# Distributed nested in Pmap
**URL:** https://discourse.julialang.org/t/distributed-nested-in-pmap/21081
**Category:** Julia at Scale
**Created:** [February 22, 2019, 2:25pm UTC](https://discourse.julialang.org/t/distributed-nested-in-pmap/21081 "2019-02-22T14:25:25Z")
**Posts on this page:** 3
**Page:** 1
### Author:  [@Alessio\_Bellisomi](https://discourse.julialang.org/u/Alessio_Bellisomi)
#### Post date: [February 22, 2019, 2:25pm UTC](https://discourse.julialang.org/t/distributed-nested-in-pmap/21081/1 "2019-02-22T14:25:25Z")
Hello,
I am fairly new to Julia, I am trying to understand a little better how the workload is distributed, so I prepared this little script to help me out.
An external pmap splits the work between different processes, each one internally will distribute some calculation over the array and aggregate.
```julia
using Distributed
addprocs(4)
list = [rand(1, 3), rand(1, 3), rand(1, 3), rand(1, 3), rand(1, 3)]
@everywhere function aggregate_data(i)
i*i
end
@everywhere function loop_and_sum(data::Array)
@distributed (+) for val in data
aggregate_data(val)
end
end
result = pmap(x->loop_and_sum(x), list)
rmprocs(4)
```
In the documentation for the [@distributed](https://docs.julialang.org/en/v1/stdlib/Distributed/index.html#Distributed.@distributed) macro I read that it spawns independent tasks over all available workers. This is what I need to clarify: who are these workers? Aren’t those the processes that I created at the beginning of the scripts, which are already busy _pmapping_ stuff? Is this efficient?
Also, as you can see above I call the _rmprocs_ to cleanup the workers when I am done. What I see in the Task Manager though, is that those process are never killed until I don’t kill the main process (~restart the Julia REPL). How is that meant to work?
Thanks a lot
---
### Author:  [@dmolina](https://discourse.julialang.org/u/dmolina)
#### Post date: [February 22, 2019, 5:11pm UTC](https://discourse.julialang.org/t/distributed-nested-in-pmap/21081/2 "2019-02-22T17:11:41Z")
I am not an expert, but:
- Worker are different processes that run your code. You can indicate how many do you want to use or with “julia -p ” or with addprocs(). You should have cpus in that computer to get the right performance. Because all process need store its own data, sometimes it could be better to use thread, because they share memory (in conjuntion with SharedArrays), but sometimes you can get strange results if you do not are careful.
The use of everywehere is right, however, with “pmap” you are already using the processing in the different workers, so you should not use distributed, I think.
About rmprocs, you are not removing all processes, you are only removing the 4th process, you can remove all with:
[rmprocs(n) for n in 1:4]
or in a simpler way, storing ids obtained by addprocs:
np = addprocs(4)
…
rmprocs(np)
---
### Author:  [@Alessio\_Bellisomi](https://discourse.julialang.org/u/Alessio_Bellisomi)
#### Post date: [February 22, 2019, 5:54pm UTC](https://discourse.julialang.org/t/distributed-nested-in-pmap/21081/3 "2019-02-22T17:54:59Z")
> About rmprocs, you are not removing all processes, you are only removing the 4th process
Ah, that makes total sense 🙂
Thanks for your sharing your thoughts. I am working on a script which is nesting pmaps and distributed and other quirks, you see why I am getting confused!