Check for block shape/grid_shapes that would make empty localarrays.

[cowlicks]

No description provided.

[cowlicks]

Closes issue #442

[coveralls]

Coverage decreased (-0.17%) when pulling 0f36738 on cowlicks:some-weird-shapes into 4405136 on enthought:master.

[kwmsmith]

Won't this issue show up for multi-dimensional block-distributed arrays for certain sizes? The test here needs to be more general to cover the multi-dimensional cases. If it can be refactored so it handles things on a per-dimension basis, that would be ideal.

For example, say we have a 16 engine cluster running. Currently, we get a lot of empty localarrays with a 6 x 9 block-distributed array:

In [5]: d = Distribution.from_shape(c, (6, 9), ('b', 'b'))

In [6]: d.localshapes()
Out[7]:
[(2, 3),
 (2, 3),
 (2, 3),
 (2, 0),
 (2, 3),
 (2, 3),
 (2, 3),
 (2, 0),
 (2, 3),
 (2, 3),
 (2, 3),
 (2, 0),
 (0, 3),
 (0, 3),
 (0, 3),
 (0, 0)]

[cowlicks]

Yes this needs to be done on a per dimension basis.

[kwmsmith]

@cowlicks what's the latest on this one?

[cowlicks]

My solution here is pretty flawed. I need to incorporate a fix into the "main case" branch in make_grid_shape. So it accounts for all dimensions. I'm still figuring out how the factoring stuff in there works.

[bgrant]

Postponing to 0.7.