very slow barriers (possibly futex?)

[SteffenSeckler]

Windows Build Number

Microsoft Windows [Version 10.0.21390.1010]

WSL Version

• WSL 2
• WSL 1

Kernel Version

5.10.16

Distro Version

Ubuntu 18.04, Ubuntu 20.04

Other Software

Docker Desktop 3.3.3

g++11 (required!)

Repro Steps

I have been testing c++11's std::barrier and it appeared VERY slow on WSL2.

Steps to reproduce:

1. download and unpack test.zip
2. mkdir build && cd build
3. CC=gcc-11 CXX=g++-11 cmake .. && cmake --build .
4. ./test

Expected Behavior

Timing results are similar to both docker on Linux (using gcc) and msvc on windows.

docker on Linux using gcc-11:

executing with_race:
 bool with_race(): 15889 instead of 20000 at iteration 0
 took: INVALID -- has a race.
executing with_atomic:
 took: 234 ms.
executing with_lock:
 took: 3489 ms.
executing with_omp_barrier:
 took: 11503 ms.
executing with_barrier:
 took: 27354 ms.

msvc (requires V16.10) on windows:

executing with_race:
 with_race: 19463 instead of 20000 at iteration 56
 took: INVALID -- has a race.
executing with_atomic:
 took: 238 ms.
executing with_lock:
 took: 1055 ms.
executing with_omp_barrier:
 took: 4144 ms.
executing with_barrier:
 took: 6422 ms.

Actual Behavior

Execution of the variant with std::barrier takes significantly (40x) longer compared to variant using openmp barriers.

executing with_race:
 bool with_race(): 10742 instead of 20000 at iteration 204
 took: INVALID -- has a race.
executing with_atomic:
 took: 177 ms.
executing with_lock:
 took: 806 ms.
executing with_omp_barrier:
 took: 2457 ms.
executing with_barrier:
 took: 96298 ms.

I am not completely certain how to determine the cause of this slowness, but when pausing the application in debug mode, it mostly resides in syscalls to futexes.

Diagnostic Logs

No response

[therealkenc]

Thank-you for the submission and tight repro.

[microsoft-github-policy-service]

This issue has been automatically closed since it has not had any activity for the past year. If you're still experiencing this issue please re-file this as a new issue or feature request.

Thank you!