Yes, many problems can be expressed as dense linear algebra, and so measuring and comparing LINPACK perf for these makes sense for those problems. However, many problems don't map well to dense linear algebra.
Sure, but as far as I've seen, linear algebra problems dominate the runtime of these very large systems. That's what I use them for.
At least the first 6 on that dwarfs list are done daily on top500 machines. I write parallel spectral methods, and use structured and unstructured grids. Achieving high scaling on these on massively parallel machines is not at all what I would call an open problem (as far as correctly using a given network for a problem, or designing a network for a given problem). For any given network topology, breaking these problems into parallel chunks that are best for the particular topology is just a graph partitioning problem with a trivially sized graph. That little white paper strikes me as being somewhere between "obvious to anyone that's done high performance computing in the last 40 years" , "putting a cute name on old hat", and just plain silly.