Open Access

ARTICLE

Hybrid Parallelism of Multifrontal Linear Solution Algorithm with Out Of Core Capability for Finite Element Analysis

by Min Ki Kim¹, Seung Jo Kim²

1 Seoul National University, Seoul, Mechanical Aerospace Engineering, Korea
2 Korea Aerospace Research Institute, Daejeon, President, Korea

Computer Modeling in Engineering & Sciences 2012, 84(4), 297-332. https://doi.org/10.3970/cmes.2012.084.297

Abstract

Hybrid parallelization of multifrontal solution method and its parallel performances in a multicore distributed parallel computing architecture are represented in this paper. To utilize a state-of-the-art multicore computing architecture, parallelization of the multifrontal method for a symmetric multiprocessor machine is required. Multifrontal method is easier to parallelize than other direct solution methods because the solution procedure implies that the elimination of unknowns can be executed simultaneously. This paper focuses on the multithreaded parallelism and mixing distributed algorithm and multithreaded algorithm together in a unified software. To implement the hybrid parallelized algorithm in a distributed shared memory environment, two innovative ideas are proposed to reduce the required physical memory. The first idea is pairing the factorization matrix. Pairing two factorization matrices in two threads with a square matrix for parallel factorizing reduces approximately half of the required physical memory. The second idea is splitting the factorization matrix, which enhances the solver performance by removing additional memory. The use of an out-of-core storage is necessary for the structural analysis with limited computing resources. To improve the computing efficiency with the out-of-core storage, it is essential to cache the factored data into free physical memory. To maximize the amount of cached data, selective data caching and recovery is used. Parallel performances of the proposed method are shown. The multithreaded parallel multifrontal algorithm is more computationally efficient than the serial algorithm.

---

Keywords

---