performance issue with TeX Live

[Emmanuel Beffara]

Hello Guix devel,

I am facing a severe performance issue with TeX Live: compilation of any
document is an order of magnitude slower with a Guix installed system as
compared to a manual installation. Is anyone confronted to this phenomenon, or
is there a way to fix this ?

I suspect the problem comes from kpathsea, but I may be misinterpeting my
observations. The fact is that there is an enormous amount of file accesses
for any compilation.

As an experiment, I compared two installations.

- One is a pure Guix installation, managed with `guix home`, that contains a
  rather large collection of packages that I include through a custom package:

    (define-public texlive-scheme-eb
      (package
        (name "texlive-scheme-eb")
        (version (number->string %texlive-revision))
        (source #f)
        (build-system trivial-build-system)
        (arguments (list #:builder #~(mkdir #$output)))
        (propagated-inputs
          (list texlive-scheme-medium
                texlive-collection-fontsextra
                texlive-collection-latexextra
                texlive-collection-pictures))
        (home-page "https://www.tug.org/texlive/";)
        (synopsis "EB's custom installation scheme")
        (description "This my a TeX Live scheme with what I use.")
        (license (license:fsf-free 
"https://www.tug.org/texlive/copying.html";))))

- The other is manual, I set up a container with

     $ guix shell --container -F --network --share=$HOME/opt=/opt bash 
coreutils curl grep gzip ncurses perl sed tar wget

  in which I installed TeX Live using `install-tl` as found on
  <https://tug.org/texlive/>, in the folder `/opt` exposed in the container.
  I installed the same package set by selecting the medium scheme in
  `install-tl` and installing the collections above with the provided `tlmgr`.

On my machine (a rather recent and powerful Dell Latitude 7410 with 16GiB RAM
and 12 cores), I observe the following timings:

- For a minimal LaTeX document (title, author and one word of body text)
  running `time pdflatex doc.tex` I get the following timings in the Guix
  version:

    real    0m4,288s
    user    0m3,140s
    sys     0m1,148s

  and in the manual version:

    real    0m0.773s
    user    0m0.729s
    sys     0m0.044s

- For a small beamer slideshow with nothing exotic (474 source lines, 17
  slides, no tikz graphics or anything), I get the following timings in the
  Guix version:

    real    1m0,122s
    user    0m14,337s
    sys     0m44,279s

  and in the manual version:

    real    0m4.554s
    user    0m4.446s
    sys     0m0.108s

I tried doing `strace` on the pdflatex calls to investigate further and it
appears that the behaviours of the two versions are largely different.
Counting the number of system calls of each kind gives the following, for the
most frequent calls:

- minimal document, Guix version:

   112860 newfstatat
    10491 getdents64
     5247 openat
     5246 close
     4397 access
     3141 read

- minimal document, manual version:

     2772 read
       90 openat
       72 access
       64 close
       64 newfstatat
       60 getdents64

- slideshow, Guix version:

  2831722 getdents64
  1538560 newfstatat
  1498287 access
  1415296 openat
  1415295 close
     4283 read

- slideshow, Guix version:

     3913 read
     1288 getdents64
     1136 access
      960 openat
      925 close
      920 newfstatat

So apparently no file hash is used in the Guix version and a large part of the
`texmf-dist` folder is browsed, probably several times.

-- 
Emmanuel