I would generally use a numpy array (presumably of short ints, 2 bytes each, since you may need more than 256 distinct values) -- that would take less than 4MB for the <2 million cells.

If for some reason I couldn't afford the numpy dependency (e.g on App Engine, which doesn't support numpy), I'd use the standard library array module - it only supports 1-dimensional arrays, but it's just as space-efficient as numpy for large homogeneous arrays, and the getter/setter routines you mention can perfectly well "linearize" a 3-items tuple that's your natural index into the single integer index into the 1-D array.

In general, consider numpy (or array) any time you have large homogeneous, dense vectors or matrices of numbers -- Python built-in lists are highly wasteful of space in this use case (due to their generality which you're not using and don't need here!-), and saving memory indirectly translates to saving time too (better caching, fewer levels of indirection, etc, etc).

I would generally use a numpy array (presumably of short ints, 2 bytes each, since you may need more than 256 distinct values) -- that would take less than 4MB for the <2 million cells.

If for some reason I couldn't afford the numpy dependency (e.g on App Engine, which doesn't support numpy), I'd use the standard library array module - it only supports 1-dimensional arrays, but it's just as space-efficient as numpy for large homogeneous arrays, and the getter/setter routines you mention can perfectly well "linearize" a 3-items tuple that's your natural index into the single integer index into the 1-D array.

In general, consider numpy (or array) any time you have large homogeneous, dense vectors or matrices of numbers -- Python built-in lists are highly wasteful of space in this use case (due to their generality which you're not using and don't need here!-), and saving memory indirectly translates to saving time too (better caching, fewer levels of indirection, etc, etc).