Load Balancing for Molecular Dynamics Simulations on Heterogeneous Architectures

Upcoming exascale compute systems are expected to be built from heterogeneous hardware architectures. According to this trend, there exist various methods to handle clusters composed of CPUs, GPUs or other accelerators. Most of these assume that each node has the same structure - for example a dual socket system with an accelerator (GPU or Xeon Phi). The workload is then distributed homogeneously among the nodes. However, not all clusters fulfill this requirement. They might contain different partitions with and without accelerators. Furthermore, depending on the underlying problem to be solved, accelerator cards may perform better in native mode compared to offloading. Besides, various aspects such as cooling may influence the performance of individual nodes. It therefore cannot always be assumed, that the structure and performance of each node and hence the performance of every MPI rank is the same. In this contribution, we apply a k-d tree decomposition method to balance load on heterogeneous compute clusters. The algorithm is incorporated into the molecular dynamics simulation program ls1 mardyn. We present performance results for simulations executed on hybrid AMD Bulldozer-Intel Sandy Bridge, Intel Westmere-Intel Sandy Bridge and Intel Xeon-Intel Xeon Phi-architectures. The only prerequisite for the proposed algorithm is a cost estimation for different decompositions. It is hence expected to be applicable to a variety of n-body scenarios, for which a domain decomposition is possible.