FloydWarshall is very memory inefficient

Further to our sending Dijkstra out to lunch:

FloydWarshall starts with:

    // init the backtrace matrix
    backtrace = new Object[n][n];

You know what’s coming next 😀 Let’s assume that N = 586165 (our “small” graph):

586165 * 586165 * 8 // on a 64-bit machine
2748715257800 // bytes
./1048576
2621379 // megabytes...
./1024
2559 // gigabytes ...

Since our graph is directed and largely acyclic, I think we can do much better. Interestingly, we haven’t even hit hash tables, this is a straight array. So, raising a ticket as a placeholder while we figure this out.

It looks like a lovely candidate for a ForkJoinPool, too.

[jsichi]

(I think the duplicated code noted by @shevek has finally been removed in #630.)

[jsichi]

I’ve merged #632. Remaining items here are:

• fastutil variant
• Alternative all-pairs algorithms for special cases (e.g. integer edges) where sub-cubic bounds are known; these should be addressed in their own issues.
• Cases where the outer loop can terminate early; I do not believe this is applicable in general, but this may be relevant for some of the special cases (like in TransitiveClosure).

Anything else?