disable_polyester_threads #86

Krastanov · 2022-07-20T16:05:16Z

As mentioned on discourse: https://discourse.julialang.org/t/nesting-threads-thread-and-polyester-batch-or-context-manager-to-limit-polyester-threads/84490/4

Benchmarks with and without it below:

using Base.Threads
using BenchmarkTools
using Revise
using Polyester

\##

function inner(x,y,j)
    for i ∈ axes(x,1)
        y[i,j] = sin(x[i,j])
    end
end

function inner_polyester(x,y,j)
    @batch for i ∈ axes(x,1)
        y[i,j] = sin(x[i,j])
    end
end

function inner_thread(x,y,j)
    @threads for i ∈ axes(x,1)
        y[i,j] = sin(x[i,j])
    end
end

function sequential_sequential(x,y)
    for j ∈ axes(x,2)
        inner(x,y,j)
    end
end

function sequential_polyester(x,y)
    for j ∈ axes(x,2)
        inner_polyester(x,y,j)
    end
end

function sequential_thread(x,y)
    for j ∈ axes(x,2)
        inner_thread(x,y,j)
    end
end

function threads_of_polyester(x,y)
    @threads for j ∈ axes(x,2)
        inner_polyester(x,y,j)
    end
end

function threads_of_polyester_inner_disable(x,y)
    @threads for j ∈ axes(x,2)
        Polyester.disable_polyester_threads() do
            inner_polyester(x,y,j)
        end
    end
end

function threads_of_thread(x,y)
    @threads for j ∈ axes(x,2)
        inner_thread(x,y,j)
    end
end

function threads_of_thread(x,y)
    @threads for j ∈ axes(x,2)
        inner_thread(x,y,j)
    end
end

function threads_of_sequential(x,y)
    @threads for j ∈ axes(x,2)
        inner(x,y,j)
    end
end

 # Big inner problem, repeated only a few times

y = rand(10000000,4);
x = rand(size(y)...);

@btime inner($x,$y,1) # 73.319 ms (0 allocations: 0 bytes)
@btime inner_polyester($x,$y,1) # 8.936 ms (0 allocations: 0 bytes)
@btime inner_thread($x,$y,1) # 11.206 ms (49 allocations: 4.56 KiB)

@btime sequential_sequential($x,$y) # 274.926 ms (0 allocations: 0 bytes)
@btime sequential_polyester($x,$y) # 36.963 ms (0 allocations: 0 bytes)
@btime sequential_thread($x,$y) # 49.373 ms (196 allocations: 18.25 KiB)

@btime threads_of_polyester($x,$y) # 78.828 ms (58 allocations: 4.84 KiB)
@btime threads_of_polyester_inner_disable($x,$y) # 70.182 ms (47 allocations: 4.50 KiB)
@btime Polyester.disable_polyester_threads() do; threads_of_polyester($x,$y) end; # 71.141 ms (47 allocations: 4.50 KiB)
@btime threads_of_sequential($x,$y) # 70.857 ms (46 allocations: 4.47 KiB)
@btime threads_of_thread($x,$y) # 45.116 ms (219 allocations: 22.00 KiB)

 # Small inner problem, repated many times

y = rand(1000,1000);
x = rand(size(y)...);

@btime inner($x,$y,1) # 7.028 μs (0 allocations: 0 bytes)
@btime inner_polyester($x,$y,1) # 1.917 μs (0 allocations: 0 bytes)
@btime inner_thread($x,$y,1) # 7.544 μs (45 allocations: 4.44 KiB)

@btime sequential_sequential($x,$y) # 6.790 ms (0 allocations: 0 bytes)
@btime sequential_polyester($x,$y) # 2.070 ms (0 allocations: 0 bytes)
@btime sequential_thread($x,$y) # 9.296 ms (49002 allocations: 4.46 MiB)

@btime threads_of_polyester($x,$y) # 2.090 ms (42 allocations: 4.34 KiB)
@btime threads_of_polyester_inner_disable($x,$y) # 1.065 ms (42 allocations: 4.34 KiB)
@btime Polyester.disable_polyester_threads() do; threads_of_polyester($x,$y) end; # 997.918 μs (49 allocations: 4.56 KiB)
@btime threads_of_sequential($x,$y) # 1.057 ms (48 allocations: 4.53 KiB)
@btime threads_of_thread($x,$y) # 4.105 ms (42059 allocations: 4.25 MiB)

@Batch

Benchmarks with and without it below: ``` using Base.Threads using BenchmarkTools using Revise using Polyester \## function inner(x,y,j) for i ∈ axes(x,1) y[i,j] = sin(x[i,j]) end end function inner_polyester(x,y,j) @Batch for i ∈ axes(x,1) y[i,j] = sin(x[i,j]) end end function inner_thread(x,y,j) @threads for i ∈ axes(x,1) y[i,j] = sin(x[i,j]) end end function sequential_sequential(x,y) for j ∈ axes(x,2) inner(x,y,j) end end function sequential_polyester(x,y) for j ∈ axes(x,2) inner_polyester(x,y,j) end end function sequential_thread(x,y) for j ∈ axes(x,2) inner_thread(x,y,j) end end function threads_of_polyester(x,y) @threads for j ∈ axes(x,2) inner_polyester(x,y,j) end end function threads_of_polyester_inner_disable(x,y) @threads for j ∈ axes(x,2) Polyester.disable_polyester_threads() do inner_polyester(x,y,j) end end end function threads_of_thread(x,y) @threads for j ∈ axes(x,2) inner_thread(x,y,j) end end function threads_of_thread(x,y) @threads for j ∈ axes(x,2) inner_thread(x,y,j) end end function threads_of_sequential(x,y) @threads for j ∈ axes(x,2) inner(x,y,j) end end # Big inner problem, repeated only a few times y = rand(10000000,4); x = rand(size(y)...); @Btime inner($x,$y,1) # 73.319 ms (0 allocations: 0 bytes) @Btime inner_polyester($x,$y,1) # 8.936 ms (0 allocations: 0 bytes) @Btime inner_thread($x,$y,1) # 11.206 ms (49 allocations: 4.56 KiB) @Btime sequential_sequential($x,$y) # 274.926 ms (0 allocations: 0 bytes) @Btime sequential_polyester($x,$y) # 36.963 ms (0 allocations: 0 bytes) @Btime sequential_thread($x,$y) # 49.373 ms (196 allocations: 18.25 KiB) @Btime threads_of_polyester($x,$y) # 78.828 ms (58 allocations: 4.84 KiB) @Btime threads_of_polyester_inner_disable($x,$y) # 70.182 ms (47 allocations: 4.50 KiB) @Btime Polyester.disable_polyester_threads() do; threads_of_polyester($x,$y) end; # 71.141 ms (47 allocations: 4.50 KiB) @Btime threads_of_sequential($x,$y) # 70.857 ms (46 allocations: 4.47 KiB) @Btime threads_of_thread($x,$y) # 45.116 ms (219 allocations: 22.00 KiB) # Small inner problem, repated many times y = rand(1000,1000); x = rand(size(y)...); @Btime inner($x,$y,1) # 7.028 μs (0 allocations: 0 bytes) @Btime inner_polyester($x,$y,1) # 1.917 μs (0 allocations: 0 bytes) @Btime inner_thread($x,$y,1) # 7.544 μs (45 allocations: 4.44 KiB) @Btime sequential_sequential($x,$y) # 6.790 ms (0 allocations: 0 bytes) @Btime sequential_polyester($x,$y) # 2.070 ms (0 allocations: 0 bytes) @Btime sequential_thread($x,$y) # 9.296 ms (49002 allocations: 4.46 MiB) @Btime threads_of_polyester($x,$y) # 2.090 ms (42 allocations: 4.34 KiB) @Btime threads_of_polyester_inner_disable($x,$y) # 1.065 ms (42 allocations: 4.34 KiB) @Btime Polyester.disable_polyester_threads() do; threads_of_polyester($x,$y) end; # 997.918 μs (49 allocations: 4.56 KiB) @Btime threads_of_sequential($x,$y) # 1.057 ms (48 allocations: 4.53 KiB) @Btime threads_of_thread($x,$y) # 4.105 ms (42059 allocations: 4.25 MiB) ```

codecov · 2022-07-20T16:08:24Z

Codecov Report

Merging #86 (36aacb8) into master (259c20f) will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master      #86   +/-   ##
=======================================
  Coverage   88.46%   88.46%           
=======================================
  Files           2        2           
  Lines         416      416           
=======================================
  Hits          368      368           
  Misses         48       48

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 259c20f...36aacb8. Read the comment docs.

chriselrod · 2022-07-20T18:16:37Z

src/utility.jl

+ t, r = request_threads(num_threads())
+ f()
+ foreach(free_threads!, r)
+end


Perhaps this should be

t, r = request_threads(num_threads()) try f() finally foreach(free_threads!, r) end

Also, maybe this should be in PolyesterWeave itself?
It could still be re-exported here.

Should also document that it turns off threading for LoopVectorization.@tturbo and Octavian.matmul, even though neither depend on Polyester (but they do depend on PolyesterWeave).

Do benchmark my suggestion though.
If it is slower, feel free to ignore it.

The advantage of my suggestion is that, if your code throws an error, it'll still free the threads. Otherwise, they'll be turned off until you manually call PolyesterWeave.reset_workers!().
Polyester.@batch doesn't use try/finally either, meaning it requires the manual calls after you get errors.

Will do:

documenting disabling threads in tturbo and matmul

checking whether a try block is much slower

move to PolyesterWeave after the above is done

After moving to PolyesterWeave, you can of course import and reexport in this PR.

The benchmarks did not show any slowdown with a try block (especially and obviously if the disable command is executed once in the outermost scope).

chriselrod · 2022-07-20T18:21:01Z

README.md

+ Polyester.disable_polyester_threads() do
+ inner_polyester(x,y,j)
+ end
+ end


Suggested change

@threads for j ∈ axes(x,2)

Polyester.disable_polyester_threads() do

inner_polyester(x,y,j)

end

end

Polyester.disable_polyester_threads() do

@threads for j ∈ axes(x,2)

inner_polyester(x,y,j)

end

end

May as well hoist the disabling and re-enabling out of the loop, to avoid doing atomic operations to the same address in parallel.

Although that is effectively what you're benchmarking with Polyester.disable_polyester_threads() do; threads_of_polyester($x,$y) end;, my concern is that people are just going to copy/paste the examples, which means they'll likely use this slightly worse form, because it is there.

Noted! The two versions (inside and outside the loop) were given exactly so that this can be compared. I will add a note to the readme to warn people about using it outside the loop.

chriselrod · 2022-07-20T18:23:49Z

README.md

+@btime Polyester.disable_polyester_threads() do; threads_of_polyester($x,$y) end; # 71.141 ms (47 allocations: 4.50 KiB)
+@btime threads_of_sequential($x,$y) # 70.857 ms (46 allocations: 4.47 KiB)
+@btime threads_of_thread($x,$y) # 45.116 ms (219 allocations: 22.00 KiB)


I'm guessing this is fastest because you have more than 4 threads?

Yes. I thought it is a reasonable edge case to show.

chriselrod · 2022-07-20T18:26:15Z

README.md

+function inner_thread(x,y,j)
+ @threads for i ∈ axes(x,1)
+ y[i,j] = sin(x[i,j])


Obviously just an example, but @tturbo should win these benchmarks (and would also be single threaded from this!). ;)

Indeed! I will mention it as a caveat in the benchmarks readme.

Krastanov · 2022-07-20T19:18:36Z

The tests will probably fail due to the dependence of the new PolyesterWeave. See here JuliaSIMD/PolyesterWeave.jl#4

chriselrod · 2022-07-20T20:59:51Z

Great, thanks!

chriselrod reviewed Jul 20, 2022

View reviewed changes

Krastanov added 2 commits July 20, 2022 15:00

use try block in disable_polyester_threads

fe7ef9e

move disable_polyester_threads to PolyesterWeave

6535fe6

Krastanov mentioned this pull request Jul 20, 2022

Implement disable_polyester_threads JuliaSIMD/PolyesterWeave.jl#4

Merged

Mention @tturbo in the examples.

36aacb8

Krastanov force-pushed the master branch from ea45737 to 36aacb8 Compare July 20, 2022 20:17

chriselrod merged commit 2926208 into JuliaSIMD:master Jul 20, 2022

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

disable_polyester_threads #86

disable_polyester_threads #86

Krastanov commented Jul 20, 2022

codecov bot commented Jul 20, 2022 •

edited

Loading

chriselrod Jul 20, 2022

Krastanov Jul 20, 2022

chriselrod Jul 20, 2022

Krastanov Jul 20, 2022

chriselrod Jul 20, 2022

chriselrod Jul 20, 2022

Krastanov Jul 20, 2022

chriselrod Jul 20, 2022

Krastanov Jul 20, 2022

chriselrod Jul 20, 2022

Krastanov Jul 20, 2022

Krastanov commented Jul 20, 2022

chriselrod commented Jul 20, 2022

disable_polyester_threads #86

disable_polyester_threads #86

Conversation

Krastanov commented Jul 20, 2022

codecov bot commented Jul 20, 2022 • edited Loading

Codecov Report

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Krastanov commented Jul 20, 2022

chriselrod commented Jul 20, 2022

codecov bot commented Jul 20, 2022 •

edited

Loading