Towards Understanding Optimal Load-Balancing of Heterogeneous Short- Range Molecular Dynamics

S Hirschmann and D Pfluger and CW Glass, 2016 23RD IEEE INTERNATIONAL CONFERENCE ON HIGH PERFORMANCE COMPUTING WORKSHOPS (HIPCW 2016), 130-141 (2016).

DOI: 10.1109/HiPCW.2016.24

For heterogeneous dynamic short-range molecular dynamics simulations it is critical to employ suitable loadbalancing methods to minimize the time to solution. However, designing, selecting and parametrizing the optimal loadbalancing method is a complex task which depends on detailed properties of the simulation scenario. The main challenge in balancing the load of molecular dynamics simulations is the extreme difference in load for scenarios with a heterogeneous particle density, which can easily reach 4-6 orders of magnitude. Therefore, heterogeneity is deemed to be a relevant property. In this paper, we formulate a suitable metric to reliably quantify heterogeneity. We apply this metric, which is based on the binning of particles and the evaluation of statistical moments, to example scenarios, and we correlate the results to the performance of five load balancing methods. Furthermore, we identify the load dynamics as a second relevant property. It quantifies how rapidly the load varies over time, and we introduce corresponding metrics. We show that the load dynamics can be used to determine how long the benefits of a specific partitioning are expected to last. The results indicate that these metrics are useful to differentiate between scenarios, and that they facilitate reasoning over the complex relationship between particle simulation scenarios and optimal load balancing methods. This work is a first step towards understanding this relationship, and it introduces key concepts that we regard as crucial in this process.

Return to Publications page