The similar method for SharedArray returns an Array

[dmbates]

Wouldn't it make more sense to have it return a SharedArray of the same element type and dimensions with the index chunks for the pids aligned?

[amitmurthy]

I guess we we want the same behavior for a DArray too, i.e., similar(d::DArray,element_type, dims) returning an uninitialized darray?

Also once #5452 is merged similar(S::SharedArray, element_type, dist::SharedDist) would return an uninitialized SharedArray while similar(D::DArray, element_type, dist::ChunkedDist) would return a DArray. dist describes the distribution of the array (pids and chunks in case of a darray, only pids in case of a sharedarray). Both element_type and dist are optional.

[amitmurthy]

Since similar returns an uninitialized array, and I guess the intent of this is to make the same code play nicely whether it is an Array, DArray or SharedArray, I am wondering what would be the best strategy to achieve this.

Towards this end should we define
procs(a::AbstractArray) = [myid()] ,
localpart(a::AbstractArray) = a and localindexes(a::AbstractArray) which again returns the complete indexes of a?

Similarly for SharedArray too. @timholy had some reservations in describing a chunking scheme for a SharedArray since it was not really relevant, but I am thinking now that it may make sense if folks want to write algorithms that compose the same way with a local or distributed array and just work. SharedArray will have a localpart returning a SubArray, and localindexes, which would be based on a partitioning scheme that plays well with SharedArrays.

Users can then write code assuming it to be distributed and it will work the same way for a local array:

for p in procs(a)
   remotecall_fetch(p, ....)
end

Thoughts?

[timholy]

My point was that some algorithms may just ignore it, but I don't think it hurts anything to have it. So if this makes life better, no objections from me.