PARALUTION is a library for sparse iterative methods which can be performed on various parallel devices, including multi-core CPU, GPU (CUDA and OpenCL) and Intel Xeon Phi.
The 1.0 version of the PARALUTION Library supports multi-node and multi-GPU configuration via MPI. All iterative solvers support global operations (i.e. distributed matrices and vectors) and all preconditioners can be used in a block-Jacobi fashion locally on each node/GPU. In addition, the software provides a global (fully distributed) Pair-Wise AMG solver.
The global AMG solver builds all the levels (and the corresponding operators) in parallel (i.e. by utilizing all MPI ranks). All AMG solvers have optimized support for non-linear/time stepping schemes. Typically, in such configuration the matrix values are changed while the sparsity pattern of the matrix remains the same. In this case, after the first call, the our AMG solvers need significantly less time to re-build. Also, in the new version all multi-grid solvers support V, W and K-cycles.
With the new release, we also provide multi-node benchmarks available on http://www.paralution.com/multi-node-benchmarks/
The library comes with a dual license model, where the user can choose between three versions.
For details, visit http://www.paralution.com