[tex-live] TeX performance over the years going downhill

[Gerben Wierda]

I have been busy trying to get some performance measurements for TeX so 
I could see if the Mac Intel version I have performs OK compared to teh 
Mac powerpc version. So, I created a very small initial test (which 
runs the cration of pdflatex.fmt for 10 times in a row with output to 
/dev/null) to see if there were basic differences at 
file/filesystem/etc level.

The results I got were weird. Mostly they were in line with the single 
CPU integer speed of the hardware, but some testers reported times that 
seemed blazingly fast compared to others. First, the suspicion was that 
the OS version was to blame. But it turned out that a very large part 
of the performance development was just TeX itself.

As one person reported (for 10 times "fmtutil --byfmt pdflatex" on a 
system for the rest completely equal):

> ===================================================
> TeXlive 7 (2002)
>
> $ pdftex --version
> pdfTeX (Web2C 7.3.7x) 3.14159-1.00b-pretest-20020211
> kpathsea version 3.3.7
>
>       28.06 real        18.31 user         4.61 sys
> ===================================================
>
>
> ===================================================
> TeXlive 2003
>
> $ pdftex --version
> pdfTeX (Web2C 7.5.2) 3.141592-1.11a
> kpathsea version 3.5.2
>
>       32.32 real        25.01 user         5.30 sys
> ===================================================
>
> ===================================================
> TeXlive 2004
>
> $ pdftex --version
> pdfeTeX 3.141592-1.20b-2.2 (Web2C 7.5.3)
> kpathsea version 3.5.3
>
>       56.57 real        45.47 user         7.46 sys
> ===================================================

And I have seen more reports corroborating this. Let me stress: apart 
from TeX, this system is completely the same. Sometimes people report 
that their older TeX is about twice as fast as the newer. The big 
difference seems to have been with some web2c release, probably the one 
where most binaries were folded into pdfetex.

My personal idea is that first we need some sort of reliable test for 
an *installed* TeX. That is, a test that uses local binaries, local 
pool files, but that for the rest is independent of what is or is not 
on the system (search differences, size of texmf tree differences, 
language setting differences, font differences, etc).

G