Before:
https://github.com/tikkss/test-unit/actions/runs/20645608219/job/59282994734#step:5:27
Finished in 5.3867068 seconds.

After:
https://github.com/test-unit/test-unit/actions/runs/20816091253/job/59792001115?pr=368#step:5:27
Finished in 4.0076588 seconds.

I investigated that elapsed time per processing in Windows CI:
(Note that some processes run in parallel, so the sum of all elapsed times is not equal to the total test execution time)

     collected in main process ┤■■■■■■■■■■■■ 0.683695
             spawned a process ┤■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■ 1.708945
   collected in worker process ┤■■■■■■■■■■■■■■■■■■■■■■■■■■■■ 1.608906
                      finished ┤■■■■■■■■■■ 0.56229

According to the above investigation, the next bottle neck is collecting test suite in worker process (Spawning processes is the slowest, but it seems to be beyond what we can improve).

In the main process, collecting the test suite takes about 0.6 seconds on Windows and 0.1 seconds on Linux.
In the worker process, collecting the test suite takes about 1.6 seconds on Windows and 0.2 seconds on Linux.

I do not know why there is such a large difference in elapsed time for collecting the test suite between the main process and worker processes only on Windows. I would like to investigate this.

By the way, on Windows, collecting the test suite becomes slower as the number of worker processes increases:

• 1 worker: 0.6s
• 2 workers: 0.9s
• 3 workers: 1.2s
• 4 workers: 1.6s

On Linux, it stays constant.