ON TUNING OF FINITE ELEMENT LOAD BALANCING FRAMEWORK

Abstract

: Many engineering problems are extremely demanding to solve even on recent hardware. Numerical solutions of these problems on parallel computers can significantly reduce computational time by performing selected tasks concurrently. This paper deals with tuning up the parallel load balancing framework of the finite element software, which is based on domain decomposition paradigm for distributed memory model. The paper describes the technique to determine the actual weights comparing computational performance of individual processing units. These weights are fundamental inputs for mesh (re)partitioning that has to be performed at the beginning of the simulation and whenever the load imbalance is significant. The capabilities and performance of the proposed technique are evaluated on the benchmark problem and discussed. Abstract: Many engineering problems are extremely demanding to solve even on recent hardware. Numerical solutions of these problems on parallel computers can significantly reduce computational time by performing selected tasks concurrently. This paper deals with tuning up the parallel load balancing framework of the finite element software, which is based on domain decomposition paradigm for distributed memory model. The paper describes the technique to determine the actual weights comparing computational performance of individual processing units. These weights are fundamental inputs for mesh (re)partitioning that has to be performed at the beginning of the simulation and whenever the load imbalance is significant. The capabilities and performance of the proposed technique are evaluated on the benchmark problem and discussed.