Move broken Parallel.gdistances function to Experimental submodule

[simonschoelly]

The Parallel.gdistances functions seems to have some incorrect assumptions about threaded memory. This leads to tests failing from time to time. In order to not completely throw away the code, this PR moves it to the Experimental.Parallel submodule and deactivates the tests for this function.

See #10

[codecov]

Codecov Report

Merging #94 (e5b8ffb) into master (abdfceb) will decrease coverage by 1.03%.
The diff coverage is 50.00%.

@@            Coverage Diff             @@
##           master      #94      +/-   ##
==========================================
- Coverage   98.82%   97.79%   -1.04%     
==========================================
  Files         109      110       +1     
  Lines        6201     6203       +2     
==========================================
- Hits         6128     6066      -62     
- Misses         73      137      +64     

[etiennedeg]

There is also gdistances tests in
Graphs.jl/test/traversals/bfs.jl (lines 29-35)

gdistances should not be globally exported, but only exported in Experimental module.

We should also reflect this change in the documentation (gdistances appears in
Graphs.jl/docs/src/pathing.md for example)

period is calling gdistances, what should we do about it ?

@ViralBShah mentionned that he also saw bfs failing, but I couldn't find such failing test.

[ViralBShah]

The bfs failures only happen sporadically in CI for me. I could not reproduce on my own machine.

[simonschoelly]

The tests in Graphs.jl/test/traversals/bfs.jl are for the non-parallel version of gdistances which is is not broken, so it still should be exported:

Graphs.jl/src/traversals/bfs.jl

Lines 126 to 141 in abdfceb

	"""
	gdistances(g, source; sort_alg=QuickSort)
	Return a vector filled with the geodesic distances of vertices in `g` from
	`source`. If `source` is a collection of vertices each element should be unique.
	For vertices in disconnected components the default distance is `typemax(T)`.
	An optional sorting algorithm may be specified (see Performance section).
	### Performance
	`gdistances` uses `QuickSort` internally for its default sorting algorithm, since it performs
	the best of the algorithms built into Julia Base. However, passing a `RadixSort` (available via
	[SortingAlgorithms.jl](https://github.com/JuliaCollections/SortingAlgorithms.jl)) will provide
	significant performance improvements on larger graphs.
	"""
	gdistances(g::AbstractGraph{T}, source; sort_alg = Base.Sort.QuickSort) where T = gdistances!(g, source, fill(typemax(T), nv(g)); sort_alg = sort_alg)

[ViralBShah]

IIRC, it was parallel bfs that was broken - not the sequential one. And it happened only on windows. The logs are all gone now.

[simonschoelly]

The testset in which the failing gdistances test where defined, is called

Graphs.jl/test/parallel/traversals/gdistances.jl

Line 1 in abdfceb

@testset "Parallel.BFS" begin

Is it possible, that this is the reason why you think that bfs fails from time to time?

[ViralBShah]

Yes that would explain it. I was just going by what I was seeing in the logs.

[etiennedeg]

Ok, my bad, all my remarks were on sequential gdistances, so I think its ok for me. (gdistances was not even mentionned in the documentation of parallel algorithms)

[simonschoelly]

Thanks for reviewing :) Will merge this now, so hopefully we won't see that many failing tests anymore.