Suggest experiment with mumps parallel direct LU solver as well. Hong
On Mon, Nov 14, 2011 at 7:34 AM, Matthew Knepley <knepley at gmail.com> wrote: > On Mon, Nov 14, 2011 at 12:02 PM, Bogdan Dita <bogdan at lmn.pub.ro> wrote: >> >> ?Hello, >> >> ?Below is my post from a few days ago and this time I've attached the >> output from log_summary. > > The time increase comes completely from SuperLU_dist during the > factorization > phase. You should use -ksp_view so we can see what solver options are used. > ? ?Matt > >> >> " >> ?Until a few days ago I've only be using PETSc in debug mode and when I >> switch to the optimised version(--with-debugging=0) I got a strange >> result regarding the solve time, what I mean is that it was 10-15 % >> higher then in debug mode. >> ?I'm trying to solve a linear system in parallel with superlu_dist, and >> I've tested my program on a Beowulf cluster, so far only using a single >> node with 2 quad-core Intel processors. >> ?From what I know the "no debug" version should be faster and I know it >> should be faster because on my laptop(dual-core Intel) for the same >> program and even the same matrices the solve time for the optimised >> version is 2 times faster, but when I use the cluster the optimised >> version time is slower then the debug version. >> ?Any thoughts? >> >> " >> ?Best regards, >> ?Bogdan Dita >> >> >> > > > > -- > What most experimenters take for granted before they begin their experiments > is infinitely more interesting than any results to which their experiments > lead. > -- Norbert Wiener >
