2

I did 4 clean installations of debian 6 and measured time of installing some average virtual package. FS options are default.

time apt-get install build-essential

I got very strange results (min:sec, less is faster):

XFS:   3:12
btrfs: 2:45
ext3:  0:30
ext4:  0:50

What is wrong with XFS and btrfs? 6 times slower than ext3? Am I doing something wrong?

enter image description here

Upd (some details):

All LVM volumes are local to VM and are on idle RAID. CDROM image is local and the same, Internet connection is stable and it factors max 10-15 seconds. All the visible slowdowns are after downloads: XFS and btrfs guests think over 1 second every Unpacking. Low-level caching is disabled. Host node is idle every installation, no active guests but the one.

Michael Hampton
  • 237,123
  • 42
  • 477
  • 940
ujifgc
  • 186
  • 10
  • 2
    I see you are using VMs, how do you factor out contention with other VMs? Are you sure you didn't have something else running at the same time as your XFS test? Did you pre-download, have a local APT cache, or a CD-ROM image so you can factor out any issues with retrieving packages from the random mirror you would get from the Internet? – Zoredache May 23 '12 at 19:34
  • I'd add to this that some amount of lower level caching would likely take place unless the storage is unmounted and remounted between each individual run... – rnxrx May 23 '12 at 19:55
  • I fail to notice any factor that could drop performance even 2 times. All LVM volumes are local to VM and are on idle RAID. CDROM image is local and the same, Internet connection is stable and it factors max 10-15 seconds. All the visible slowdowns are after downloads: XFS and btrfs guests think over 1 second every `Unpacking`. Low-level caching is disabled. I decided to run these tests because I was unhappy with package manager performance on metal host. – ujifgc May 24 '12 at 05:44
  • @ujifgc: Unhappy with package manager performance? apt too slow? Try installing bigger stuff with Gentoo's Portage or *BSD ports and talk about waiting ... :-) that said, mount options such as `barrier=0` might help with XFS -- ext3 does not enable that by default, unless things have changed recently. – Janne Pikkarainen May 24 '12 at 08:46
  • 1
    Nope, `barrier=0` mount option doesn't affect performance. I got 3:13 of pain while installing build-essential on XFS after editing `fstab` and rebooting. `mtab` shows the option ok. – ujifgc May 24 '12 at 15:35
  • Can you provide other performance metrics about the four different systems with perhaps a comparison report on Openbenchmark.org ? – Stefan Lasiewski Jun 11 '12 at 19:13
  • You mean synthetic benchmarks? I'm sure the new systems will rock on that. – ujifgc Jun 12 '12 at 06:18
  • Well, I am just referring to benchmark software such as the tests provided by the [Phoronix Test Suite](http://www.phoronix-test-suite.com/), which provides both real and synthetic tests. It pretty popular. You can compare your system against other systems on www.openbenchmarking.org, and invite others to run the same test and compare the results against your own. – Stefan Lasiewski Jun 16 '12 at 19:18
  • 1
    FYI, my own tests also show a huge discrepancy between btrfs and ext4 (running apt on ext4 vs btrfs; I did not test ext3 or xfs). – Stefan Lasiewski Jun 16 '12 at 19:19

1 Answers1

1

Which kernel version do you use? If I read this correct, something still 2.6.X, right, so XFS@2009-2010?

XFS has been known to have a serious bottleneck in its log writing mechanism. This is something which has finally been solved in kernel version 3.3 once and for all and the solution is enabled there by default.

So you should update your system's kernel and then you should be on the safe side.

Marc Stürmer
  • 1,894
  • 12
  • 15