong rank revealing pivoting is then combined with tournament pivoting strategy to produce a communication avoiding LU factorization that is more stable than CALU. For hierarchical systems, multiple levels of parallelism are available. However, none of the previously cited methods fully exploit these hierarchical systems. We propose and study two recursive algorithms based on the communication avoiding LU algorithm, which are more suitable for architectures with multiple levels of parallelism. For an accurate and realistic cost analysis of these hierarchical algorithms, we introduce a hierarchical parallel performance model that takes into account processor and network hierarchies. This analysis enables us to accurately predict the performance of the hierarchical LU factorization on an exascale platform.