Reducing the effect of global communication in \(\text{GMRES} (m)\) and CG on parallel distributed memory computers (Q1904023)

The paper studies the reduction of the communication overhead introduced by inner products when the iterative methods like CG and GMRES are parallelized. Two ways of improvement are suggested. The first way consists in assembling the results of a number of inner products collectively after grouping some orthogonalization steps. The second way uses some rearranging of the computation steps to achieve a possibility of overlapping of communication with computation. The effect of both ways of improvement is assessed by deriving computer time estimations and by running test problems.