[petsc-dev] [petsc-users] Multiple solves with PCMG fail

[Jed Brown]

Mark Adams <mfadams at lbl.gov> writes:
> OK, I know resistance is futile but I'll just go on the record as saying
> this is not a good idea.  GAMG uses a random vector and gives this to
> Cheby.  The EPSI project that I work on has a zero RHS in their first
> solve.  What if you do the setup before the solve and before you have a RHS.

Cheby doesn't determine the bounds in KSPSetUp_Chebyshev.

My real concern with randomness is that results are not reproducible.
If convergence is sensitive to the RHS used for the estimate, the last
thing I want is to be unable to reproduce that failure.  One could
simply hash the index, but then answers would be different if the
variables are permuted.

Using the RHS also needs one less vector (not a memory concern for GAMG,
but the cost is not trivial for HPGMG, for example).

> Also, the random vector is high frequency and should converge faster.  The
> RHS can often be low frequency.

Good point, and it may get rougher after a few iterations.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 818 bytes
Desc: not available
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20141022/fd4e2072/attachment.sig>