Benchmarks Of Clear Linux's Native Kernel Against Current/Mainline/Preempt-RT Kernels

[phoronix]

Phoronix: Benchmarks Of Clear Linux's Native Kernel Against Current/Mainline/Preempt-RT Kernels

Given the recent discussions over the default performance of the Linux scheduler, the Liquorix patches to the Linux kernel, and other recent forum discussions over different kernel configurations and flavors, here are some reference benchmarks looking at the performance of some of the kernel options available to Clear Linux users...

http://www.phoronix.com/scan.php?page=news_item&px=Clear-Linux-Kernel-Reference

[Volta]

Current kernel: 5.4.8, mainline kernel: 5.5-rc5. What did you just benchmark?

[metak]

Current kernel: 5.4.8, mainline kernel: 5.5-rc5. What did you just benchmark?

He benchmarked what is packaged in Clear linux repos. Although, the mainline and current haven't been updated in a while.

 
 linux-5.4.8-886 linux-current-5.3.1-11 linux-lts2018-4.19.93-102 linux-mainline-5.3.0-10 linux-preempt-rt-5.2.21-47

[woife]

@Michael: If preempt-rt is tested, the most interesting property to test would be ..... preemption

As far as I see all the tests in this benchmark test different kinds of throughput. I would assume that a preempt-rt kernel would loose in most of those tests, and indeed, it does.

An interesting RT-test is cyclictest, which also has a Phoronix Testsuite profile: https://openbenchmarking.org/test/pts/cyclictest
But running cyclictest on an idle system is no help. What would be really interesting would be results of cyclictest under different load conditions (network, CPU, graphics, memory, ...) and then comparing the results of different kernels.

[vladpetric]

This is as good a time as any to thank Michael for his long term contributions to FOSS benchmarking - Keeping it Honest and Keeping It Scientific.

Yes, I know, I should probably be more thankful with my wallet . Working on it!

[bug77]

@Michael: If preempt-rt is tested, the most interesting property to test would be ..... preemption

As far as I see all the tests in this benchmark test different kinds of throughput. I would assume that a preempt-rt kernel would loose in most of those tests, and indeed, it does.

An interesting RT-test is cyclictest, which also has a Phoronix Testsuite profile: https://openbenchmarking.org/test/pts/cyclictest
But running cyclictest on an idle system is no help. What would be really interesting would be results of cyclictest under different load conditions (network, CPU, graphics, memory, ...) and then comparing the results of different kernels.

Well, seeing what and how much (if anything) you sacrifice to get rt is also interesting

[Ipkh]

There can never be a clear winner in scheduling tests. Real world workloads are just too varied and can not really be benchmarked. Only actual server admins or developers could test production equivalent systems to see which scheduler works best.

And stupid programmers that don't check the documentation before trying things just add more flame bait to the conversation. Linux's strong language was right on point.

[vladpetric]

There can never be a clear winner in scheduling tests. Real world workloads are just too varied and can not really be benchmarked. Only actual server admins or developers could test production equivalent systems to see which scheduler works best.

Maybe 10 years from now apps could publish their own meaningful business metrics to the kernel (e.g., end-to-end latency, throughput for real work, not idling) and the scheduler would figure out a close-to-ideal plan, using some straightforward machine learning.

One can dream ...

[insilications]

It seems the linux-preempt-rt package is running 5.4.5 now (4 days ago)

https://github.com/clearlinux-pkgs/linux-preempt-rt