[Xen-users] cpu utilizations of xentop inflated in comparison to mpstat

[tim benke <tbenke@xxxxxxxxxxxx>]

Hi all,

when comparing the utilizations in Dom0 reported by xentop and mpstat I experience extremely high CPU utilizations reported by xentop when performing hard disk I/O. In the experiment of copying a 1GB in a virtual machine backed by a tap:aio: backend we measured the CPU utilizations in Dom0 without considering the first and last 10 seconds. The experiment was repeated with 2 VMs.

mpstat values are user+sys

#VMs runtime utilization(xentop) utilization(mpstat)
1 44s 13%   0.01+0.35%
2 110s 8%   0.015+0.14%

It seems like xentop is reporting extremely high CPU utilizations in comparison to mpstat. Why is that so and can I validate one of the two values?
One other interesting thing I cannot explain is the fact that the CPU utilization is going down for 2 VMs, although the request rate should be identical. Why does the CPU utilization decrease when the amount of work being done is twice the original amount.


The results repeatable and the variance for xentop is about 1% utilization. The machine is running Xen 3.1 has 4 physical CPUs. The default VCPU allocation was used i.e., 4 VCPUs for Dom0 and one for each DomU.

There's a performance report on Xen 3.0 by Fujitsu Siemens under (http://docs.fujitsu-siemens.com/dl.aspx?id=7d0f32df-d32c-4ec9-b4a1-aab797689882) that says on page 8:

With the OS version used the following problem occurs in the Dom0 :
The »xentop« command reports dubious values in the CPU load of the Dom0 as a result of the I/O
activities. These values do not correlate with the values of commands, such as »mpstat«, namely to
the extent that the xentop values are in comparison considerably excessive. No decisive clarification
could be found as to which command is supplying authentic values. Analyses, which were performed
to clarify the problems by means of network load between the SLES10 system and a comparable
Windows counterpart, resulted in values for the Windows system that are in fact close to the
»mpstat« values. However, on account of the quite different architecture of both operating systems
this can only be an indication. As a consequence of this problem it is not possible in this document to
provide any details about the effective system load during a performance measurement.

..Tim Benke

_______________________________________________
Xen-users mailing list
Xen-users@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-users