EIS Lab homeResearch Snapshots, Accolades, ...Faculty, Staff, Students, and AlumniPapers, Reports, Theses, ...Analysis Theory & Methodology, X-Analysis Integration (XAI), Change Management, Engineering Information Technology, ... Projects, Sponsors, Toolkits, ...Conferences, Workshops, Thesis Presentations, ... Georgia Tech search engineCourses, Tools, Related Organizations, Directions & Locale Guides, ...

Parallel Subdomain Method For Massively Parallel Computers (Finite Element)


Su, (Philip) Shin-Chen (1992) Doctoral Thesis, Georgia Institute of Technology, Atlanta.


The presented work provides the synthesis of a parallel subdomain method for the finite element method (FEM) considering massively parallel computer environments. Massively parallel processing will provide effective computing speeds needed for future large-scale analysis. The key to achieving improved performance over the traditional global model method on this new class of computer is the development of parallel subdomain software designs tailored to massively parallel computing. A new parallel subdomain method has been developed and performed on the realistic numerical examples on the selected local/shared memory type moderately and massively parallel computers for linear static FEM solutions.

A completely parallel global model for linear static analysis is proposed which consists of four parallel program modules. These parallel modules incorporate the local/shared data storage technique on the basis of a local/shared memory base. Two key modules, matrix decomposition and displacement solution, involve substantial synchronization. For these procedures an in-depth performance study is performed which includes the implementation of the parallel matrix decomposition module. The multiprocessing performance is evaluated for two- and three-dimensional linear static problems.

A new parallel subdomain method design is presented for linear static analysis and consists of eight parallel program modules. Most of the parallel modules are similar to the parallel global model method. The domain partition is performed in one and two direction cases. The processor mapping is performed through a simulated procedure. Two types of shared memory, distributed and virtual shared memory, are incorporated with the local/shared data storage technique. An in-depth study is performed for the important memory management research issue. The parallel subdomain method performance is evaluated and compared with the parallel global model for two- and three-dimensional linear static problems under moderately and massively parallel computer environments.

A comparative study showed that the proposed parallel subdomain method is well suited for general-purpose FEM systems using massively parallel local/shared memory computers, and that it provides a more solid foundation to the concept that subdomain strategy can provide significant improvement in speed-up for large-scale FEM problems running on massively parallel computers.

Manuscript: order via UMI