[CentOS] ext3/ext4 random write performance on ioDrive duo

Sat Apr 17 18:29:14 UTC 2010
Ryousei Takano <takano-ryousei at aist.go.jp>

Hi all,

I have evaluated Fusion-io ioDrive duo (PCIe direct attached NAND
flash memory storage) on the CentOS 5.4.
The experimental result shows the write performance with multiple
jobs (processes) on the ext3/ext4 filesystems is lower than that of XFS.
The question I have is what is the cause of the ext3/ext4 performance
limitation.

I have measured the random write performance on several filesystems
using fio benchmark (http://freshmeat.net/projects/fio/).
My PC box has dual Intel Xeon W5590 3.33 GHz (8 cores), 48 GB memory,
and Fusion-io ioDriveDuo (320 GB SLC).

The XFS performance grows as the number of jobs increases. On the
other hands, the ext3/ext4 do not.
In the both cases, I mount filesystems with the "noatime" option only.

The results are shown below:

block size      XFS     XFS     ext4    ext4
                       (1job)  (8job)  (1job)  (8job)
1024    21      71      22      21
2048    40      144     42      40
4096    74      290     76      76
8192    131     635     133     131
16384   212     1029    219     215
32768   310     1169    326     316
65536   399     1236    374     374
131072  748     1332    718     676
262144  895     1336    902     804
524288  1000    1327    974     905
1048576 1100    1308    1136    995
2097152 1220    1276    1179    1082
4194304 1230    1233    1270    1149
8388608 1309    1233    1243    1188

(The unit is MB/s.)

You can see the more detail information from here:
       http://code.google.com/p/pspacer/wiki/ioDriveDuo

Any comments and suggestion will be welcome.
If this list is not proper place to talk about this problem, can
anyone tell me the proper one?

Best regards,
Ryousei