000200820 001__ 200820
000200820 005__ 20210129215543.0
000200820 0247_ $$2Handle$$a2128/8618
000200820 037__ $$aFZJ-2015-03205
000200820 082__ $$a004
000200820 1001_ $$0P:(DE-HGF)0$$aSchelthoff, Christof$$b0
000200820 1112_ $$a13. Workshop über Parallelverarbeitung$$cLessach$$d1994-09-25 - 1994-10-01$$wAustria
000200820 245__ $$aPolynomial Preconditioning for the Conjugate Gradient Method on Massively Parallel Systems
000200820 260__ $$aClausthal-Zellerfeld$$bInstitut für Informatik$$c1995
000200820 29510 $$aProceedings des 13. Workshops über Parallelverarbeitung
000200820 300__ $$a150-167
000200820 3367_ $$0PUB:(DE-HGF)8$$2PUB:(DE-HGF)$$aContribution to a conference proceedings$$bcontrib$$mcontrib$$s1432105111_25856
000200820 3367_ $$033$$2EndNote$$aConference Paper
000200820 3367_ $$2ORCID$$aCONFERENCE_PAPER
000200820 3367_ $$2DataCite$$aOutput Types/Conference Paper
000200820 3367_ $$2DRIVER$$aconferenceObject
000200820 3367_ $$2BibTeX$$aINPROCEEDINGS
000200820 4900_ $$aInformatik-Bericht$$v95/1
000200820 520__ $$aA frequently used iterative algorithm for solving large, sparse, symmetric and positiv definite systems of linear equations is the method of conjugate gradients (CG).This method requires one matrix-vector product and some dot products in each iteration. Convergence is dependent on the condition number of the coefficient matrix. So preconditioning techniques are used to reduce the number of iterations.In this context, polynomial preconditioning was developed. This method decreases the total number of dot products by reducing the total number of iterations. Of course, some additional work has to be done for the preconditioning. When a polynomial of degree k is used, k matrix-vector products per iteration have to be calculated rather than one. On scalar machines, this shift between matrix-vector products and dot products influences the performance of the algorithm only slightly. On massively parallel systems, dot products require global synchronization, while the calculation of matrix-vector products merely results in communication with a small number of processors. Hence, polynomial preconditioned CG seems to scale better than CG without preconditioning. Of course, this is not the case in general. The performance of this preconditioner depends on several issues, e.g., the sparsity pattern and the eigenvalue distribution of the matrix, an efficient communication scheme for the matrix-vector products and the time needed for global synchronization of the specific parallel machine. The actual implementation used here is based on Chebyshev polynomials. Performance tests were carried out on the Intel Paragon XP/S 10 with 140 nodes at the Research Centre Jülich (KFA). The CG method with polynomial preconditioning shows better performance and scalability than the basic method on massively parallel machines. Additionally there are some numerical advantages like a higher accuracy and an increased stability.
000200820 536__ $$0G:(DE-HGF)POF2-899$$a899 - ohne Topic (POF2-899)$$cPOF2-899$$fPOF I$$x0
000200820 7001_ $$0P:(DE-HGF)0$$aBasermann, Achim$$b1
000200820 8564_ $$uhttps://juser.fz-juelich.de/record/200820/files/ib-9423.pdf$$yOpenAccess
000200820 8564_ $$uhttps://juser.fz-juelich.de/record/200820/files/ib-9423.gif?subformat=icon$$xicon$$yOpenAccess
000200820 8564_ $$uhttps://juser.fz-juelich.de/record/200820/files/ib-9423.jpg?subformat=icon-1440$$xicon-1440$$yOpenAccess
000200820 8564_ $$uhttps://juser.fz-juelich.de/record/200820/files/ib-9423.jpg?subformat=icon-180$$xicon-180$$yOpenAccess
000200820 8564_ $$uhttps://juser.fz-juelich.de/record/200820/files/ib-9423.jpg?subformat=icon-640$$xicon-640$$yOpenAccess
000200820 8564_ $$uhttps://juser.fz-juelich.de/record/200820/files/ib-9423.jpg?subformat=icon-144$$xicon-144$$yOpenAccess
000200820 909CO $$ooai:juser.fz-juelich.de:200820$$pdnbdelivery$$pVDB$$pdriver$$popen_access$$popenaire
000200820 9132_ $$0G:(DE-HGF)POF3-899$$1G:(DE-HGF)POF3-890$$2G:(DE-HGF)POF3-800$$aDE-HGF$$bForschungsbereich Materie$$lForschungsbereich Materie$$vohne Topic$$x0
000200820 9131_ $$0G:(DE-HGF)POF2-899$$1G:(DE-HGF)POF2-890$$2G:(DE-HGF)POF2-800$$3G:(DE-HGF)POF2$$4G:(DE-HGF)POF$$aDE-HGF$$bProgrammungebundene Forschung$$lohne Programm$$vohne Topic$$x0
000200820 915__ $$0StatID:(DE-HGF)0510$$aOpenAccess
000200820 9201_ $$0I:(DE-Juel1)JSC-20090406$$kJSC$$lJülich Supercomputing Center$$x0
000200820 9201_ $$0I:(DE-Juel1)VDB62$$kZAM$$lZentralinstitut für Angewandte Mathematik$$x1
000200820 980__ $$acontrib
000200820 980__ $$aVDB
000200820 980__ $$aFullTexts
000200820 980__ $$aUNRESTRICTED
000200820 980__ $$aI:(DE-Juel1)JSC-20090406
000200820 980__ $$aI:(DE-Juel1)VDB62
000200820 9801_ $$aFullTexts
000200820 981__ $$aI:(DE-Juel1)VDB62