Windows PowerShell 5.1 tests are much slower on Server 2025 than on Server 2022

[nohwnd]

The PS_5_1_Windows_Server2025 leg takes longer than PS_5_1_Windows_Server2022 on every build I looked at. In build 3940 the unit test phase took 80.34s on 2022 and 135.37s on 2025, with the same 2913 tests.

It is not the image being slower. PS7 runs on the same two images in the same build:

	Server 2022	Server 2025
PS7	53.35s	55.41s
PS 5.1	80.34s	135.37s

PS7 is 4% slower on 2025, Windows PowerShell 5.1 is 69% slower. A slower VM, a disk difference or Defender scanning the working folder would slow down both, so this looks like Windows PowerShell 5.1 or the .NET Framework on that image, and not the OS.

Almost every test file is slower, three carry most of it:

file	2022	2025
Help.Tests.ps1	14.95s	28.96s	1.9x
Format2.Tests.ps1	5.07s	16.81s	3.3x
Mock.Tests.ps1	13.21s	21.00s	1.6x

The gap is noisy. Over 8 recent builds 2025 was slower every single time, so the direction is real, but the size went from +0:22 to +6:48. Measure over several builds, one sample is not enough to tell whether a change helped.

This matters now because after #2984 the 5.1 legs are the slowest in the matrix and they set the wall time of a PR build, 12:07 out of 13:13.

Builds the numbers come from: 3940 (#2984), 3939, 3936, 3914, 3913, 3910, 3908, 3906.

🤖

[nohwnd]

I pulled the same logs from 7 builds instead of 1, and the picture changes. Correcting myself.

Unit test phase, seconds, sorted:

image	runs	median
Server 2022	75, 78, 79, 80, 80, 81, 86	79.5
Server 2025	82, 85, 90, 106, 106, 126, 135	105.9

Server 2022 is stable, an 11s spread over 7 builds. Server 2025 goes from 82 to 135 on the same code, and 3 of the 7 runs are inside the 2022 range. So the big gap is not there every time, it comes and goes, and that says runner variance rather than something the code or the PowerShell version does differently.

The per-file numbers I posted above came from build 3940, which is the slowest 2025 run in the set. Over 7 builds:

file	2022	2025	median ratio
Mock.Tests.ps1	12.7 .. 13.8	14.8 .. 28.3	1.52x
Help.Tests.ps1	13.8 .. 16.7	14.6 .. 29.0	1.18x
Format2.Tests.ps1	5.1 .. 6.8	4.9 .. 16.8	1.02x

Format2.Tests.ps1 is not slower on 2025, it was 16.8s in the one build I looked at and 4.9s in another. Drop it, I read one sample as a pattern.

Mock.Tests.ps1 is the one that holds up. It is slower on 2025 in all 7 builds, the slowest 2022 run is still faster than the fastest 2025 run. But in the good runs it is only about 2s of the difference, so it does not explain the 50s gap either.

What is left is: the Server 2025 runners are sometimes much slower for Windows PowerShell 5.1 and consistently fine for PS7 on the same image in the same build. I still do not have a cause. Anyone picking this up should collect over many builds, a single build tells you nothing here, as shown above.

🤖