Regressions in Linux 5.9 kernel

Here is an article from Phoronix on the regressions.


Tests show that the regressions are serious. These issues were introduced by Linus Torvalds. There are also good quotes from Linus. It appears to be a real problem.

gary

7 Likes

Yeah, i also heard and read about this. Hope it gets figured out.

As happens so often with these things - there are no right answers: a patch fixed a very unfair behaviour for bad and worst cases but this also has negative impacts on overall performance.
Looks like some middle ground has to be found - some unfairness needs to be allowed to not regress too badly:
https://marc.info/?l=linux-ext4&m=160010568530799&w=2

1 Like

this is a dev kernel at this time , not a stable kernel
it unfortunately reveals ,
patches without any tests bench for somes regressions and investigation ,
only mickael from phoronix have done that with heavy cpu & bisect

in case of a stable new version , that check is missing just before the release
mickael with phoronix suite provides all tools for benchmark and performance

1 Like

The last comment so far in the Phoronix comments details using kernel 5.4 instead, as there’s been a 5-10% performance drop since that kernel. Had put the kernels on the ignore package list just now, but wondering now about going to 5 4 … last time I tried the LTS it ran badly, but will try a boot with it and see how it goes.

EDIT: it’s running really well so far. :+1:t2:

1 Like

I can confirm a little performance drop from 5.8.9 to 5.9rc5, however nowhere near 5-10%. More like 0,5% in my benchmarks.

2 Likes

I am sorry that I was so imprecise in my posting. The part that is important to me and educational to others are the comments by Linus Torvalds addressed to Phoronix. A great example of his thinking and how can this problem be solved.

I also posted this because I believe that users should be aware of the “deep” issues in how operating systems work and the difficult in choosing the “right solution” when there maybe no clear solution.

1 Like

You don’t have to be sorry, it was very useful info after all.
We have to wait for the final release before we draw conclusions, I’m confident the issues will be mostly fixed by then.

1 Like

Yes, I totally agree but I did not state my purpose clearly enough. I will do that from now on.

2 Likes

Here are the results created by Linus. And @freggel.doe wins with his suggestion that:

Blockquote
Looks like some middle ground has to be found - some unfairness needs to be allowed to not regress too badly
Blockquote
https://www.phoronix.com/scan.php?page=article&item=linux-59-unfairness&num=1
Blockquote
Also, some good stuff in the comments as well.

Btw, kernel 5.9 now supports ZTSD compression for both kernel and initram, resulting in rather small filesize and very fast decompression.

1 Like

Tested again yesterday: no performance regression at all with 5.9-rc6.
Benchmark results are almost identical to 5.8.

1 Like