> now you have a block-diagonal matrix and distributed every block to a > processor, right? What Solver did you use to solve the single blocks?
Well, I can not give a right answer. From step-32 tut. , If I followed it up correctly, I am using the AMG for the part of laplacian eqn. , and IC for the part of pressure-pressure coupling when making the preconditioner. for stokes equation. And then, when solving the stokes eqn , I am using the GMRES solver. So that, I do not exaclty if every blocks are distributed on the each processors. Do you have a any idea? Just wondering how to avoid to make global sparsity pattern not to allocate big size memory on each processors when solving the matrix whatever solvers are employed. This should be close to real distributed parallelization. Thank you. On Wed, February 11, 2009 9:20 pm, Markus Bürg wrote: > Hello Sungho, > >> Making pattern locally -> success >> Putting global entries in local range -> success > OK, now you have a block-diagonal matrix and distributed every block to a > processor, right? What Solver did you use to solve the single blocks? > > Best Regards, > Markus > > > > On Wed February 11 2009 8:00 pm, [email protected] wrote: >> Message: 2 >> Date: Wed, 11 Feb 2009 18:53:05 -0000 (GMT) >> From: [email protected] >> To: [email protected] >> Subject: [deal.II] Parallelization for Trilinos lib.?? >> Message-ID: <[email protected]> >> Content-Type: text/plain;charset=iso-8859-1 >> >> >> Dear All, >> >> I am trying to make a parallized code with trilinos lib that is much >> similar to step-32 tutorial. It now can be done to make it easily, but >> at >> the same time, gives another problem , parallized optimization. As a >> presumed result of getting study the performance of this code( similar >> to >> step-32' one), every process MUST hold the global whole of sparsity >> pattern that gives a bad things in which we are supposed to solve the >> big >> sized domain. >> >> In my guess, >> For example, I am using the node of cluster that has 4 processors on >> each >> one. If I use the 4 proc., each one holds the same size of sparsity >> pattern which turned out that we need to have a four times memory size. >> >> Such problem is caused from that each proc. has whole of global matrix >> pattern. That was why I try that each proc. has just local matrix >> pattern >> in the code , especially, the part of making sparsity pattern. >> >> Making pattern locally -> success >> Putting global entries in local range -> success >> >> But, It could not be solved the part of trilinos solver. >> >> >> Anyone has been thought of such a similar problem/trouble? If then , >> please share me with your precious experience or give a your advice for >> good performance.! >> >> >> Thank you >> Sungho Yoon >> Leeds >> UK > _______________________________________________ > Dealii mailing list > [email protected] > http://poisson.dealii.org/mailman/listinfo/dealii > _______________________________________________ Dealii mailing list [email protected] http://poisson.dealii.org/mailman/listinfo/dealii
