As Jed mentioned you should not see a significant difference with '-random_seed 1'. This fix may just be papering over a problem that will bite you eventually. If you run with '-XXX_gamg_verbose 2' the eigenvalues used will be printed out. Try running with and without '-random_seed 1' and send the output. We are looking to see if the computed eigenvalues on these second (bad) solves is different in the two runs.
And I'm a bit puzzled because I don't know how '-random_seed 1' could be used in GAMG to compute eigen estimates... On Jul 11, 2013, at 4:47 PM, John Mousel <[email protected]> wrote: > I have two KSP contexts, helm%ksp and proj%ksp. I switch between the two by > calling KSPSetOperators. There was no performance difference, just a > different answer. > > > On Thu, Jul 11, 2013 at 3:41 PM, Jed Brown <[email protected]> wrote: > John Mousel <[email protected]> writes: > > > Jed, > > > > I alternate between solving Helmholtz and Poisson equations in an outer > > loop. The matrix is fixed for both operators during the process, and only > > the right-hand side is updated for each system. > > Do you have one KSP or two? If one, I assume you're trying to preserve > memory, but there is likely more memory in the preconditioner than the > Krylov space and if you want to reuse the preconditioner, you have to > set it up anew each time. > > > I just tried -random_seed 1, and it seems to have made all the difference > > in the world! > > That's disconcerting because we would like the algorithm to be > insensitive to such things. How big was the performance difference? > Can you give us some information to reproduce? (Maybe a smallish > example matrix that demonstrates this problem.) >
