More File-System Tests Of The Linux 3.18 Kernel

[phoronix]

Phoronix: More File-System Tests Of The Linux 3.18 Kernel

Earlier this week on Phoronix I posted benchmarks indicating potential block/file-system performance regressions using the Linux 3.18 kernel. Since then I've been carrying out more tests looking for any file-system performance problems on other hardware...

http://www.phoronix.com/vr.php?view=MTg0NTc

[Fracta1L]

Omg, so terrible results for Btrfs...

[WorBlux]

Omg, so terrible results for Btrfs...

Not really, Looking at the same tests, it does random reads just as fast as the othere, the problem being that it doesn't really gain any speed then doing sequential reads. If you dig into the results deeper you see it won the postMark test though a slight performance regression did show up in the git pull.

[sdack]

Omg, so terrible results for Btrfs...

Btrfs' performance is usually on par with EXT and XFS. I suspect there is an issue with the benchmark itself or how the file systems have been set up for the tests.

Honestly, I am a bit worried for Phoronix here. When one wants to talk about a regression between kernel versions and the regression shows to be in the range of about 10% then why would you use a benchmark for testing where the tested file systems show an unusual difference to one another in the range of 400% to 2000%?

It makes any discussion on a regression as absurd as wanting to discuss positions of the Kama Sutra by using whales instead of humans.

[droste]

[...]Honestly, I am a bit worried for Phoronix here. When one wants to talk about a regression between kernel versions and the regression shows to be in the range of about 10% then why would you use a benchmark for testing where the tested file systems show an unusual difference to one another in the range of 400% to 2000%?[...]

Did you actually have a look at the graphs? The fs with a regression is ext4 not btrfs and ext4 seems to be in the same range as XFS.

[sdack]

Did you actually have a look at the graphs? The fs with a regression is ext4 not btrfs and ext4 seems to be in the same range as XFS.

The much better question here is: did you actually look at the performance difference between Btrfs, EXT and XFS?

Whatever regression you may think you see should the difference between the file systems be of a much greater concern to you and should be examined before you start talking about a regression when using that benchmark.

If you cannot see this or do not understand it then there is no hope in discussing this with you. Your view would then be too narrow to grasp the whole picture.

[droste]

Are you suggesting that we should ignore that ext4 lost 10% performance just because btrfs is generally slow?
You realize that these are somewhat separate products developed by a separate set of people.
Yes btrfs is slow and something should be done (and this is not news!). Why can't something be done in parallel to check what change in ext4 did this so that it does not slowly adjust its speed to the current speed of btrfs. If ext4 and xfs would be as slow as btrfs everything is ok, right? Because now they have no difference in speed!

The regression is the 10% speed we lost in ext4. The tests of the other file systems are in there to see if this is a common speed loss for all file systems or just ext4. So what matters here is the relative difference between the kernel version for each file system and not the absolute difference of the file systems.

[sdack]

Are you suggesting that we should ignore that ext4 lost 10% performance just because btrfs is generally slow?

Do you seriously believe Btrfs to be 4x-20x times slower than Ext and XFS?

[tehehe]

Do you seriously believe Btrfs to be 4x-20x times slower than Ext and XFS?

I agree with sdack. I would like to know why btrfs is SO MUCH slower than other file systems in this test.

[droste]

Do you seriously believe Btrfs to be 4x-20x times slower than Ext and XFS?

No (to some extend because I use it as my file system for my root drive), at least not with use cases that happen in reality. I also don't believe that it writes 10x times as fast. If you check all graphs you see that btrfs in this system writes 10x as fast for random writes.

http://openbenchmarking.org/result/1411189-LI-FSLINUX3180

That's the reason why I would not look at the absolute values of this test! But the weirdness of this test case / test scenario does not explain why it performance changes between kernel versions.