Tweaking Ubuntu 17.10 To Try To Run Like Clear Linux

[phoronix]

Phoronix: Tweaking Ubuntu 17.10 To Try To Run Like Clear Linux

Even with the overhead of having both KPTI and Retpoline kernel support in place, our recent Linux distribution benchmarks have shown Intel's Clear Linux generally outperforming the more popular distributions. But if applying some basic performance tweaks, can Ubuntu 17.10 perform like Clear Linux? Here are some benchmarks looking at a few factors.

http://www.phoronix.com/vr.php?view=25852

[dungeon]

For fun I may rebuild the Ubuntu package archive with some performance optimized compiler flags compared to the safe flags used by Debian/Ubuntu, but that isn't something realistic for most individuals without having very high core count systems, but would be fun to benchmark.

Fire up 'apt-build world' and wait a day if something does not break in between, but it will for sure if you go too much crazy with flags

[caligula]

What about the perf of two other ricer distros, Arch and Gentoo?

[ext73]

Something poorly worked it out :P Despite this, you can do much more efficiently and responsibly on Ubuntu / Kubuntu:

http://www.netext73.pl/2018/01/launchpad-ruszy.html

[vegabook]

Honestly these results are shocking. 20% area performance improvements from Clear Linux. It's like I didn't have to upgrade my CPU in the past 4 years.

This surely is showing up a deficiency in the Linux kernel developers's priorities if Clear Linux can whack 'em so easily and so hard. 98% of data centers use Intel processors. Surely the Linux people should be spending a lot of resource on performance?? I mean Red Hat market capitalisation is 22 _billion_ dollars!!! They'rere making money hand over fist. So I'm not sympathetic to the idea of some poor guy coding assembly in his garage. This is professional stuff.

Seems like somebody is asleep on the job here. Intel has faced some bad press recently, but hey, their devs are running circles around the Linux core devs here.

With this kind of perf delta, no sysadmin worth his salt can motivate to be running anything BUT Clear Linux.

[indepe]

Great comparison! Can’t wait for more tests in this direction. It may require many small steps to get to ClearLinux’ performance level. The differences are large enough to influence choice of OS (vs Windows), but ClearLinux itself doesn’t even try to be of general use.

[Ropid]

What about the perf of two other ricer distros, Arch and Gentoo?

You might have some misconceptions about those. Arch is about being "simple" which in practice means being lazy when packaging software. The packagers try to change as little as possible about upstream releases. The compile flags are modest, pretty much just -O2. The configs are usually exactly what came by default with the upstream software. The end result is that things aren't riced at all by default, meaning you wouldn't test anything special when benchmarking in an automated fashion.

Gentoo is (I think) a similar situation and installing while using defaults would mean that benchmarking will not test anything special.

[GreatEmerald]

Gentoo is (I think) a similar situation and installing while using defaults would mean that benchmarking will not test anything special.

The defaults defaults, yes. If you can even say that there are defaults. But you're supposed to at least add `-march=native` to CFLAGS. That alone makes it so you don't need things like function multiversioning, and every single package is built that way, not just the immediate bechmarking app.

There was also that effort to make everything compile with LTO and some things that need it with -O3 which seemed to go fairly well.

[Michael]

Fire up 'apt-build world' and wait a day if something does not break in between, but it will for sure if you go too much crazy with flags

Yep that's the plan.

[eva2000]

But you're supposed to at least add `-march=native` to CFLAGS.

I use that for my source compiled nginx and php-fpm installs but need to switch off from them on cloud providers like linode where you can easily migrate from one vps host node to another where old and new host nodes have different cpu models/families which with march=native will give segfaults and illegal instruction errors.

Honestly these results are shocking. 20% area performance improvements from Clear Linux. It's like I didn't have to upgrade my CPU in the past 4 years.

Not really that surprised. I remember building my own MariaDB 5.2 RPMs way back optimised for Intel cpus instead of generic and some MySQL benchmarks were up to 50% faster

Thanks Michael for sharing more awesome benchmark comparison results. Love your work