Lapack Working Note 56 Conjugate Gradient Algorithms with Reduced Synchronization Overhead on Distributed Memory Multiprocessors

The standard formulation of the conjugate gradient algorithm involves two inner product computations. The results of these two inner products are needed to update the search direction and the computed solution. Since these inner products are mutually interdependent, in a distributed memory parallel environment their computation and subsequent distribution requires two separate communication and synchronization phases. In this paper, we present three related mathematically equivalent rearrangements of the standard algorithm that reduce the number of communication phases. We present empirical evidence that two of these rearrangements are numerically stable. This claim is further substantiated by a proof that one of the empirically stable rearrangements arises naturally in the symmetric Lanczos method for linear systems, which is equivalent to the conjugate gradient method.