VFS accounting performance benchmarks

[Dmitrij D. Czarkoff]

On Tue, Dec 20, 2011 at 10:07:22AM +0100, Francois Tigeot wrote:
> Most of the VFS accounting work is now in the tree (and disabled by default,
> enable it with vfs.accounting_enabled="1" in /boot/loader.conf).
> Only a potentially disruptive commit to enable it for / has not been pushed.
> 
> Since VFS accounting uses a per mount-point spinlock to protect its counters, 
> there is some concern it will degrade write performance on SMP machines.
> 
> To measure if this was the case or not, I've run some blogbench benchmarks on
> a small Xeon system; the details of the test and the results are in the
> attached PDF.
> 
> VFS accounting kernel was from 2011-12-19
> Kernel without VFS accounting was from 2011-11-19
> 
> For my part I think the results are inconclusive: either there's no measurable
> performance difference or the choice of benchmark was bad.

I've made a plot which shows the results in more evident way (if I correctly
understand the test results data).

-- 
Dmitrij D. Czarkoff
Attachment:
vfsa.svg
Description: image/svg
Attachment:
-------------- next part --------------
A non-text attachment was scrubbed...
Name: png00000.png
Type: application/octet-stream
Size: 32924 bytes
Desc: "Description: PNG image"
URL: <http://lists.dragonflybsd.org/pipermail/kernel/attachments/20111220/8082821b/attachment-0020.obj>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: bin00000.bin
Type: application/octet-stream
Size: 65813 bytes
Desc: "vfsa.png"
URL: <http://lists.dragonflybsd.org/pipermail/kernel/attachments/20111220/8082821b/attachment-0020.bin>