Hi,
Encountering some difficulties to exploit SAR results in a virtualized
environment, I need some precisions about measures given by the command
"sar
-u" (sysstat version : 8.1.8).
I hope you will have time to answer my few questions.
Hereafter are usefull information to present the problem.
Context
-------
My test run on a system virtualized with Xen 3.0.1.
5 OS are installed on this system : Dom0 and 4 identical virtual machines.
The virtual CPU of each VM is caped to 25% of the physical CPU. Dom0 has no
CPU limitation.
The command "sar -u" is executed on each OS during the test.
(1)Average values given by the test are :
user nice system iowait steal idle
Dom0 0,02 0 0,58 18,82 1,57 79,01
vm1 14,93 0 2,18 13,43 33,38 36,08
vm2 0,01 0 0,02 5,43 0,63 93,9
vm3 0,01 0 0,02 2,64 0,64 96,7
vm4 0,01 0 0,01 4,22 0,14 95,63
(2)Next values are an extract from vm1 results :
user nice system iowait steal idle
...
13,47 0,00 0,00 11,72 55,86 18,95
51,85 0,00 1,23 7,41 36,79 2,72
29,46 0,00 0,33 15,55 40,59 14,08
...
Questions
---------
- Is the CPU load of a virtual machine equal to the sum : %user + %system +
%iowait ? Otherwise which sum could be representative ?
- As the CPU allocated to each VM is 25% of the physical CPU, is the SAR
measure related to 25% of the physical CPU or to 100% of the physical CPU ?
- Inside a VM, is it normal to get measures overriding the fixed CPU limit
(25%) (cf. (2))?
- Does the "iowait" measure correspond to a process wait for hard
drive IO
response ?
- Is the CPU idle during "iowait" (ie : available for other uses) ?
- What does "steal" measure mean ? Does it figure :
* a physical CPU ressource needed by the VM beyond the 25% threshold and
refused to the VM ?
* a fraction of the 25% physical CPU ressource allocated to the VM and
which is not disposable for the VM processes ?
Despite our researches (manual, sites, forums), we didn''t find a
complete
answer. I''d be grateful if you could provide me some.
Thanks.
--
View this message in context:
http://www.nabble.com/Need-for-precisions-on-SAR-measures-tp22315119p22315119.html
Sent from the Xen - User mailing list archive at Nabble.com.
_______________________________________________
Xen-users mailing list
Xen-users@lists.xensource.com
http://lists.xensource.com/xen-users
some points to remember: - sar doesn''t know a bit about virtualization, therefore all its measures are relative to the resources available to that VM - Dom0 is also a VM managed by the underlying Xen hypervisor, so DomU''s CPU time is used ''from outside'' Dom0, it doesn''t appear in Dom0''s sar reports. - all IO pass through Dom0 userspace (except for PCI passthrough), so DomU''s IO does appear in Dom0''s sar reports. -- Javier _______________________________________________ Xen-users mailing list Xen-users@lists.xensource.com http://lists.xensource.com/xen-users
Hi, Javier Guerra Giraldez wrote:> > - sar doesn''t know a bit about virtualization, therefore all its > measures are relative to the resources available to that VM >A question concerning this point : If all the measures are relative to the resources available, how can we explain that : On the non virtualized platform, the consumption with a process is : user nice system iowait idle 43,54 0 8,56 6,7 41,12 On the virtualized platform without CPU cap by xen (physical CPU accessible for every domain), the consumption with the same process is : user nice system iowait steal idle Dom0 0,01 0 0,97 26,44 1,97 70,61 vm1 21,28 0 8,91 23,48 0,5 45,83 (activity only on vm1) vm2 0,01 0 0,1 8,33 0,84 90,81 vm3 0 0 0,01 4,82 0,74 94,34 vm4 0 0 0,01 3,75 0,31 95,92 On the virtualized platform with CPU caped by xen (25% physical CPU accessible for each domain, except Dom0 which has full access), the consumption with the same process is : user nice system iowait steal idle Dom0 0,02 0 0,58 18,82 1,57 79,01 vm1 14,93 0 2,18 13,43 33,38 36,08 (activity only on vm1) vm2 0,01 0 0,02 5,43 0,63 93,9 vm3 0,01 0 0,02 2,64 0,64 96,7 vm4 0,01 0 0,01 4,22 0,14 95,63 So, given these results, caping the CPU for each domain at 25% of the physical CPU has globaly : - decrased the user, system and iowait consumptions - incrased the steal consumption - incrased a bit the user + system + iowait + steal sum If all the measures are relative to the resources available, couldn''t we expect, as fewer CPU ressources are given to vm1, an increase of the CPU activity measured on vm1 (at least user and system) ? -- View this message in context: http://www.nabble.com/Need-for-precisions-on-SAR-measures-tp22315119p22334827.html Sent from the Xen - User mailing list archive at Nabble.com. _______________________________________________ Xen-users mailing list Xen-users@lists.xensource.com http://lists.xensource.com/xen-users
On Wed, Mar 4, 2009 at 12:20 PM, Nicolas80 <dedestri@hotmail.com> wrote:> > Hi, > > > Javier Guerra Giraldez wrote: >> >> - sar doesn''t know a bit about virtualization, therefore all its >> measures are relative to the resources available to that VM >> > > A question concerning this point : > If all the measures are relative to the resources available, how can we > explain that :i stand corrected. sar does know a small bit about paravirtualization: the steal time is time that the (real) CPU is used for other VMs. there''s a very readable explanation at http://virt.kernelnewbies.org/ParavirtBenefits -- Javier _______________________________________________ Xen-users mailing list Xen-users@lists.xensource.com http://lists.xensource.com/xen-users