On Fri, Dec 3, 2010 at 22:15, Hong Zhang <hzhang at mcs.anl.gov> wrote:
> Yes, old ICC was buggy. Unlike Cholesky which reveals error immediately, > incorrect ICC may still gives convergence. I roughly recall that > comparing convergence of > icc for aij and sbaij, I fixed bug in one of the routines (likely for > sbaij). > What block size of sbaij do you use? > This was for bs=2. Note that I plain AIJ was requiring a shift and SBAIJ(2) had negative pivots (but didn't do anything about them). So they would have had to both be buggy. > While changing data structures in matrix factorization routine, I > cleaned or rewrote > part of factorization routines, mainly for bs=1. bs>1 might need more work, > no > guarantee there :-( > The new one is working, I have identical results with AIJ and SBAIJ(2). Neither of them need a shift. This is good, except that I no longer have a nice explanation for why ICC performs so poorly on this problem when used with 1-level ASM (no coarse level). The difference between direct subdomain solves and incomplete solves, even with small subdomains, is an order of magnitude for ASM, but negligible when there is a coarse level. I know this is not overly weird, but having ICC produce pivots that needed shifting (despite the matrix being SPD) was a tidy explanation that apparently I have to discard in favor of "incomplete factorization is unpredictable". :-( Thanks for running through the history with me. Jed -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20101203/2e6139c7/attachment.html>
