Preconditioner that works to 512 cores and falls apart at 1024

Strong scaling is close to ideal up to 512 ranks. At 1024 the iteration count roughly doubles and the wall time goes up. The partitioning is the same algorithm at both sizes and the subdomains are still large enough that I would not expect the coarse problem to dominate.

Before I start instrumenting: is the doubling of the iteration count a known signature of something specific, or is it just what happens when the coarse level stops being coarse enough?