Liquorix vs. Linux 6.12 Upstream Kernel Performance Across Many Workloads

[phoronix]

Phoronix: Liquorix vs. Linux 6.12 Upstream Kernel Performance Across Many Workloads

A Phoronix Premium subscriber a while back requested some fresh benchmarks of how the Liquorix downstream of the Linux kernel is comparing against the latest upstream kernel... Here are some benchmarks looking at the Liquorix flavor of the Linux kernel compared to upstream Linux 6.12...

https://www.phoronix.com/news/Linux-6.12-Liquorix-Performance

[moonwalker]

Do I understand correctly that most of the benchmarks test for throughput rather than latency/responsiveness, and thus the results should be completely expected with Liquorix tuning towards lower latency?

[damentz]

Do I understand correctly that most of the benchmarks test for throughput rather than latency/responsiveness, and thus the results should be completely expected with Liquorix tuning towards lower latency?

Yes and no, Liquorix _should_ perform worse in most throughput benchmarks, but I believe we're also seeing a regression in these benchmarks.

There's an outstanding bug report for a regression caused by the newly added code to support asymmetric CPUs, it probably has a hand in what you're seeing: https://gitlab.com/alfredchen/linux-prjc/-/issues/104

Effectively, benchmark/synthetic type workloads will run processes and threads that get stuck on queues and prevent higher priority tasks from running. The developer behind PDS understands what the issue is and is working on a solution, but currently this bug is in Liquorix. The way this manifests is that any benchmark that depends on careful ordering of nice levels or priorities will behave badly when all cores on a CPU are 100% saturated.

[pWe00Iri3e7Z9lHOX2Qx]

This is part of the challenge for optimized kernels. There are many orders of magnitude fewer users running these vs. vanilla or distro kernels. You really need a broad perf test automation suite to catch regressions.

[damentz]

Another interesting note, the benchmark forces amd-pstate with the powersave profile, the default for Liquorix is performance with acpi-cpufreq. PDS spreads its work against more cores so frequency scaling may not work in some cases, that might be what happened with the LiteRT benchmarks. I'll need to consider disabling intel-pstate and amd-pstate by default, or maybe just remove entirely if distributions and benchmarking suites continue to override the preferred scaling governers.

This issue actually happened on the last benchmark for Liquorix as well: https://www.phoronix.com/forums/foru...56#post1334056

[skeevy420]

Another interesting note, the benchmark forces amd-pstate with the powersave profile, the default for Liquorix is performance with acpi-cpufreq. PDS spreads its work against more cores so frequency scaling may not work in some cases, that might be what happened with the LiteRT benchmarks. I'll need to consider disabling intel-pstate and amd-pstate by default, or maybe just remove entirely if distributions and benchmarking suites continue to override the preferred scaling governers.

This issue actually happened on the last benchmark for Liquorix as well: https://www.phoronix.com/forums/foru...56#post1334056

That's just the way Michael does benchmarks. He goes with the distribution default settings to a fault. If I remember Ubuntu correctly, that isn't the fault of Pstate as much as it is Ubuntu's default settings with Power Profile Daemon. Even though you have it set as Performance in the kernel's config, Ubuntu's userspace governor handling overrides that. You might just need to update the install instructions so your kernel's users know to tweak their PPD or CPUFreq settings.

[nado]

Afaik, I've never heard of anyone who's serious about computing that actually uses these alternatives kernels.

[V1tol]

Alternative kernels are good for gaming. Or DAW (if you don't use macOS). Everyone else either use whatever is provided by their distro or use in-house systems built from scratch to meet their specific requirements. In any case 99% of users are perfectly happy with default settings.

[nado]

Alternative kernels are good for gaming. Or DAW (if you don't use macOS). Everyone else either use whatever is provided by their distro or use in-house systems built from scratch to meet their specific requirements. In any case 99% of users are perfectly happy with default settings.

I've generally read articles and benchmarks that show negligible differences, which in most real world scenarios are redundant. I don't know enough about it, but I can imagine it has some niche use cases where it might make sense. The other question is what side effects it might exhibit?

[kylew77]

In any case 99% of users are perfectly happy with default settings.

As much as it is a right of passage to compile your owner kernel, it really isn't needed nowadays for any FLOSS OS, not even Linux, for the vast majority of users. Every time I get tempted to compile a custom kernel for Linux tuned to that machine's CPU and hardware I get to thinking is that really going to help me in the long run. If that hardware breaks I need to find an exactly the same system to recover on and I have to recompile the kernel for each release of a bug fix kernel or newer version. It has too small of a performance boost for too much effort.