Upgrading Linux 5.15 LTS To Linux 6.1 Can Yield Additional Performance For AMD EPYC "Milan" Servers

[phoronix]

Phoronix: Upgrading Linux 5.15 LTS To Linux 6.1 Can Yield Additional Performance For AMD EPYC "Milan" Servers

Released on Sunday was Linux 6.1 and in addition to having many new features making it all the more exciting is that it's expected to be this year's Long-Term Support (LTS) kernel release. As such it will see widespread adoption particularly among servers and much interest from the hyperscalers. For those weighing an upgrade from last year's Linux 5.15 LTS kernel, Linux 6.1 with initial testing on an AMD EPYC Milan-X 2P server has shown a nice speed bump is possible across a wide-range of workloads.

https://www.phoronix.com/review/linux-61-amd-epyc

[schmidtbag]

Pretty impressive differences for a kernel update.

[CochainComplex]

This is quite in stark contrast to the latest kernel bench on gen 13 Intel CPU. As far as I remember in that benchmark the newer the kernel the worse the performance?
Is this enough "evidence" to speculate if the recent kernel improvements do help large corecount server CPUs but not small desktopesque systems or was the regression on the Intel just an outlier ?

Edit: https://www.phoronix.com/review/raptor-lake-kernels this is the mentioned bench..ok overall there is not much of a regression.

[Danny3]

Does the Linux 6.1 kernel come with the MGLRU feature enabled by default?
Because I remember seeing some spectacular improvements, on some tests, when it was in development and now, I'm not sure I can see them anymore.

[Michael]

Does the Linux 6.1 kernel come with the MGLRU feature enabled by default?
Because I remember seeing some spectacular improvements, on some tests, when it was in development and now, I'm not sure I can see them anymore.

No not yet default

[kozman]

This is quite in stark contrast to the latest kernel bench on gen 13 Intel CPU. As far as I remember in that benchmark the newer the kernel the worse the performance?
Is this enough "evidence" to speculate if the recent kernel improvements do help large corecount server CPUs but not small desktopesque systems or was the regression on the Intel just an outlier ?

Edit: https://www.phoronix.com/review/raptor-lake-kernels this is the mentioned bench..ok overall there is not much of a regression.

I think the bad benches we saw were on a mid-6.1rc kernel and much of the regressions were noticed around rc5 > rc7 and fixed in rc8 and final 6.1. Hard to say if there's lingering regressions but it seems like the worst ones got taken care of with the rest likely get shaken out in a few dot releases. 6.1 gonna be LTS so plenty of time to work on odds and ends.