Steps towards GPU Accelerated Aggregation AMG

Abstract

We present an implementation of AMG with simple aggregation techniques on multiple GPUs. It supports the parallel matrix representations typically used for finite volume discretisation. We employ the ICRS sparse matrix format and the asynchronous exchange mechanism of MPI on CPUs that has been modified to make it suitable for the GPU coprocessors. We show that the solution phase of the standard v-cycle AMG with simple aggregation is accelerated by a factor of up to 12. The solution phase of the more advanced Krylov-accelerated AMG runs faster by a factor of up to 7 on Nvidia TESLA C2070 compared to calculation on Intel X5650 CPUs.