The XanMod Kernel Is Working Well To Boost Ubuntu Desktop / Workstation Performance

[Etherman]

I read to page 3 before installing xanmod.

Edit: "evbug" spammed logs with _lots_ of errors so had to do a workaround by blacklisting evbug module.

[Raka555]

Ugh, throughput tests being used to benchmark low-latency kernel again... Thanks for testing it for performance (throughput) regressions I guess.

We really need ways to measure LATENCY to test things like this. Like test long does launching gedit take when your system is at 100% CPU and disk usage? What is the worst 1% FPS for a given game? How long does a click on the start menu take until the menu opens? How long does it take to switch from LibreOffice to Chrome to Firefox? How long does it take to open a new tab and open a simple page from disk in Chrome/Firefox? And so on and so forth.

I know it's difficult to test for latency, but there's very little point testing throughput on low-latency kernels...

Amen brother !

[perpetually high]

Nice! I've been a big fan of xanmod as I borrow a lot of its changes for own my config.

I noticed this while looking at its recent changelog for 5.4.8-xanmod5:

----
commit c7efd7e5d687f2cea431b8080fa98180323d3f6e
Author: Alexandre Frade <[email protected]>
Date: Sun Jan 5 10:01:50 2020 -0300

sched/core: Increase nr_migrate to 256 tasks to iterate in a single balance run

Signed-off-by: Alexandre Frade <[email protected]>
----

I know a few of us were discussing CFS tweaks and the previous xanmod change of 32 -> 128. So it looks like he's experimenting from 128 -> 256 now.

Anyways, good to know the kernel tinkering for desktops is paying off. Thanks for testing Michael.

[spirit]

Hi,
maybe it could be great to test some dxvk game to compare

I found this blog post benchmarking cfs vs bmq vs MuQss vs PDS scheduler,
and cfs was always the slowest. (don't known if cfs has been tuned like in xanmod)

https://flightlessmango.com/benchmarks/XDXMkN8m_FY

https://www.youtube.com/watch?v=XDXMkN8m_FY

[Alliancemd]

Not as buggy as Liquorix but Netflix does not work with this kernel(xanmod)... Error: M7020
Had to return to the default one...

[TuxTuxTux]

Not as buggy as Liquorix but Netflix does not work with this kernel(xanmod)... Error: M7020
Had to return to the default one...

I'm pretty sure that has nothing to do with the kernel, Netflix works perfectly fine for me with Xanmod.

[damentz]

Very interesting, so Liquorix regressed with 5.4.0-8.2 on workstation type benchmarks. That must be the switch from MC LLC (especially tuned for Ryzen processors), to SMT queue sharing. Maybe I should revisit if the gains for SMT on conventional hardware are worth it, even though SMT queue sharing got the best average frame rates and minimum frame times for my system (5960x, 3200mhz DDR4).

The second thing I noticed in the introduction is that the CPU governor is set to acpi-cpufreq + ondemand. On Liquorix, I set the default to performance to force the CPU cores to run at higher frequencies. This reason why is MuQSS does a relatively bad job keeping busy threads on the same core, which prevents cpufreq from raising the core frequencies, making it lose in single threaded benchmarks.

Of course, you can override the governor. I do that myself on my work laptop and personal to get better battery life, but I set the AC profile to run with a higher minimum frequency. But for my desktop, performance all the way, for acpi-cpufreq/intel-pstate.

Either way, the benchmarks look correct. Liquorix should always lose on throughput benchmarks. The only anomaly was the SQLite benchmark, possibly a regression in BFQ itself. I know there's a low latency flag - perhaps in this day and age it's worth disabling lowlatency by default for BFQ? Almost everyone uses SSDs for their primary storage now, so being so aggressive on the IO scheduler by default may not be ideal anymore.

[Linuxxx]

Very interesting, so Liquorix regressed with 5.4.0-8.2 on workstation type benchmarks. That must be the switch from MC LLC (especially tuned for Ryzen processors), to SMT queue sharing. Maybe I should revisit if the gains for SMT on conventional hardware are worth it, even though SMT queue sharing got the best average frame rates and minimum frame times for my system (5960x, 3200mhz DDR4).

The second thing I noticed in the introduction is that the CPU governor is set to acpi-cpufreq + ondemand. On Liquorix, I set the default to performance to force the CPU cores to run at higher frequencies. This reason why is MuQSS does a relatively bad job keeping busy threads on the same core, which prevents cpufreq from raising the core frequencies, making it lose in single threaded benchmarks.

Of course, you can override the governor. I do that myself on my work laptop and personal to get better battery life, but I set the AC profile to run with a higher minimum frequency. But for my desktop, performance all the way, for acpi-cpufreq/intel-pstate.

Either way, the benchmarks look correct. Liquorix should always lose on throughput benchmarks. The only anomaly was the SQLite benchmark, possibly a regression in BFQ itself. I know there's a low latency flag - perhaps in this day and age it's worth disabling lowlatency by default for BFQ? Almost everyone uses SSDs for their primary storage now, so being so aggressive on the IO scheduler by default may not be ideal anymore.

Just recently I configured an AMD Ryzen laptop for someone, and this is what I observed:

- the performance governor is not able to hit the boost speeds
- ondemand would lead to noticable stuttering
- best config: schedutil with /sys/devices/system/cpu/cpufreq/schedutil/rate_limit_us set to 0 (zero)

All of the above with Ubuntu 19.10 with Linux 5.3 - lowlatency.

[Guest]

Why? If you don't use ZFS, the modules aren't loaded/used are they?

This is purely a licensing problem. I don't even want to have these modules on my system.

[polarathene]

- the performance governor is not able to hit the boost speeds
- ondemand would lead to noticable stuttering

Assuming you hit the boost speeds with govenors other than performance, perhaps pushing all cores to as high as they could constantly is what prevented that due to either thermals or power? I assume if the other govenors achieved the boosts, that other cores were clocked lower, or if all cores were boosted the frequencies for them weren't as consistently high as performance mode would keep them at? This is a laptop after all so these issues would be more likely to be encountered?

The ondemand stuttering might be from latency to ramp up frequency? I saw an article recently(not phoronix) about the new 4xxx series having a massive improvement on this front, 80% faster or something.