Add concurrency safe iterator wrapper ChannelLike

[fredrikekre]

Just to show that the idea from JuliaFolds2/ChunkSplitters.jl#49 can be used to improve the (chunked) GreedyScheduler.
The advantages are that the iterator doesn't have to be copied into a Channel, and that the AtomicChunk-thingy is more efficient (but not as general, of course) than a Channel. See some benchmarks in the test.jl file.

[MasonProtter]

Very nice idea, I had wanted something like this before.

[fredrikekre]

Ok, cleaned it up and added some docs. If you prefer to keep it internal we can do that, but it isn't really much code to maintain. On the other hand, perhaps this isn't a package of "generally useful things". Let me know what you think.

Posting the benchmarks here for posterity since I removed the file with them from the commit:

Benchmark:

using OhMyThreads: tmapreduce, GreedyScheduler
using BenchmarkTools: @btime

function mc_parallel(N; ntasks, chunksize)
    scheduler = GreedyScheduler(; ntasks, chunksize, split = :consecutive)
    M = tmapreduce(+, 1:N; scheduler) do i
        rand()^2 + rand()^2 < 1.0
    end
    pi = 4 * M / N
    return pi
end

const N = 100_000_000

for chunksize in (10, 1_000, 100_000), ntasks in (1, 2, 4)
    @btime mc_parallel(N; ntasks = $ntasks, chunksize = $chunksize)
end

PR:

950.227 ms (15 allocations: 944 bytes)
656.752 ms (21 allocations: 1.48 KiB)
562.387 ms (33 allocations: 2.61 KiB)

506.860 ms (15 allocations: 944 bytes)
252.174 ms (21 allocations: 1.48 KiB)
129.447 ms (33 allocations: 2.61 KiB)

499.351 ms (15 allocations: 944 bytes)
249.004 ms (21 allocations: 1.48 KiB)
127.825 ms (33 allocations: 2.61 KiB)

master:

2.575 s (53 allocations: 439.07 MiB)
3.593 s (59 allocations: 219.72 MiB)
3.346 s (70 allocations: 110.09 MiB)

518.885 ms (46 allocations: 4.06 MiB)
263.877 ms (52 allocations: 4.06 MiB)
141.339 ms (60 allocations: 69.31 KiB)

500.587 ms (41 allocations: 67.19 KiB)
252.641 ms (47 allocations: 67.75 KiB)
127.858 ms (59 allocations: 26.23 KiB)

[carstenbauer]

LGTM and I'm fine with documenting this as API. However, I think we need to modify the GreedyScheduler docstring which currently mentions Channel explicitly. Probably we shouldn't mention the specific implementation of the "storage" at all.

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

Attention: Patch coverage is 80.00000% with 2 lines in your changes missing coverage. Please review.

Project coverage is 85.81%. Comparing base (cba9127) to head (1776729).
Report is 127 commits behind head on master.

Files with missing lines	Patch %	Lines
src/types.jl	77.77%	2 Missing ⚠️

❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #121      +/-   ##
==========================================
- Coverage   90.24%   85.81%   -4.44%     
==========================================
  Files           3        7       +4     
  Lines          82      578     +496     
==========================================
+ Hits           74      496     +422     
- Misses          8       82      +74     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[carstenbauer]

BTW, could perhapse make sense to mention this in the docs under https://juliafolds2.github.io/OhMyThreads.jl/stable/literate/tls/tls/#Another-safe-way-based-on-Channel.

[fredrikekre]

Added to changelog and updated the GreedyScheduler docstring.

[carstenbauer]

Unfortunately, we mention "channel" more times in the docstring in the keyword arguments section. Would be great if you could "fix" this.

[fredrikekre]

BTW, could perhapse make sense to mention this in the docs under

Done.

Unfortunately, we mention "channel" more times in the docstring

Fixed, replaced with "shared workqueue".