Benchmarking Clear Linux With KPTI + Retpoline Support

[andyprough]

The fact of the matter is, these tests showed a lot of performance losses.

Agree with you on your other points. But I have yet to see "a lot of performance losses". So far, I haven't seen anything different than the typical regressions that set in when a new version of the kernel or GCC is being tested. A few tests show minimal improvements, most show either no effect or minimal impact, and some show a bit larger performance loss.

Assuming that the software fixes will be improved over a short period of time, this all looks like it will turn out to be a big nothing burger in terms of performance. Fact is, speculative caching has needed to be re-engineered correctly for a significant number of years, and this appears to be the catalyst to get that work done. So in the long run, this is most likely a net positive.

[R00KIE]

I'd say this is still not the end of the story, Intel has still not published any microcode updates in their download page[1] but my Lenovo has a very recent bios update which comes with a microcode update and lenovo itself says a microcode update is needed to fix/mitigate one of the problems[2]. I suspect some workloads may still get impacted and I suppose it's not just the kernel that needs to be patched/recompiled to include the retpoline fixes or similar. Don't forget gcc has pending patches to deal with this.

[1] https://downloadcenter.intel.com/dow...code-Data-File
[2] https://support.lenovo.com/pt/en/solutions/len-18282

[schmidtbag]

Agree with you on your other points. But I have yet to see "a lot of performance losses". So far, I haven't seen anything different than the typical regressions that set in when a new version of the kernel or GCC is being tested. A few tests show minimal improvements, most show either no effect or minimal impact, and some show a bit larger performance loss.

I didn't specify the amount of losses or how significant they were. I just said there were a lot of them, which is true - there are many tests that show measurable degradation. The fact I mentioned that these tests don't really reflect my opinion of Intel implies that I too found [most of] these to be minimal.

Assuming that the software fixes will be improved over a short period of time, this all looks like it will turn out to be a big nothing burger in terms of performance. Fact is, speculative caching has needed to be re-engineered correctly for a significant number of years, and this appears to be the catalyst to get that work done. So in the long run, this is most likely a net positive.

I wouldn't count on it. Seeing as this is a vulnerability at the hardware level, I doubt there's much you can do to fix the vulnerability other than disabling it entirely. Future processors may find a way to fix this problem securely, but I think this performance loss will be permanent. But again - for the most part, it's not that big of a deal. The only thing to be especially worried about are systems that can't be patched.

[arjan_intel]

I'd say this is still not the end of the story, Intel has still not published any microcode updates in their download page[1] but my Lenovo has a very recent bios update which comes with a microcode update and lenovo itself says a microcode update is needed to fix/mitigate one of the problems[2]. I suspect some workloads may still get impacted and I suppose it's not just the kernel that needs to be patched/recompiled to include the retpoline fixes or similar. Don't forget gcc has pending patches to deal with this.

[1] https://downloadcenter.intel.com/dow...code-Data-File
[2] https://support.lenovo.com/pt/en/solutions/len-18282

https://downloadcenter.intel.com/download/27431/Linux-Processor-Microcode-Data-File

[sa666666]

Seriously did AMD murder your family? Did you invest in them right before the release of Bulldozer? This article has NOTHING to do with AMD and yet they're the only thing you can focus on lately. If you're going to draw negative attention to them, you may want to stop comparing apples to oranges while cherry-picking results (fruit theme unintentional).

The fact of the matter is, these tests showed a lot of performance losses. It's not a big deal: these tests aren't going to prevent me from recommending Intel. Hell, this whole fiasco isn't something that would prevent me from recommending Intel - it was a mistake. But for whatever reason you're getting super insecure and it's really starting to get pathetic, and annoying. Relax - Intel isn't going anywhere. Quit your whining and just accept that everyone from all sides are taking a loss to some degree.

Agreed. This guy is starting to get as bad as GhostOfFunks was about Gnome. I begin to wonder if they're the same person, since they have the same (irrational) zeal. Oh, how I wish blocklists were supported in these forums.

[franglais125]

the post-fix consumer-grade 8700K annihilates a $4000 Epyc part in both Redis & ApacheBench.

https://www.phoronix.com/scan.php?pa...nchmarks&num=1

This is not so, read the last few comments on that article. As Michael clarified, in that article you cannot compare across CPUs, as the systems use different components (different SSD, RAM, etc). The point of that article was to see how each CPU was affected vs itself.

[DanL]

This guy is starting to get as bad as GhostOfFunks was about Gnome. I begin to wonder if they're the same person, since they have the same (irrational) zeal.

Different posting style.

[smitty3268]

So "from 0% up to 30%" actually means that out of 1001 apps tested average is 5%

So, on average negligible

Same like someone compare distros performance, these are also on average negligible with just couple weird cases here and there

Counterpoint: 5% is basically what you got out of an entire CPU generation upgrade from Intel over the last few years. So even a "small" 5% slowdown is years of progress wiped out overnight.

[linuxgeex]

Michael

Excuse me, Clear Linux is taking 282,978ms - that's 4 minutes and 43 seconds - to boot and you're saying that it's one of the fastest? On what planet does any stock linux distro take nearly 5 minutes to boot and that is considered fast?!

Obviously something is broken and this needs to be retested.

[edwaleni]

So "from 0% up to 30%" actually means that out of 1001 apps tested average is 5%

So, on average negligible

Same like someone compare distros performance, these are also on average negligible with just couple weird cases here and there

5% can be significant for a database that processes 24x7 or if a system processes voice traffic. As everyone says, your experience will vary.