tex-live performance

Anton Ertl anton at mips.complang.tuwien.ac.at
Sat Apr 3 16:18:50 CEST 2021


I have been using LaTeX for benchmarking (as well as production, but
this mail is about benchmarking) for many years, resulting in results
on machines from the early 1990s until today:

https://www.complang.tuwien.ac.at/franz/latex-bench

For the input used for this benchmark and instructions see

https://www.complang.tuwien.ac.at/anton/latex-bench/

Of course, if we want to use this for comparing the hardware, the
actual work done for this benchmark should be relatively similar.  And
indeed, I have done some runs using performance counters on AMD64
systems from Debian 5-10 where I always found that the benchmark
executed around 2000M instructions (specifically, between 1949M
instructions and 2108M instructions, where the number of LaTeX
packages installed have apparently quite a bit of influence on the
executed instructions: On the system where I saw 1949M instructions
earlier I now see 2072M).

Today I ran the benchmark on a Debian Testing (to become Debian 11)
system (with just texlive-base and texlive-latex-base installed) and
found that it executes 2606M user-level instructions; I measured this
with

perf stat -e instructions:u latex bench

The Debian 10 texlive package is based on TeX Live 2018, the Debian
Testing one on TeX Live 2020.  So I wonder: Has there been a change
between these two TeX Live releases that could explain this
difference?  Other explanations are possible, e.g., the Debian texlive
maintainers might build the Testing texlive package with less
optimization than they used for Debian 10.  But I have to start asking
somewhere, so I am doing it here.

Thanks in advance and thanks for providing TeX Live.

- anton


More information about the tex-live mailing list.