Parallel parametric study of finite-deformation elastic–plastic analyses using dynamic load balancing with execution orders

Abstract

This study developed a Message-Passing-Interface-based dynamic load balancing algorithm with four execution orders for the parallel parametric study of nonlinear finite element analyses. The dynamic load balancing algorithm is based on the master–worker method. The execution orders are descending and ascending orders for the number of nodes and for the number of incremental steps. In particular, Execution Orders A and B are the node-major step-minor descending order and the step-major node-minor descending order, respectively. These methods were applied to two numerical examples of finite-deformation elastic–plastic problems, namely, a compact tension specimen and a notched cylinder. The dynamic load balancing algorithm with Execution Order B was the fastest in the first numerical example, whereas that with Execution Order A was the fastest in the second. However, the differences between the results for Execution Orders A and B were slight in both numerical examples. Hence, both Execution Orders A and B are able to achieve practically short total computational time. The speedups from the static load balancing algorithm to the dynamic one for the two numerical examples were 1.54 and 1.37, respectively. The speedups from the ascending execution order to the descending execution order were 1.24 and 1.29, respectively. The results demonstrated that both dynamic load balancing and an appropriate execution order are necessary to reduce the load imbalance and thereby achieve short total computational time.