Towards Distributed Heterogenous High-Performance Computing with ViennaCL

Abstract. One of the major drawbacks of computing with graphics adapters is the limited available memory for relevant problem sizes. To overcome this limitation for the ViennaCL library, we investigate a partitioning approach for one of the standard benchmark problems in High-Performance Computing (HPC), namely the dense matrix-matrix product. We apply this partitioning approach to problems exceeding the available memory on graphics adapters. Moreover, we investigate the applicability on distributed memory systems by facilitating the Message Passing Interface (MPI). Our approach is presented in detail and bench-mark results are given.