ArrayDist storing distribution info for darrays. Deprecated d* functions...

[amitmurthy]

This PR does the following:

• Added ArrayDist which is an abstract type. Subtypes are SharedDist and ChunkedDist. Both store the array dimensions and list of processes participating in the distribution.
• ChunkedDist also stores the chunked indexes for the DArray
• d* functions have been deprecated in favor of the regular convenience constructors accepting a ChunkedDist

Will update documentation if the above is OK.

Updates to SharedArray using a SharedDist will be made in a separate PR.

[amitmurthy]

cc: @JeffBezanson - please have a look.

[JeffBezanson]

Not necessary, since default constructors now look like this. The default outer constructor will also infer N given dims.

[JeffBezanson]

Good. Just a couple small comments. Will definitely merge this after doc updates and rebase.

[amitmurthy]

Updated docs and rebased.
Also closes #5481 .

[amitmurthy]

Bump @JeffBezanson

[JeffBezanson]

This seems to be inconsistent with getindex below; i.e. the size does not give the valid range of indexes for getindex. That could be confusing.

[amitmurthy]

You are right. How about size(ChunkedDist) returning the size of the indexes, while size(SharedDist) returns the length of the pids vector. Both describe the size of the distribution.

size(DArray) and size(SharedArray) will return the size of the complete array. There will not be a function to retrieve dims from an ArrayDist - folks have to use ad.dims

[amitmurthy]

Just pushed an update.

[JeffBezanson]

@StefanKarpinski @ViralBShah
I think this is definitely the way forward, since we can't keep adding letter prefixes to zeros (dzeros, etc.), but the syntax zeros(ChunkedDist(dims)) is a bit verbose. My first thought is that it won't pass the Viral and/or Alan tests. Maybe wrap ChunkedDist in a function dist for example? Any other name ideas?

[StefanKarpinski]

Yeah, this has always bothered me. It's a clear sign of lack of composability every time you have parallel universes of functions that differ by prefix or suffix. How about something like this zeros(Dist,dims) and zeros(Shared,dims).

[JeffBezanson]

Come to think of it, the name SharedDist is kind of an oxymoron, since "shared" and "distributed" are opposites in parallel programming terms. Renaming them Shared and Dist would make sense.

[amitmurthy]

@StefanKarpinski , Dist represents 3 things

• the dimensions of the array,
• the pids participating in the distribution and
• the distribution of the array over pids

So it can be something like zeros(Dist(dims; pids, dist)) where optional keyword arguments allow the user to specify the pids and dist, i.e., number of parts in each dimension. Default values for the same are workers() and a split along the larger dimension first.

For SharedArrays it can be zeros(Shared(dims; pids))

[amitmurthy]

Closed due to #10334