Software and Performance Engineering for Iterative Eigensolvers

The complexity of the latest HPC architectures increasingly limits the productivity of researchers in numerical algorithms and the `time to market' for parallel algorithms. Implementing a new method on a supercomputer today involves at least three levels of parallelism and typically several programming models like MPI, OpenMP and CUDA. Frameworks like Trilinos and PETSc have since many years been useful for testing new ideas in parallel algorithms. But when it comes to e.g. CPU/GPU clusters they fail to deliver convincing performance to date. We look at sparse solvers from a software engineer's point of view and advocate a programming model we call `SPMD+OK', introducing performance models in the test-driven development process and a strategy to facilitate the integration of algorithmic developments into existing applications. Harnessing the peta-scale with the block Jacobi-Davidson method is used as a running example in this talk, and the libraries GHOST and PHIST are presented (https://bitbucket.org/essex/)