build results vary from parallelism

[bmwiedemann]

Bug report

Bug summary

After fixing #4609, there is some other remaining issue and my tools said, it is about the number of cores I give the build-VM.

Code for reproduction

build once each in a 1-core-VM and a 2-core-VM

osc co openSUSE:Factory/python-yt && cd $_
for N in 1 2 ; do
    osc build --noservice --vm-type=kvm -j$N --keep-pkg=RPMS.$N
    md5sum `find /var/tmp/build-root/standard-x86_64/home/abuild/rpmbuild/ -name \*.c|sort` > $N.md5
done
diff -u ?.md5

Actual outcome

bounding_volume_hierarchy.cpython-310-x86_64-linux-gnu.so and other binaries vary

/home/abuild/rpmbuild/BUILD/yt-4.1.4/build/lib.linux-x86_64-cpython-311/yt/utilities/lib/bounding_volume_hierarchy.cpp        2039-08-30 07:17:28.000000000 +0000
@@ -7059,7 +7059,7 @@            
                                 goto __pyx_L18;
                                 __pyx_L18:;
                                 #ifdef _OPENMP
-                                #pragma omp critical(__pyx_parallel_lastprivates1)                              
+                                #pragma omp critical(__pyx_parallel_lastprivates0)                                  
                                 #endif /* _OPENMP */
                                 {   
                                     __pyx_parallel_temp0 = __pyx_v_i;

Expected outcome

It should be possible to create bit-identical results (currently, that works only by doing all builds in 1-core-VMs)

Version Information

• Operating System: openSUSE-Tumbleweed 20230724
• Python Version: 3.9 - 3.11
• yt version: 4.1.4
• Other Libraries (if applicable):

This bug was found while working on reproducible builds for openSUSE.

[welcome[bot]]

Hi, and welcome to yt! Thanks for opening your first issue. We have an issue template that helps us to gather relevant information to help diagnosing and fixing the issue.

[neutrinoceros]

I half-remember similar problems being raised in the past by downstream packagers (sometimes parallel builds may even crash). Any leads on what we should be on the look out for ?

[bmwiedemann]

From the context, it could come somewhere out of cython - it would not be the first compiler that has trouble with parallel processing.

But I know too little about your setup. python-scikit-learn seems to have a similar problem.

[neutrinoceros]

related issue: #4278