Linux 6.6 Delivers Some Impressive Gains For AMD EPYC 9754 "Bergamo" Server Performance

[phoronix]

Phoronix: Linux 6.6 Delivers Some Impressive Gains For AMD EPYC 9754 "Bergamo" Server Performance

From my early testing thus far of the Linux 6.6 kernel in its very early state, some of the most impressive gains are happening on AMD's high core count server processors, the EPYC 9754 "Bergamo" in particular is enjoying some stellar improvements for various server workloads on this forthcoming kernel.

https://www.phoronix.com/review/linux-66-bergamo

[sdack]

The BFS scheduler is said to be based on EEVDF. However, with a name like the "Brain Fuck Scheduler" will many not have given a "fuck" about it and it first needed an Intel engineer to create something with a more respectable name before the larger community felt safe enough to get behind it. It would be interesting to see how EEVDF compares to BFS, but BFS looks abandoned now. There may be a lesson here though... Do not give your code a stupid name?!

[darkbasic]

Wow, this is like getting the performance benefits of a brand new architecture for free.

[schmidtbag]

Pretty remarkable considering these were already very powerful chips to begin wtih.

[Errinwright]

Introduction of BORE would also boost the lower core count parts I presume?

[damentz]

This is great, our first glimpse of what EEVDF is doing in 6.6. Pretty much it's either a minor or massive improvement across the board.

The BFS scheduler is said to be based on EEVDF. However, with a name like "Brain Fuck Scheduler" will many not have given a "fuck" about it and it first needed an Intel engineer to create something with a more respectable name before the larger community felt safe enough to get behind it. It would be interesting to see how EEVDF compares to BFS, but BFS looks abandoned now. There may be a lesson here... Do not give your code stupid names?!

The closest actively maintained scheduler that resembles EEVDF is PDS in Project-C. Not sure how hard it would be for Michael to add this into a benchmark, maybe compare v6.5 + PDS, v6.5 stock, and 6.6-rcN? That'll give us the best picture of what improvement mainline EEVDF is doing over out-of-tree attempts to implement a similar scheduling strategy.

[avis]

Would be nice to compare CFS with EEVDF in terms of the number of lines of code.

Could be that EEVDF allows these massive gains because it's literally 10 times less code.

That's not all of course, various loops and logic are equally important. EEVDF could be just more streamlined and logical.

[sdack]

...
The closest actively maintained scheduler that resembles EEVDF is PDS in Project-C. Not sure how hard it would be for Michael to add this into a benchmark, maybe compare v6.5 + PDS, v6.5 stock, and 6.6-rcN? That'll give us the best picture of what improvement mainline EEVDF is doing over out-of-tree attempts to implement a similar scheduling strategy.

This would be nice. Rumour (aka Reddit) also has it EEVDF would provide gains with Linux gaming in particular for CPU-bound games, but it also is said to help with stutters and rubberbanding. There are quite a few interesting benchmarking opportunities here.

[Volta]

Would be nice to compare CFS with EEVDF in terms of the number of lines of code.

Could be that EEVDF allows these massive gains because it's literally 10 times less code.

That's not all of course, various loops and logic are equally important. EEVDF could be just more streamlined and logical.

When there are gains with simpler code it's nothing else, but beauty. What's interesting change from CFS to EEVDF was very silent. Not even close to whatever CFS replaced.

[Kjell]

CachyOS has been using EEVDF with BORE tweaks for quite some time now