EXT4 vs XFS for Oracle, which one performs better?

[UPDATE] Since publishing this post I received requests to benchmark ASM and Oracle Enterprise Linux. An updated post for this article EXT4 vs XFS vs ASM vs ASM + OEL, which one performs better? Taking it to the next level.

My past few blog posts have all been on Oracle and SLOB (Silly Little Oracle Benchmark) create by Kevin Closson. As part of a number of runs with SLOB, I also was curious to benchmark EXT4 and XFS file systems.

EXT4 vs XFS, which file system performs better?

As you can imagine there is not a single and simple answer to these questions because it depends on a number of variables – so I tried to eliminate most of them.

• The same server
• The same VM configuration
• The same Hypervisor
• The same Oracle configuration
• The same SLOB configuration (30min with a 70:30 Read/Write ratio)
• The same LVM configuration
• The same storage configuration (I am using Datrium DVX, which uses localhost SSDs to perpetually read data I/O locally at bus speeds. More info on the exact config here)
  

The only aspect changing between tests is the datadisk filesystem, being either EXT4 or XFS. Here is an image of the disk configuration.

–CentOS/7 (XFS)
|–250 GB
–Oracle U01 (XFS)
|-250 GB
–Oracle U02 (XFS)
|–250 GB
–Oracle Datadisk (LVM)
|-500 GB
|-500 GB
|-500 GB
|-500 GB
|-500 GB
|-500 GB
|-500 GB
|-500 GB

Additionally, considering that file systems can be optimized in countless ways I have accepted all the default configurations for fdisk, LVM, FSTAB and the file systems themselves. That means that this comparison is only valid for the pristine default installations.

--- Volume group ---
   VG Name               vg-01
   System ID             
   Format                lvm2
   Metadata Areas        8
   Metadata Sequence No  2
   VG Access             read/write
   VG Status             resizable
   MAX LV                0
   Cur LV                1
   Open LV               1
   Max PV                0
   Cur PV                8
   Act PV                8
   VG Size               <3.91 TiB
   PE Size               4.00 MiB
   Total PE              1023992
   Alloc PE / Size       1023992 / <3.91 TiB
   Free  PE / Size       0 / 0   
   VG UUID               fC2LdY-IBs9-AfKd-LevI-WbfX-2PCx-VxSZei

--- Logical volume ---
   LV Path                /dev/vg-01/oracle
   LV Name                oracle
   VG Name                vg-01
   LV UUID                o3V0Ca-Qiv0-3tDI-bWsp-2mrA-Btci-zbckEu
   LV Write Access        read/write
   LV Creation host, time oradb02.slab.datrium.com, 2019-01-30 17:53:44 -0800
   LV Status              available
   # open                 1
   LV Size                <3.91 TiB
   Current LE             1023992
   Segments               1
   Allocation             inherit
   Read ahead sectors     auto
 currently set to     16384
 Block device           253:3 

To collect the data I used Chris Buckel article on SLOB Sustained Throughput Test: Interpreting SLOB Results, and I graphed the same way.

However, looking at every data point in a graph doesn't really give us a solid understanding of what is happening because the SLOB workload is dynamic and there are peaks and deeps during a 30 minutes run.

To eliminate outliers I replaced the time series with a trendline that uses a moving average of 10 data points (there are roughly 605 data points for each SLOB run). This approach removed anomalies while maintaining data fidelity.

Conclusion

The conclusion for this Oracle SLOB test that uses 8Kb block size I/O is that XFS performs better than EXT4 under the exact same default configuration conditions - further, XFS is able to better utilize the CPU available to drive performance, due to the parallel I/O based on allocation groups. As always, your mileage may vary :-)