Parallel out-of-core matrix inversion

(eng) This paper presents a parallel out-of-core algorithm to invert huge matrices, that is when size of matrices is larger than the available physical memory by one or more orders of magnitude. Preliminary performance results are shown for a commodity cluster. An accurate prediction performance model of the algorithm is given. Thanks to the prediction model, optimizations which avoid the overhead of the out-of-core algorithm are derived. Performances of the optimized algorithm using a $O(N)$ memory size are similar to the performances of the best known parallel in-core algorithm using a $O(N^2)$ memory size (where $N$ is the matrix order). There is no memory restriction for inversion of huge matrices