[Linux-cluster] GFS vs EXT3 performance

[Cahill, Ben M]

Hi all,

Based on running iostat, it looks like EXT3 uses significantly fewer
block writes than GFS to do the tar/rm exercise:

dotar.sh:

cd /gfsmount
cp /path/to/linux-2.6.7.tar.gz .
tar -xzf linux-2.6.7.tar.gz
sync


dorm.sh:

rm -rf *
sync


I did this with virgin GFS and virgin EXT3 filesystems, using the same
physical drive (sdd), remaking each fs before running dotar.sh.

iostat sdd 60 2 showed that:

GFS uses ~873,000 block writes for dotar (took avg 50 sec)
GFS uses ~280,000 block writes for dorm (took avg 25 sec)

EXT3 uses ~556,000 block writes for dotar (took avg 18 sec)
EXT3 uses ~  8,900 block writes for dorm (took avg 2.0 sec)

This shows at least some symptom as to why GFS is not as fast as EXT3,
especially for the rm.  Does anyone know why GFS couldn't (at least in
theory, using one node) be as block efficient as EXT3?

These tests were run on one node, using nolock, no volume management.  1
GHz dual Xeon, 1 GB mem, using 1 FC drive in JBOD tower.  CVS 10/25/04,
2.6.9 kernel, with debugging options enabled (this makes a significant
difference in CPU usage).

-- Ben --