Fresh Linux 4.16 Kernel Benchmarks With KPTI & Retpolines

[darkbasic]

Yep, only benchmark uses that and no one else, so it represents - nothing

...

It represents the status before spectre and meltdown got discovered. What's wrong with that? What people want to know is how hard the performance impact is and that's how you're going to measure it: 4.16 has all the fixes while 4.14.0 has none.

[dungeon]

It represents the status before spectre and meltdown got discovered. What's wrong with that?

Wrong is that people will make wrong conclusion "how 4.14 is better than something else" aka immediately first comment by PLum get this way

While point should be "how much performice-wise mitigation patches affects same kernel version" and this article does not answering that, so conclusion is nowhere - as article mostly talks about some it seems unrelated regression

[Steffo]

Linux has some serious regressions since 4.15! Hope they fix it!

[darkbasic]

Wrong is that people will make wrong conclusion "how 4.14 is better than something else" aka immediately first comment by PLum get this way

While point should be "how much performice-wise mitigation patches affects same kernel version" and this article does not answering that, so conclusion is nowhere - as article mostly talks about some it seems unrelated regression

I think that it is not possible to completely disable mitigation patches for all attack variants, otherwise Michael would have done it.

[toyotabedzrock]

Some of the spectre patches seem to be permanent. This will be a problem if they are mitigated in hardware and the expected performance doesn't return.

It seems odd Gimp performance went down, I wouldn't have expected that.

[foppe]

This will be a problem if they are mitigated in hardware

Not to worry.

[HenryM]

I wonder if kernel 4.14.0 performance equals that of 4.14.24. After all those repotoline/ktpi patches have been backported to the 4.14 branch.

I have compared an up-to-date 4.14 (4.14.23-1-lts on Arch) to a few versions of 4.15 and 4.14 always comes out ahead, sometimes significantly so. and according to PTS, that is with "KPTI + __user pointer sanitization + Full generic retpoline Protection".

[sireangelus]

how do you disable them?

[dungeon]

I think that it is not possible to completely disable mitigation patches for all attack variants, otherwise Michael would have done it.

Well, we shouldn't promote disabling of anything as in reality most people won't disable anything and just run default . That is why i think most people are interested to see "before & now" default comparison like this:

https://www.phoronix.com/forums/foru...28#post1011228

[darkbasic]

Well, we shouldn't promote disabling of anything as in reality most people won't disable anything and just run default . That is why i think most people are interested to see "before & now" default comparison like this:

https://www.phoronix.com/forums/foru...28#post1011228

Backports don't contain ALL mitigation patches yet, so if you want to know the cumulative impact of ALL mitigation patches you still need to disable everything on top of latest development kernel. The alternative is to benchmark 4.14.0 against 4.16-rc like Michael did, but that means you can possibly hit other regressions as well.