Fix distribute function

[amitmurthy]

This bug was causing the entire array to be distributed to be serialized to and from the workers in the init function.

[coveralls]

Coverage decreased (-20.1%) to 50.0% when pulling c89fbce on amitm/expt into ec9ab98 on master.

[codecov-io]

Current coverage is 69.01%

Merging #72 into master will decrease coverage by 1.10%

@@             master        #72   diff @@
==========================================
  Files             1          1          
  Lines           569        568     -1   
  Methods           0          0          
  Messages          0          0          
  Branches          0          0          
==========================================
- Hits            399        392     -7   
- Misses          170        176     +6   
  Partials          0          0          

Powered by Codecov. Last updated by ec9ab98...28707f5

[coveralls]

Coverage decreased (-1.1%) to 69.014% when pulling 28707f5 on amitm/expt into ec9ab98 on master.

[andreasnoack]

Ah. So there is actual a difference in the code between 0.4 and 0.5. I don't really get the difference here between the channel and future here. It would be great if you could elaborate. Thanks.

[amitmurthy]

 owner = myid()
 rr = Future()
 put!(rr, A)
 d = DArray(size(A), procs, dist) do I
       remotecall_fetch(() -> fetch(rr)[I...], owner)
 end

Since serialization of "filled" Futures also includes the data of the Future, the previous code (shown above) was resulting in the entire array A, first being sent as part of the closure (due to rr being captured in the DArray init function) to each worker, then from each worker back to the master (remotecall_fetch on the owner) and finally the requested sub-array was being returned to the worker.

Changing it to RemoteChannel, which always only sends 3 integers ( where, whence and id ) fixed it.

In 0.4, this was a RemoteRef (which behaves like a RemoteChannel) and hence was not an issue. The bug crept in during the 0.5 port.