Linux 5.0 Kernel Performance Is Sliding In The Wrong Direction

[phoronix]

Phoronix: Linux 5.0 Kernel Performance Is Sliding In The Wrong Direction

With the Linux 5.0 kernel performance approaching the finish line, the past few days I've been ramping up my tests of this new kernel in our benchmarking farm. Unfortunately, when looking at the results at a macro level it's pointing towards Linux 5.0 yielding lower performance than previous kernel releases.

http://www.phoronix.com/vr.php?view=27563

[andyprough]

Is it possible there's something wonky with whatever is going into that PPA? I've compiled quite a few of my own linux- libre 5.0 rc kernels, and haven't noticed any difference between them and the 4.20 kernels I've been compiling. Of course, I haven't run any PTS benchmarks either.

[DanglingPointer]

Would be good to script-compile v5 source using best parameters and compare to Ubuntu mainline v5 kernel.

[sa666666]

Would also be nice to test the latest 5.x with all mitigations turned off, to see if it can get back to 4.13 levels, or if this is with us for good.

[Space Heater]

Some of these tests are producing very strange results that frankly don't make sense.

e.g.

NAS Parallel Benchmarks are about 20% slower after 4.12 - what would possibly account for that massive drop off?

Stress-NG SysV Message passing, a giant nearly 2x increase from 4.12->4.13 and a similar-sized drop from 4.13->4.14.

[andyprough]

Would be good to script-compile v5 source using best parameters and compare to Ubuntu mainline v5 kernel.

Yup. The problem could be on Ubuntu's side.

[SofS]

I understand the difficulties related to bisecting the kernel looking for the commits related to the regressions and that investigating such wide spread issue would be too much to ask for without a supercomputer. However, the most extreme cases could hint at something, three in particular seem like good candidates for bisecting, sockperf latency under load, postgre buffer read-write and sockperf throughput.

[Marc Driftmeyer]

Or the fact the more complete a modern OS Linux becomes the more it can no longer be just a Server OS w/o tailoring it from the mainline kernel. You want all these modern services, well they come with a price.

[tildearrow]

This is why I am reluctant to upgrading from 4.18...

[Wojcian]

Or the fact the more complete a modern OS Linux becomes the more it can no longer be just a Server OS w/o tailoring it from the mainline kernel. You want all these modern services, well they come with a price.

It doesn't make much sense. It's a regression, mitigations or problem on Ubuntu's side. Linux was never just a server OS and performance matters for Linus the most. Ubuntu has some debugging enabled, so there could be some change that affects performance as well. Btw. Good job Phoronix. However, sometimes tests are so random it's hard to tell if we're good or bad. It would be nice to have few main benchmarks plus additional ones.