Back to search

Aristotle University Of Thessaloniki (AUTH)

Εφαρμογή μεθόδων παράλληλης επεξεργασίας στη μελέτη της ταλαντωτικής συμπεριφοράς σύνθετων μηχανικών συστημάτων

Abstract

dc:description

The use of Automated Multilevel Substructuring (AMLS), for the analysis of constantly bigger models created the need for the extension of the method to systems of parallel processing. The main characteristic of the method is that for an efficient implementation a number of choices have to be made regarding the algorithms used in the intermediate steps. In this dissertation a fully distributed version of the general AMLS methodology has been designed and implemented (called MLDS) and the factors which influence the efficiency of the algorithm have been analyzed. For this purpose, the solution of static problems and symmetric eigenproblems has been studied as well. Initially, the Cholesky factorization of dense matrices and the multifrontal method for the factorization of sparse matrices were presented. For the parallelization of the latter a static weighted subtree to subcube mapping has been used. As the number of processors increases the efficiency decreases due to the growth of the volume of communication and the load imbalance because of the static mapping. Afterwards, the Lanczos method was described and two methods of its parallelization were discussed. With the first method, the frequency interval of interest is divided in subintervals each of which is processed by one processor. Although, no communication between the processors is required, the efficiency is limited because of the different calculation cost if each subinterval. In addition, the distributed execution usually involves a greater number of numerical operations than the corresponding serial one. With the second method, the most time consuming operations of the solution of static problems and orthogonalization are individually parallelized. The efficiency is low because of the high cost of communication. Finally, the MLDS algorithm was presented. First the identification of the mathematical substructures and the creation of the assembly tree were discussed. Then the parallelization of the intermediate steps was described. Briefly, all the processors are divided in groups, except for one, which is called main global process. Each substructure is assigned to a group. The purpose of the main global process is to prepare, send and collect the data involved in the various staged of the processing of each substructure. Generally, the accuracy and efficiency of the algorithm depends on the number of levels of the assembly tree and the cut-off frequency of the individual eigenproblems.

Degree

thesis:*
Grantor dc:publisher
Aristotle University Of Thessaloniki (AUTH)
Year dc:date
2009

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Sikelis, Konstantinos
  • Σικέλης, Κωνσταντίνος

Subjects

dc:subject × 20

Rights

Language dc:language
gre

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/19310

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Sikelis, Konstantinos; Σικέλης, Κωνσταντίνος. Εφαρμογή μεθόδων παράλληλης επεξεργασίας στη μελέτη της ταλαντωτικής συμπεριφοράς σύνθετων μηχανικών συστημάτων. Aristotle University Of Thessaloniki (AUTH), 2009. http://hdl.handle.net/10442/hedi/19310