On Fri, 20 Nov 2009, Roy Stogner wrote:
> On Fri, 20 Nov 2009, Vijay S. Mahadevan wrote:
>> I'm still trying to understand what is broken in the xda IO. And is
>> this the same format as .xdr ? Is the mesh not written out properly or
>> is it not converging to the right solution ? I do not have two version
>> of libmesh enabled and disabled with Hilbert but am curious to know
>> what is failing in ex10 though.
> There's either a bug in reading or in writing - it appears that the
> solution on an AMR mesh is getting "scrambled" when read out and back
> With the --disable-libHilbert bug, this happens after just one or two
> refinement steps.
> With the --enable-libHilbert bug, things seem to be okay until after
> 10 or 11 steps.
A quick update: the --disable-libHilbert bug may now be fixed in the
SVN head. At least, I finally understood the primary cause of the
problem, applied a trivial fix, and can no longer trigger read()
corruption with any of the meshes and different parallel partitionings
I've tried. I'd appreciate testing from others as well.
One warning: mesh/solution pairs written with --disable-libHilbert
won't currently be read correctly with --enable-libHilbert. And
whereas the failure cases where writing with libHilbert will produce a
bad file are hard to reproduce, reading a non-libHilbert solution with
--enable-libHilbert is currently expected to give a corrupt result
most of the time. We're still working on both libHilbert integration