Kernel Bisecting Has Never Been Faster Than With AMD EPYC + AMD Threadripper

[phoronix]

Phoronix: Kernel Bisecting Has Never Been Faster Than With AMD EPYC + AMD Threadripper

While Zen 3 is just around the corner, current-generation Ryzen Threadripper and EPYC processors continue to impress particularly for build boxes and tasks like bisecting the Linux kernel's massive codebase...

http://www.phoronix.com/scan.php?page=news_item&px=Kernel-Bisecting-5.9-Ongoing

[microcode]

This is spectacular. The Linux Foundation should be funding your lab, Michael.

[aphysically]

Looking forward to hearing about the TensorFlow regression and what happens with the Apache/HB/nginx regression. It would be cool if the TF fixes if needed made it into 5.9 since it's LTS but we'll have to see.

[Mitch]

The most fascinating bit here is that Linus alludes to regressions AND improvements due to the patch but also a staggering increase in context switching. I just can't imagine what workload would improve with such a patch.

[Danny3]

This is spectacular. The Linux Foundation should be funding your lab, Michael.

Completely agree with you on this one!
Michael is doing one of the most important jobs in the kernel development process.
It would be great if he wouldn't have to beg for people to turn off their adblockers or to subscribe to premium since the Linux foundation has a lot of money from the big companies which are helped a lot by Michael's regressions and performance problem findings.
Hopefully they will understand one day how important Michael's job is and decide to help him a bit.

[Mark Rose]

Does PTS automatically update the kernel, reboot, run the test, and so on, during bisecting?

[Michael]

Does PTS automatically update the kernel, reboot, run the test, and so on, during bisecting?

Yes it can with an extra PTS module.

[milkylainen]

The most fascinating bit here is that Linus alludes to regressions AND improvements due to the patch but also a staggering increase in context switching. I just can't imagine what workload would improve with such a patch.

Just because you're Linus Torvalds does not make you immune from performance regressions in a complex environment.
I think this more than ever stresses the point that Linux needs an official performance regression suite.

[mcoffin]

I've also been infinitely more productive in bisects in general, not just with the Linux kernel. I recently had to bisect across 2 full major revisions of mesa, and at < 1min of build time to build the full pacman package, I was able to get finished pretty quickly with my 3960X compared to the waiting hell that I used to go through before upgrading.

[Mitch]

Just because you're Linus Torvalds does not make you immune from performance regressions in a complex environment.
I think this more than ever stresses the point that Linux needs an official performance regression suite.

It's not a blame thing. I didn't know who even submitted or wrote the patch. I'm saying that in light of Linus's observation plus Michael's testing, it's unexpected to me that the patch improves anything. Not trying to rock the boat or anything. I like Linus and everyone who contributes to the Kernel. I generally assume they're doing their best with good intentions.