benchmarking-gimp

benchmarking-gimp.rst

Goal

Address Jakub's concerns regarding the performance impact of PIE. See :

https://lists.fedoraproject.org/pipermail/devel/2013-April/181171.html

Why Gimp

Gimp has lot of dependencies :

$ ldd /usr/bin/gimp-2.8 | wc -l
77

$ du -hs /usr/bin/gimp-2.8
5.8M    /usr/bin/gimp-2.8

Now comparing this to libreoffice :

$ ldd /usr/lib64/libreoffice/program/soffice.bin | wc -l
81

$ du -hs /usr/lib64/libreoffice/program/soffice.bin
40K  /usr/lib64/libreoffice/program/soffice.bin

So Gimp is a more "suitable" target for benchmarking purposes.

Benchmarking Gimp

• Get test data from :

  http://www.eecs.berkeley.edu/Research/Projects/CS/vision/grouping/BSR/
  http://www.eecs.berkeley.edu/Research/Projects/CS/vision/grouping/BSR/BSR_bsds500.tgz

• Patch for building Gimp as PIE can be obtained from :

  https://github.com/kholia/unSPEC

• Benchmarking scripts can be obtained form :

  https://github.com/kholia/gimp-bench

• Copy flipscale.py to ~/.gimp-2.8/plug-ins/ and do chmod +x ~/.gimp-2.8/plug-ins/flipscale.py
• Copy fubatchpy.scm to ~/.gimp-2.8/scripts/
• time gimp -i -b '(call_py_flipscale "pics/*.jpg" 120 150)' -b '(gimp-quit 0)'
• For best results, keep the test data in RAMDISK.

Gimp Results

• ~ 39.160s ==> upstream compiled by me (PIE)
• ~ 39.750s ==> upstream stock (PIE)
• ~ 40.850s ==> "non-PIE" my build (no PIE)

Notes

• Running prelink on the "non-PIE" build resulted in no visible difference.
• This test is borrowed from http://www.exp-media.com/fr/node/19