speedup @everywhere

[amitmurthy]

Current implementation executes two calls - a remotecall for expression execution and a remotecall_fetch to wait for execution completion on each worker. Now there is a single call of remotecall_fetch.

@andreasnoack , the linear slowdown in the first @everywhere 1+1 was because all code paths related to remotecall_fetch was being compiled for the first time and it was being executed serially.
Actually remotecall_fetch was being called as part of the addprocs. The extra time may just have been due to the serial execution of remotecall_fetch against all workers.

cc @JeffBezanson

[andreasnoack]

Great. I can confirm that this is much faster. On the Slurm managed cluster I now get

nworkers	second
4	0.630646
8	0.309988
16	0.31393
32	0.325648
64	0.334107
128	0.374203
256	0.38756
512	0.443842
1024	💣

but the last row could easily be Slurm related and would have to be investigated further.

[ViralBShah]

👍

[ViralBShah]

cc @alanedelman

[tkelman]

backported in 426c5be