Hello.
I have been doing some performance tests with Xen and CentOS and have found some strange statistics when comparing the virtual memory (/proc/vmstat) and I/O (/proc/diskstat) statistics between the Domain 0 and guests.
When I look at the "pgpgin" statistic it is exactly the same as the "read sectors" in Dom0. However, when I look at the same statistics in the guest, the "read sectors" is exactly 2x "pgpgin".
I have tried this on 2 different versions, each on a different physical server. Here is an example of the statistics over a 30 second period (while running a benchmark in Dom1) in the Dom0 and Guest Domain. pgpgin read sectors ======= ========= Dom0 1698024 1698024 Dom1 848412 1696824
I have looked at some various settings to see if I could explain it. I can't tell if the Dom1 guest kernel is reporting incorrectly, or there is some other explanation. If anyone has any ideas about this I would appreciate any feedback.
Dom 0 kernels 3.10.29-11.el6.centos.alt.x86_64 3.4.54-8.el6.centos.alt.x86_64
Guest kernel 2.6.32-431.el6.x86_64
Thanks,
Deron
On Sun, 6 Apr 2014, Deron wrote:
Hello.
I have been doing some performance tests with Xen and CentOS and have found some strange statistics when comparing the virtual memory (/proc/vmstat) and I/O (/proc/diskstat) statistics between the Domain 0 and guests.
When I look at the "pgpgin" statistic it is exactly the same as the "read sectors" in Dom0. However, when I look at the same statistics in the guest, the "read sectors" is exactly 2x "pgpgin".
I have tried this on 2 different versions, each on a different physical server. Here is an example of the statistics over a 30 second period (while running a benchmark in Dom1) in the Dom0 and Guest Domain. pgpgin read sectors ======= ========= Dom0 1698024 1698024 Dom1 848412 1696824
I have looked at some various settings to see if I could explain it. I can't tell if the Dom1 guest kernel is reporting incorrectly, or there is some other explanation. If anyone has any ideas about this I would appreciate any feedback.
Dom 0 kernels 3.10.29-11.el6.centos.alt.x86_64 3.4.54-8.el6.centos.alt.x86_64
Guest kernel 2.6.32-431.el6.x86_64
The numbers that you are seeing are probably wrong due to an accounting issue in blkfront or blkback. It would be interesting to know whether this happens even with the most recent kernels.
I think I know why this is happening, and I think it is correct. I ran a test again only looking at Dom0 and found that it was similar to Dom1. The 'read sectors' from /proc/diskstats were about 2x the pgpgin from /proc/vmstat. This is whay I see on my physical hardware as well.
The reason that I am not seeing the same results when reading data from Dom1 is because I am only looking at the physical block device (/dev/sda in this case). When I look at the diskstats from physical device /dev/sda and device between used for blkback /dev/tda they show the 2 reads. Since they are both "block devices" I assume they need to page in to make the data available.
So I think this is not really an issue, just a matter of perspective.
Deron
On Wed, Apr 9, 2014 at 4:15 AM, Stefano Stabellini < stefano.stabellini@eu.citrix.com> wrote:
On Sun, 6 Apr 2014, Deron wrote:
Hello.
I have been doing some performance tests with Xen and CentOS and have
found some strange statistics when comparing the
virtual memory (/proc/vmstat) and I/O (/proc/diskstat) statistics
between the Domain 0 and guests.
When I look at the "pgpgin" statistic it is exactly the same as the
"read sectors" in Dom0. However, when I look at
the same statistics in the guest, the "read sectors" is exactly 2x
"pgpgin".
I have tried this on 2 different versions, each on a different physical
server. Here is an example of the statistics
over a 30 second period (while running a benchmark in Dom1) in the Dom0
and Guest Domain.
pgpgin read sectors ======= ========= Dom0 1698024 1698024 Dom1 848412 1696824I have looked at some various settings to see if I could explain it. I
can't tell if the Dom1 guest kernel is
reporting incorrectly, or there is some other explanation. If anyone
has any ideas about this I would appreciate any
feedback.
Dom 0 kernels 3.10.29-11.el6.centos.alt.x86_64 3.4.54-8.el6.centos.alt.x86_64
Guest kernel 2.6.32-431.el6.x86_64
The numbers that you are seeing are probably wrong due to an accounting issue in blkfront or blkback. It would be interesting to know whether this happens even with the most recent kernels. _______________________________________________ CentOS-virt mailing list CentOS-virt@centos.org http://lists.centos.org/mailman/listinfo/centos-virt
-
Deron -
Stefano Stabellini