简介:InthispaperwestudythecomputationalperformanceofvariantsofanalgebraicadditiveSchwarzpreconditionerfortheSchurcomplementforthesolutionoflargesparselinearsystems.Inearlierworks,thelocalSchurcomplementswerecomputedexactlyusingasparsedirectsolver.Therobustnessofthepreconditionercomesatthepriceofthismemoryandtimeintensivecomputationthatisthemainbottleneckoftheapproachfortacklinghugeproblems.InthisworkweinvestigatetheuseofsparseapproximationofthedenselocalSchurcomplements.TheseapproximationsarecomputedusingapartialincompleteLUfactorization.Suchanumericalcalculationisthecoreofthemulti-levelincompletefactorizationsuchastheoneimplementedinpARMS.Thenumericalandcomputingperformanceofthenewnumericalschemeisillustratedonasetoflarge3Dconvection-diffusionproblems;preliminaryexperimentsonlinearsystemsarisingfromstructuralmechanicsarealsoreported.
简介:图G的导致的路径数字(G)被定义为G的顶点集合能被划分成的子集的最小的数字以便每个子集导致一条路径。Broere等。如果G是顺序n的一张图,证明了那,那么$\sqrtn\leqslant\rho\left(G\right)+\rho\left({\barG}\right)\leqslant\left\lceil{\tfrac{{3n}}{2}}\right\rceil$。在这份报纸,我们描绘图G为哪个$\rho\left(G\right)+\rho\left({\barG}\right)=\left\lceil{\tfrac{{3n}}{2}}\right\rceil$,在$\rho\left(G\right)上改进更低的界限+\rho\left({\barG}\right)$在一个当n是一个奇怪的整数的平方时,并且为$\rho\left(G\right)决定最好的可能的上面的界限+\rho\left({\barG}\right)$当既不G也不$\barG$孤立顶点时。
简介:AparallelhybridlinearsolverbasedontheSchurcomplementmethodhasthepotentialtobalancetherobustnessofdirectsolverswiththeefficiencyofpreconditionediterativesolvers.However,whensolvinglarge-scalehighly-indefinitelinearsystems,thishybridsolveroftensuffersfromeitherslowconvergenceorlargememoryrequirementstosolvetheSchurcomplementsystems.Toovercomethischallenge,weinthispaperdiscusstechniquestopreprocesstheSchurcomplementsystemsinparallel.Numericalresultsofsolvinglarge-scalehighly-indefinitelinearsystemsfromvariousapplicationsdemonstratethatthesetechniquesimprovethereliabilityandperformanceofthehybridsolverandenableefficientsolutionsoftheselinearsystemsonhundredsofprocessors,whichwaspreviouslyinfeasibleusingexistingstate-of-the-artsolvers.