On Monday 25 July 2005 13:15, you wrote:
> > I've ran some simple 'dd' read performance tests on
> > xen-2.0.6-stable, finding quite different results between
> > image backed domains and partition backed domains. The Image
> > backed domains are slightly faster than the partion backed ones.
> Are you sure they were using similar parts of the disk? There can easily
> by a 2x or even 3x performance difference between the outside and inside
> edges of the disk.
Sure, I know. I was using different parts for the partitions and for the
images. But I've also measured the (parrallel) read-throughput of all
available partitions which didn't show a great variance: Avg read time for
1GB with 8 parrallel runs was 339 s, where the longest measurement took 386 s
and the shortest was 314 s.
> > Also I not, that for more than two domains, running dd in
> > parallel, the performance (especially of the image backed
> > domains) drops dramatically below values I've measured using
> > xen-2.0.5-stable.
> > Are you aware of this? Will it be resolved?
> No, we're not aware of any difference between the two, though there were
> some changes to the way requests were batched. It would be good if a few
> people on the list could investigate and produce some detailed
> experiemntal data.
What I did was to run
dd if=$DEV of=/dev/null bs=32k count=32k
For 1, 2, 3, 4 and 8 parrallel instance of
a) My system with an SMP kernel
here $DEV had to point to a different partition for each run, since
otherwise 8 runs took as long as 1, most probably due to caching
b) 220.127.116.11-xen0 image backed
here $DEV was /dev/hda1 used for all domains. The image files we placed
on /dev/hda12 (=20 MB/s according to hdparm)
c) 18.104.22.168-xen0 partition backed
$DEV=/dev/hda1, which were actually backed on /dev/hda[8-11].
I've just run hdparm -tT on all of my partitions, giving me values between 20
and 25 MB/s. According to this: the images appear even slower than they
should be, not faster.
If I'll have enough time, I can run some more benchmarks this week.
Xen-devel mailing list