Code review comment for lp:~jobh/dolfin/symmetric-assemble

>
> > unsymmetric component. This is less severe than for SystemAssemble, which
> > will be symmetric but wrong.
> >
>
> It's just as bad. It will produce the wrong answer in common cases,
> e.g. Cholesky factorisation.

Right you are, thanks! I'll put a warning in (like in SystemAssembler).

> > Now, topological search should be easy enough to fix, but it will make it
> > heavier (parallel comms + a search through the dofs). Maybe it's not
> worth
> > it, because the cost of pointwise search is (usually?) low compared to
> > other overhead.
> >
>
> Yes, the solution is to look at the DirichletBC implementation.
>

Not sure what you mean, "look at" as in "study" or as in "fix"? I'm leaning
towards thinking that the actual boundary search is not so important
compared to other overhead, so it's completely acceptable to just say "use
pointwise" in parallel.

-j.