Interpreting `JET` warnings on a simple `Polyester` multithreaded loop

General Usage Performance
, , ,
1

Polyester seems to be one of the bigger sources of TTFX in some of my libraries. It is similarly the main source of warnings when using JET. Below is a very simple multi-threaded loop that causes a large number of JET warnings (both in 1.7.0 and in nightly). I do not have much understanding of ThreadingUtilities and of StaticArrays, where the warnings originate. Would someone be able to guide me where to look for improvements or how I can change my function, in order to silence the JET warnings and make sure there is no risk of runtime dispatch?

julia> using JET, Polyester

julia> function a(m::Matrix, i::Int)
           @batch for r in axes(m,1)
               m[r,:] .+= i
           end
       end
a (generic function with 1 method)

julia> m = rand(Int,5,5)

julia> @report_opt a(m,1)
ā•ā•ā•ā•ā• 92 possible errors found ā•ā•ā•ā•ā•
...

With these versions

  [c3a54625] JET v0.5.16
  [f517fe37] Polyester v0.6.13
  [aedffcd0] Static v0.7.5
  [90137ffa] StaticArrays v1.5.0
  [8290d209] ThreadingUtilities v0.5.0
1 Like
2

Are you using the array slices as indices? Or do you mean to use axes(m,1) instead of eachrow(m)?

1 Like
3

Is this with Static v0.7?

4

Thanks for the correction, got confused as I was cutting down the size of the example. It is corrected in the original post now.

5

It is indeed Static v0.7, but I used the wrong MWE, so that might have been a bit mislead. Please excuse the confusion. The warnings are from StaticArrays (and ThreadingUtilities), not from Static. All versions are added to the original post.

1 Like
6

These are basically all for printing.
Iā€™m apparently on different versions (as I get 60, not 72, possible errors found).
But everything I see comes from these two lines:

This code is for diagnostics, that is, it displays and dumps a failed task so you can see the stacktrace when your threaded code runs an error.

If your programā€™s performance is bottlenecked by how quickly it can print error messages, Iā€™d suggest trying to stop it from throwing so many errors. Maybe fix the bugs that keep showing up in all those stacktraces?
Or PR base Julia to speed up display(::Task) and dump(::Task), I guess.

Of course, you probably didnā€™t hit these lines in your code.
You just got a report from JET.

Is there some way I can hide this code from JET, so we donā€™t have so many false positives/its output is more useful?

I donā€™t think dump(::Task) is all that useful, so I can probably remove that line.

1 Like
7

Iā€™ve gotten reports about Polyester being bad for latency before. If you have some examples, I may be able to look at it sometime.

1 Like
8

Indeed, I am not hitting these lines, they are just showing up in JETā€™s abstract interpretation. To great extent, I am just cargo-culting my approach to lowering TTFX: I have noticed that the fewer Aqua and JET warnings I have, the faster the package imports (while I did not have as much luck with SnoopCompile), so I am spending some time trying to figure out the source of every JET warning. Admittedly, this is not the best ā€œfirst principlesā€ approach, but it is educational enough, and I needed to start somewhere.

Thanks for your explanation! It does not seem there is anything more to see here. If I figure out how to tell the compiler and/or JET to not worry about these lines, I will send you a pull request.