performance test by Phoronix

[Michal Schmidt]

Dne Thu, 18 Dec 2008 13:18:41 +0100
Michal Schmidt <mschmidt at redhat.com> napsal:

> On Thu, 18 Dec 2008 13:51:27 +0200
> "Muayyad AlSadi" <alsadi at gmail.com> wrote:
> 
> > Hmmmmmmmm, faster processing slower IO......
> > what to consider
> > 1. SELinux
> > 2. some ext3 fs mount options
> > 3. having some daemon
> > ...etc.
> 
> 4. I wonder if they could have done some really stupid mistake, like
> having the distributions on different partitions on the disk. 
> 
> > BTW: I'm pleased by fedora performance,
> > I don't care about benchmarks
> 
> Well, benchmarks can point out real problems.
> 
> But I don't trust Phoronix. I guess I'll have to see if I can
> reproduce their results (not on the same hw, alas).

Last night I ran the IOzone benchmark on Fedora 10 and Ubuntu 8.10 to
see if I could reproduce a similarly big difference.

In short: there was no difference between Fedora and Ubuntu.

In detail:

My test machine is: Asus A8V Deluxe, Athlon 64 FX-53 (2,4 GHz), 2 GB
RAM, Seagate SATA 1TB ST31000340AS.

First I installed Fedora 10 into a 20 GB logical volume, ran the
benchmark, saved the results, then installed Ubuntu 8.10 into the same
logical volume (letting its installer mkfs it), and ran the benchmark.

I used x86_64 version of both distros, and I always installed a
base system without X11 and server daemons. In both cases I let the
installers use the current stable updates for their distros. Then I
installed the dependencies needed for the Phoronix test suite (php-cli,
gcc). I downloaded phoronix-test-suite-1.4.2, installed it, and used it
do fetch IOZone3_308 and install it.

I had a problem getting IOzone working on Ubuntu. P-T-S downloaded it
and built it, but it crashed on startup with "overflow detected". I
worked around it by copying the iozone binary from Fedora 10.

The benchmark was run always after booting to single user mode.
In all test runs the write performance was about 54 MB/s, never
differed by more than 0.3 MB/s.

Some random observations:
 - SELinux did not cause any significant slowdown in F10
   (I re-ran the test also with SELinux disabled.)
 - The mount options for the ext3 root fs as shown in /proc/mounts:
     - F10:    rw,errors=continue,user_xattr,acl,data=ordered
     - Ubuntu: rw,relatime,errors=remount-ro,data=ordered
 - When booted into the default runlevel, Fedora enabled ondemand CPU
   frequency scaling. Ubuntu did not.

Conclusion:
I still don't know what could have caused such a large difference
between Fedora and Ubuntu in Phoronix's results. It's not SELinux. The
mount options are different, but do not seem to influence IOzone results
either. Maybe Fedora has a problem only on the hardware they tested? Or
maybe they made a silly mistake.

Michal