Other Linux Distributions Begin Analyzing Clear Linux's Performance Optimizations

[fabiohl]

Considering that Clear is not a general use OS, some of the optimizations coming to those would be welcome. An average desktop computer user will not use Clear, it doesn't even have a UI.

There's a bundle available that installs a Basic xfce desktop. But, surely, clearlinux is obviously focused on server, containers and et cetera.

[chimpy]

I know Clear Linux isn't a standard desktop distro, but it looks like it would be great on a Chromebook (with xfce). Michael, any chance you'd be interested in running some benchmarks?

[peppercats]

But you know that's not even half of it. Intel's CPU's are pretty good. I wouldn't exactly call them great though. Everybody thinks that because they have the performance crown, that must mean their architecture is great, but it's just not true. Believe me their architecture has some serious scalability issues.

EDIT: If Intel was AMD's size there is no way in hell their architecture could achieve the performance they've managed to get from it.

EDIT: I think GF was the worst thing AMD ever did and I think it's Intel's fault partly. (Dirk Meyers, great CPU designer, horrible CEO)

It's not a conspiracy theory if they got sued for it and AMD won in a court of law...
The FTC even investigated it.

https://www.ftc.gov/os/adjpro/d9341/091216intelcmpt.pdf

8.
Third, Intel used its position in complementary markets to help ward off competitive threats in
the relevant CPU markets. For example, Intel redesigned its compiler and library software in or about
2003 to reduce the performance of competing CPUs

[starshipeleven]

Actually clueless. I don't care much for intel, mostly cus unfair prices. I don't use AMD cus their cpus are weaker (even if intel has that amdcripple thing in their compiler), I'm one of the people waiting for zen in hopes of fair pricing unlike that intel overpriced shit.

That's a PowerPC user right there. (yeah I know, you're buying used stuff like everyone with a brain does in the last 5 years)

[starshipeleven]

If the process of "commenting out" the unused code is performed automatically by a compiler (it can be a JIT compiler), it's a valid approach. Otherwise it is questionable.

I am not sure what kind of "commenting out the code" we are talking about here. Can somebody clarify this please?

the fedora guy said "not even patches which could become upstream (they do things like commenting out chunks of code with single line commit messages)"

commit messages means it is done by humans. Unless they got a compiler that also uses git.

[duby229]

I think comment unused code is a also a validade optimization, among others of couse.

If it's unused code there is no need to comment it out. What they do is look for previously profiled behavior and then comment out or write around it.

[starshipeleven]

If it's unused code there is no need to comment it out. What they do is look for previously profiled behavior and then comment out or write around it.

Also called "cheating" for the uninitiated.

[carewolf]

This is EXACTLY why in Clear Linux we've worked on FMV (Function Multi Versioning, where you can mark specific hot functions for the compiler to create multiple copies for different instruction sets), on making sure /usr/lib64/avx2 works (where we put specific avx2 optimized libraries that are used on AVX2 systems, while the /usr/lib64 lib is used on all other systems) etc etc...

Do you also have an extra sub-dir for SSE4? I find SSE4.1 and AVX2 to be the most significant feature steps for autovectorization. SSSE3 has the shufb instruction and SSE4.1 the zero-extend and 32-bit mullo, that really increases the number of functions GCC and Clang can autovectorize. AVX2 seems to be used a lot less automatically at the moment, mostly both compilers just uses the AVX-prefix but still only vectorizes to 128-bit.

[kurros]

Do you also have an extra sub-dir for SSE4? I find SSE4.1 and AVX2 to be the most significant feature steps for autovectorization. SSSE3 has the shufb instruction and SSE4.1 the zero-extend and 32-bit mullo, that really increases the number of functions GCC and Clang can autovectorize. AVX2 seems to be used a lot less automatically at the moment, mostly both compilers just uses the AVX-prefix but still only vectorizes to 128-bit.

Since Clear Linux is built for Westmere or newer that shouldn't be necessary, since they all have SSE4.1/4.2/POPCNT)

[cj.wijtmans]

This is EXACTLY why in Clear Linux we've worked on FMV (Function Multi Versioning, where you can mark specific hot functions for the compiler to create multiple copies for different instruction sets), on making sure /usr/lib64/avx2 works (where we put specific avx2 optimized libraries that are used on AVX2 systems, while the /usr/lib64 lib is used on all other systems) etc etc...

Yes you can just give up and throw your hands in the air (and conspiracy theories are fun) but we tackled the problem (and not by just picking AVX2 as the minimum CPU generation, that would be great for performance but not practical to other distros etc)

That is a really nice technique that all distros should use. For gentoo not that useful. I am really interested on how to tweak the vanilla kernels scheduler do you have any source for pointers? The scheduler seems to work fine for me on an i5 without hyperthreading.