Microsoft.Extensions.Diagnostics.ResourceMonitoring with version >= 8.5.0 has bug on Utilization calculation

[Weikai1997]

Description

We found that the CPU performance was very low when using ResourceMonitoring (we used IUtilizationPublisher to publish cpu metrics)
After I went through the source code of .net extension, I found the root.

Let's split the version into two parts, 8.5.0 ~ 8.7.0 and 8.8.0
For 8.5.0~8.7.0, their problem is that they divide the cpu request twice, in the LinuxUtilizationProvider, we calculate the _scaleForTrackerApi = hostCpus / availableCpus; then in the GetSnapshot we multiply it with the cgroupTime. Then we divided it by GuaranteedCpu again in Calculator. Please note that the value of GuaranteedCpu here is 1 in R9 and before .net 8.4.0, but after .net 8.5.0 it is container cpu request, so the dividing of one more cpu request causes the calculated cpu value to become smaller and incorrect.

Then the newest version 8.8.0
It almost fixed the problem, but caused new problems. In the 8.8.0, the LinuxUtilizationProvider changed, _scaleRelativeToCpuRequestForTrackerApi = hostCpus;It removes the availableCpu that is divided when calculating scale. If there are no other changes, now only one cpu request is divided and the value should be correct
But the GuaranteedCpu originally divided in Calculator has been moved to the ResourceUtilization constructor. Now the range of utilization in Calculator is 0 ~ (100 * request). But we performed a Min operation on it var cpuUtilization = Math.Min(Hundred, utilization). This ultimately results in the range of cpu in ResouceUtilization become 0 ~ (100 / request)

So from 8.5.0 to 8.8.0 the cpu value of ResourceUtilization are incorrect and need to be fixed.

Reproduction Steps

1. Use Resource Monitoring with version 8.5.0 ~ 8.8.0, and the IUtilizationPublisher to log the CpuUsedPercentage
2. Let your container use the full cpu
3. Set your linux container request and limit larger than 1 (don't use 1)
4. Run it in a k8s environment.
5. You can see the logged value will not exceed 100/cpurequest

Expected behavior

the cpu value should be 0~100

Actual behavior

the value is range in 0 ~ 100/request

Regression?

the version 8.5.0 ~ 8.8.0 has the issue

Known Workarounds

No response

Configuration

No response

Other information

No response

[RussKie]

Is this something could be replicated locally in a docker container?

[Weikai1997]

@RussKie local WSL container run with docker run --cpus maybe not replicatable, I tried but cpu.shares value is still 1 which not calculated the cpu request correctly

[Weikai1997]

@RussKie PR created for fix the issue in 8.8.0 #5388