high performance computing on graphics processing units: hgpu.org

Finite element methods (FEM) are most widely used for simulation of structural dynamics problems. Due to their highly compute intensive nature, these methods are used with domain decomposition where the problem is divided into subdomains which are individually solved and coupled together to obtain the final solution. One of the latest and most efficient approach use multi-scale Newmark methods with domain decomposition, to solve structural dynamics problem. Determining the optimal coupling order of the decomposed subdomains is very difficult and hence even the performance of the efficient FEM approaches is limited. In our previous work, we developed some heuristics based on cost model to determine the optimal coupling order. It is observed that even with optimal coupling order most of the time is spent in computing linear algebra routines. In this work, we show that the overall performance of multiscale methods can be improved substantially by speeding up linear algebra calculations by offloading them to the GPU, thereby reducing the latency of coupling operations. We show that the optimal solution is a hybrid solution that combines the strengths of both CPU and GPU platform. We achieve up to 270% performance improvement for a certain FEM problem.