high performance computing on graphics processing units: hgpu.org

We present an extension of a GPU-based matrix inversion algorithm for distributed memory contexts. Specifically, we implement and evaluate a message-passing variant of the Gauss-Jordan method (GJE) for matrix inversion on a cluster of nodes equipped with GPU hardware accelerators. The experimental evaluation of the proposal shows a significant runtime reduction when compared with both the distributed non-GPU implementation of GJE and a conventional method based on the LU factorization (as implemented in ScaLAPACK). In addition to this, our proposal leverages the aggregated capacity of the GPU memories in the cluster to overcome the constraints imposed by the reduced memory space of these devices.