Hi all,

I am using Fedora 29 and some others (Debian testing, jessie, ubuntu 16.04, 18.04, etc. in VMware) on my Core i7 4790K rig, building BOINC software with GNU C/C++ and Intel C/C++ optimizing compiler. The binary is used in a BOINC project https://csgrid.org/csg/ and does a lot of floating point calculations, and the source is open to public. So I'm trying to build a faster one. It uses only cpu. No OpenCL or cuda is necessary. But AVX2 extension is available to my rig, which can be enabled with "-O3 -mavx2" or "-O2 -mavx2 -ftree-vectorize" options.

So far, I found plain vanilla ubuntu 16.04.1 (not upgraded with "apt update; apt upgrade") is the fastest (distributed application is build with this. gcc-5.4.0), and I cannot build a faster one than this, even with Intel optimizing compiler. Compilers older than gcc-5.4.0 cannot compile the source because of syntax problems.

So I wonder which combination of gcc version, glibc version, distro is the best. Ironically, newer gcc/glibc isn't necessarily fast. The next fastest one was debian jessie, gcc-6.4(?), I forgot glibc version.

Someone give me advice? I'll try on vmware.

Thanks in advance!!

PS: BTW I tried compiling another project Amicable Numbers which uses integers mostly and found Fedora 29 (gcc-8.2.1 glibc-2.28) is the fastest. Weird world.

I won't be very helpful, but few thoughts...

Before your tests, do you get ride of everything unnecessary (services mainly, X, ...) on each distro ? Because distro keep more or less things up by default.

It seems a lot about hardware optimizations, so the obvious ones for me, would be gentoo or LFS. Among the gentoo community, you should be able to get a lot of ideas and advice to build the fastest system. This is almost their purpose.
Newest stable gcc should gives better results as it brings new and better optimizations and options.

Out of curiosity, which tests do you run ?

Thank you for your reply.

Since my base is Fedora 29, any test is run on Fedora 29. Fortunately my application is statically linked in vmware, it can be measured on the same Fedora 29 environment. Testing is very simple. Stop BOINC which consumes most of cpu power and see no other cpu consuming service is running with "top", then just let it calculate a set of test data for about 30 seconds along with the comparison application which is normally used and see which is faster and how much faster. This application makes a file at the beginning, and after that adds lines to another file, so by comparing the file time with "ls --full-time" (exactly I made an alias), the run time can be precisely measured. Repeat this some times and confirm the run time is stable.

Gentoo is certainly worth a try. If gcc is changed to 8.x and recompile glibc 2.28 or so, it would be nice. I have no experience with LFS.

Thanks!!