Memory-adaptive parallel sparse Cholesky factorization

The problem of Cholesky factorization of sparse positive-definite matrices on distributed-memory multiprocessors is considered. A column-based algorithm with the ability to adapt to the amount of memory available on each processor is presented. Exploiting the available memory allows the local computation on each processor to be ordered so that good local efficiencies and dynamic load balance are achieved. A proof that this distributed algorithm is deadlock-free is given. Experimental results of an implementation of this algorithm on an Intel iPSC/860 multiprocessor system are reported.<
>