XanMod's Linux 5.10 Kernel Helping Tap Extra Performance With The AMD Ryzen 9 5900X

[phoronix]

Phoronix: XanMod's Linux 5.10 Kernel Helping Tap Extra Performance With The AMD Ryzen 9 5900X

For those wondering how the likes of the XanMod and Liquorix kernel spins are competing these days with the mainline Linux kernel, here are some fresh benchmarks looking at these popular derivatives of the Linux kernel. XanMod in particular atop Ubuntu can easily help squeeze extra performance out of the system as shown by these benchmarks on an AMD Ryzen 9 5900X desktop.

http://www.phoronix.com/vr.php?view=29870

[zeb_]

Very nice article, thanks for the comparison.

A question regarding I/O schedulers: it seems liquorix kernel is sometimes impaired by the fact it uses bfq on an NVMe drive. But it is easy to change with a udev rule. According to Archlinux wiki at https://wiki.archlinux.org/index.php..._I/O_scheduler the advice is none for NVMe, mq-deadline for SSD and bfq for rotational. Would using that rule.improve liquorix?

[schmidtbag]

For just a patched kernel (as opposed to applications and libraries being patched in Clear), those are surprisingly good results.

[ipsirc]

Please test pfkernel as well!

[damentz]

Very nice article, thanks for the comparison.

A question regarding I/O schedulers: it seems liquorix kernel is sometimes impaired by the fact it uses bfq on an NVMe drive. But it is easy to change with a udev rule. According to Archlinux wiki at https://wiki.archlinux.org/index.php..._I/O_scheduler the advice is none for NVMe, mq-deadline for SSD and bfq for rotational. Would using that rule.improve liquorix?

Yes, you can use udev to force your nvme drive to use "none" as your IO scheduler. However, if you're doing things that need that throughput, perhaps you should not be using Liquorix (and Zen Kernel by extension)? Liquorix is configured for responsiveness and low latency behavior at the cost of power consumption (and some throughput). If it wins any throughput benchmarks it's almost by accident . It's for systems that are multitasking and the multitasking behavior must produce the least amount of side effects on neighboring processes.

If you're at the point where you're tuning options to favor throughput, something you'd expect for systems treated like servers/appliances, I'd go with XanMod. It makes enough significant changes that it's a no brainer if you want a free upgrade in those aspects. Just know that throughput is not what makes a system feels fast, it's responsiveness. That's why you don't typically run low latency kernels on servers since there's no way to "feel" the difference, so just go all in on throughput.

And one final note on the git benchmark. Most git operations are single or low threaded operations. MuQSS really doesn't do a good job keeping those threads on the same cores for fairness reasons, and in turn, cpufreq doesn't know how to set the right frequency. The only solution I've found to that is to just tell cpufreq to go full speed at 45% core utilization. It works most of the time but it's best to run the performance governor otherwise. Most modern CPUs idle very efficiently so you can get away with that. For laptops though, stick to ondemand.

EDIT: Just noticed that this time around, Michael stuck with the default kernel configuration. Liquorix in these test was using the performance governor. Seems like in this case, MuQSS just can't keep git on the same core to save its life and loses on throughput there.

[geearf]

A question regarding I/O schedulers: it seems liquorix kernel is sometimes impaired by the fact it uses bfq on an NVMe drive. But it is easy to change with a udev rule. According to Archlinux wiki at https://wiki.archlinux.org/index.php..._I/O_scheduler the advice is none for NVMe, mq-deadline for SSD and bfq for rotational. Would using that rule.improve liquorix?

Hmmm, I wonder. Every time I look at Paolo's tests, they make me use bfq for everything not just rotational.
Sure I may lose a bit of throughput but that's not a big deal compared to the responsiveness gained.
Of course his tests may be a bit biased, but after talking with him, I don't know, I just trust him.

As for this, it'd be nice to include the TKG kernel, with its default options, to see how it compares.

[theriddick]

Been using Xanmod for my Ryzen 3600 setup, seems decent so far. Only 5.10 allowed for 6800XT card to work however. Haven't bothered with the SAM stuff yet, will wait until its out of the beta phase.

[zeb_]

Yes, you can use udev to force your nvme drive to use "none" as your IO scheduler. However, if you're doing things that need that throughput, perhaps you should not be using Liquorix (and Zen Kernel by extension)? Liquorix is configured for responsiveness and low latency behavior at the cost of power consumption (and some throughput). If it wins any throughput benchmarks it's almost by accident . It's for systems that are multitasking and the multitasking behavior must produce the least amount of side effects on neighboring processes.

If you're at the point where you're tuning options to favor throughput, something you'd expect for systems treated like servers/appliances, I'd go with XanMod.

Thanks for your comments. It is now standard to find NVMe (M.2) drives on desktop motherboards, not only on servers.
Besides, I thought that Liquorix and XanMod were both optimised for desktop use, not server. Tuning the I/O scheduler overrides the default kernel option - but do you mean there are still some settings in the kernel - in addition to the selected I/O scheduler - that change the responsiveness vs throughput?

[zeb_]

The best way to select the I/O scheduler is certainly to identify relevant tests depending on your computer usage: as a desktop, with a lot of encoding or not, with a lot of compilation, or as a server. In other words, which kernel/scheduler gets the most 1st places in the relevant categories for your usage, instead of looking at the 143 benchmarks as a whole.

It would be nice to see the outcome with the linux-ck kernel too.

[sn99]

Can you put up a guide for building the most optimized kernel, I believe your experience will yield fascinating results