Improving strong scaling of the Conjugate Gradient method for solving large linear systems using global reduction pipelining
share
This paper presents performance results comparing MPI-based implementations of the popular Conjugate Gradient (CG) method and several of its communication hiding (or 'pipelined') variants. Pipelined CG methods are designed to efficiently solve SPD linear systems on massively parallel distributed memory hardware, and typically display significantly improved strong scaling compared to classic CG. This increase in parallel performance is achieved by overlapping the global reduction phase (MPI_Iallreduce) required to compute the inner products in each iteration by (chiefly local) computational work such as the matrix-vector product as well as other global communication. This work includes a brief introduction to the deep pipelined CG method for readers that may be unfamiliar with the specifics of the method. A brief overview of implementation details provides the practical tools required for implementation of the algorithm. Subsequently, easily reproducible strong scaling results on the US Department of Energy (DoE) NERSC machine 'Cori' (Phase I - Haswell nodes) on up to 1024 nodes with 16 MPI ranks per node are presented using an implementation of p(l)-CG that is available in the open source PETSc library. Observations on the staggering and overlap of the asynchronous, non-blocking global communication phases with communication and computational kernels are drawn from the experiments.
READ FULL TEXT