Xeon Phi acceleration of domain decomposition iterations via heterogeneous active messages
Lukáš Malý, Jan Zapletal, Michal Merta, Martin Čermák · AIP conference proceedings · 2018
We present an acceleration strategy for a domain-decomposition iterative solver based on local Schur complements with respect to the skeleton of the computational domain. In finite element tearing and interconnecting (FETI) this results in an iterative application of a large number of dense matrices. For the offload of such kernels to the Intel® Xeon Phi™ coprocessors we use the Heterogeneous Active Messages (HAM) library. A simple load balancing strategy is presented to efficiently utilize both the host CPU and the available coprocessors during individual iterations.