Return to search

Analysis and design development of parallel 3-D mesh refinement algorithms for finite element electromagnetics with tetrahedra

Optimal partitioning of three-dimensional (3-D) mesh applications necessitates dynamically determining and optimizing for the most time-inhibiting factors, such as load imbalance and communication volume. One challenge is to create an analytical model where the programmer can focus on optimizing load imbalance or communication volume to reduce execution time. Another challenge is the best individual performance of a specific mesh refinement demands precise study and the selection of the suitable computation strategy. Very-large-scale finite element method (FEM) applications require sophisticated capabilities for using the underlying parallel computer's resources in the most efficient way. Thus, classifying these requirements in a manner that conforms to the programmer is crucial. / This thesis contributes a simulation-based approach for the algorithm analysis and design of parallel, 3-D FEM mesh refinement that utilizes Petri Nets (PN) as the modeling and simulation tool. PN models are implemented based on detailed software prototypes and system architectures, which imitate the behaviour of the parallel meshing process. Subsequently, estimates for performance measures are derived from discrete event simulations. New communication strategies are contributed in the thesis for parallel mesh refinement that pipeline the computation and communication time by means of the workload prediction approach and task breaking point approach. To examine the performance of these new designs, PN models are created for modeling and simulating each of them and their efficiencies are justified by the simulation results. Also based on the PN modeling approach, the performance of a Random Polling Dynamic Load Balancing protocol has been examined. Finally, the PN models are validated by a MPI benchmarking program running on the real multiprocessor system. The advantages of new pipelined communication designs as well as the benefits of PN approach for evaluating and developing high performance parallel mesh refinement algorithms are demonstrated.

Identiferoai:union.ndltd.org:LACETR/oai:collectionscanada.gc.ca:QMM.103003
Date January 2006
CreatorsRen, Da Qi.
PublisherMcGill University
Source SetsLibrary and Archives Canada ETDs Repository / Centre d'archives des thèses électroniques de Bibliothèque et Archives Canada
LanguageEnglish
Detected LanguageEnglish
TypeElectronic Thesis or Dissertation
Formatapplication/pdf
CoverageDoctor of Philosophy (Department of Electrical and Computer Engineering.)
Rights© Da Qi Ren, 2006
Relationalephsysno: 002603336, proquestno: AAINR32234, Theses scanned by UMI/ProQuest.

Page generated in 0.0024 seconds