Tuesday, April 12, 2011

Comparing ext3 to ext4 benchmarks

Below is the output from bonnie++, installed from rpmforge on CentOS 5.5, running against a RAID 0 (Linux Software RAID) of two Seagate Barracuda LP ST32000542AS, formatted with ext3.
#bonnie++ -d /mnt/SOMEDRIVE -n 32:64:4:4 -q | bon_csv2html > /StorageServer/RAID0.ext3.html


Here is the same command but running on the same server, but this time running CentOS 5.6 with the same Linux software RAID 0 on two Seagate Barracuda LP ST32000542AS, but formatted with ext4.



We can see that Sequential Output Per Character is slightly higher, Block K/Sec is higher and with significantly lower latency, but Rewrite is lower and with higher latency on ext4. (Higher is better)

Sequential input on both Per Caharacter and Block K/Sec is also slightly higher and with lower latency on ext 4. Random seeks are also slightly higher on ext4. (Higher is better)

Sequential creates are higher, but require more CPU, while Sequential Reads are lower using about the same CPU. Sequential Deletes are also slower on ext4 than they where on ext3.

Where ext4 really seems to shine is on the Random Create section. Creating random files almost doubled per second on ext4 vs. ext3, all while using similar CPU. Random Reads where so fast on ext4 that I need to retest to get accurate results (shown by +++++ on the results). Random Deletes where about 25% faster on ext4 when compared to ext3 again while using less CPU, though not by much (1%).

Now these being benchmarks they give a nice indication about what kinds of performance I may be able to have, but I need to do real world testing. Hopefully I have some meaningful results after I finish upgrading all of my servers and filesystems. I will also benchmark our RAID 5 system before and after the conversion.

For our purposes it appears that ext4 is the better filesystem given our setup and procedures.

1 comment:

Steve Lippert said...

I just updated the post to include links to the full size graphics. That should help everyone that has been trying to look at these comparisons. I apologize it took so long.