[petsc-dev] LU for MPIAIJ

[Stefano Zampini]

I will also be interested in Jed's proposal: currently PCBDDC has an option
(pc_bddc_coarse_problem_type) which make it switching  between gathering
the coarse problem on one processor (sequential option), gathering to all
(replicated option), solve with PCREDUNDANT (parallel option), this when
doing exact primal coarse solves (the same holds true also for the FETI-DP
implementation which uses PCBDDC when doing MatMult). These steps are
performed using MPI on raw data.

Hiding some sort of redistribution into a PC object will simplify a lot the
ugliest part of PCBDDC, and make it readable also by any other person which
is not me ;-).  I don't mind if it will go in PCREDUNDANT or PCLU (or
PCREDISTRIBUTE, me and Jed had a talk on this also some time ago).
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20130108/10b9f69b/attachment.html>