Open jwpeterson opened 9 years ago
Alternatively, we can keep the flags out but disable the example (reduced_basis_ex5 may also need disabling) when libmesh is configured with --enable-blocked-storage
So I gather that the fix we came up with previously was to disable blocked storage by default? It'd be nice to understand why that causes new nonzeros in the matrix. I don't think this issue is specific to the RB code, other than that code uses "extra" matrices.
In any case, disabling those examples when libmesh is configured with --enable-blocked-storage sounds fine to me.
In any case, disabling those examples when libmesh is configured with --enable-blocked-storage sounds fine to me.
:+1: I'll make a commit that does this momentarily.
Example disabled in 2682ece.
Let's look at this again soon. Making blocked storage default to off and marking it as incompatible with some RB examples didn't actually fix the problem, just worked around it.
And we have users on the mailing list interested again, and honestly the memory savings would probably be non-trivial for a lot of other users who don't realize the blocked matrix option exists or who know about it but don't want to trust it until this issue is worked out.
See previous discussion here.
It seems like we used to pass
-mat_new_nonzero_allocation_err false
to these examples specifically to work around this issue, but those flags got dropped in b4d47091f992bf2d by @dknez. It sounds like we may still need them...