[mdlug] Are Linux Distros Too Bloated?

[LAP]

On Sun, 16 Apr 2023 18:56:51 +0000
Robert James Fulner <fulner at alumni.nmu.edu> wrote:

> 
> I've been meaning to try Gentoo for sometime to see how much
> improvement I get by complying directly for my machine thereby
> eliminating overhead, but I've never got around to it.
> 

Although I do not, in general, appreciate software benchmarks I decided
to try a comparison using C software produced by the NIST called
SciMark 4.0:

https://math.nist.gov/scimark2/download_c.html

SciMark is similar to standard tests for high performance computing.

On my Gentoo (low-end) Xeon workstation, using optimized compiler settings,
the results were:

FFT Mflops: 2338.10 (N=1024)
SOR Mflops: 2062.17 (100 x 100)
MonteCarlo: Mflops: 943.78
Sparse matmult Mflops: 2550.52 (N=1000, nz=5000)
LU Mflops: 8320.08 (M=100, N=100)

************************************
Composite Score: 3242.93
************************************


Then I compiled Scimark4 with the generic options that are used by
the average GNU/Linux distro.  On the same machine the results were:

FFT Mflops: 1868.69 (N=1024)
SOR Mflops: 1520.70 (100 x 100)
MonteCarlo: Mflops: 762.05
Sparse matmult Mflops: 2771.88 (N=1000, nz=5000)
LU Mflops: 4883.40 (M=100, N=100)

************************************
Composite Score: 2361.34
************************************

This score is nearly 1000 Mflops lower!

Furthermore, it still uses an optimized libm from glibc but
it is still almost 1000 Mflops lower!

All I can say is that optimization matters.

This, however, is certainly not the end of the story.  There
are many factors at play when considering software performance.
Most GUI programs, which are geared to human sensation,
would probably not be noticeably influenced by such optimization.

But since it can be done, why not just do it.  That is my take.