Libvirt: add perf event support

https://blueprints.launchpad.net/nova/+spec/support-perf-event

The purpose of this blueprint is to add nova an ability to support perf event to gain statistic (for example cpu cache usage) for each instance. These perf event data will be collected by Ceilometer. [1]

Problem description

Perf event is a Linux feature that provides a framework for analyzing performance events at both hardware and software levels. Through a list of measurable events we can measure events coming from different resources (like context-switches, cache misses, etc.) and gain statistic for each instance.

Perf has integrated to libvirt from 1.3.3 and it now supports to gain cpu cache and more event type will be added. We can enable perf support in Nova.

Use Cases

As a cloud operator, he/she wants to know if instances in a cloud occupy what kinds of resources, for example, cpu, memory, cpu cache, memory bandwidth etc., and also the amount of resource of the instance. These kinds of monitor data can be collected from Ceilometer. With these monitor data, the operatior can do some analysis to identify what is the most important resource for this instance and he/she can do further operations like migrate to some other hosts to provide better resources to meet customers SLA.

The Ceilometer spec requires to nova have perf support.[1]_

Proposed change

Add new libvirt driver list configure option enabled_perf_events, which is a list to indicate the perf event type, default is [].

Add missing elements when generating XML definition in libvirt driver to support perf event per the configuration of enabled_perf_events. Only supported event with proper Libvirt version, it will be ignored if the version of Libvirt is too old.

For example we have enabled_perf_events=[‘cmt’], the XML element will be like this:

<perf>
  <event enable="yes" name="cmt"/>
</perf>

Libvirt requires this flag in it’s XML to initialize a file descriptor before we gain the statistic, the polling won’t be started until we call Libvirt API.

In this spec, we don’t propose nova to polling statistic data itself, Ceilometer can benefit from this configuration.

Alternatives

Another solution is enable perf event per instance by using flavor’s extra_spec, like adding ‘perf:event=cmt,…’, add aggregate to host. It’s complex to enable it from operators.

The reason not using flavor’s extra_spec is this is not a spec user should take care, it’s not a feature of VM, but things the platform can provide us, so if platform can provide us such feature, we can benifit from it. It doesn’t make sense that a user wants an instance schedule to a host which can have performance monitor.

Data model impact

None

REST API impact

None

Security impact

None

Notifications impact

None

Other end user impact

None

Performance Impact

There won’t be performance impact. When perf event is enabled, the operation is just to write the count into the memory, and the impact can be almost ignored, especially we just enable events for VMs (not for each processes).

No addition API call is requires at all.

Other deployer impact

None

Developer impact

None

Implementation

Assignee(s)

Primary assignee:

Eli Qiao <liyong.qiao@intel.com>

Other contributors:

qiaowei-ren <qiaowei.ren@intel.com>

Work Items

The primary work items are:

  • Add new libvirt driver configuration option.

  • If the version of libvirt is new enough to support the flags in the xml, update the libvirt guest XML configuration when one or more perf events are specified in libvirt driver configuration.

Dependencies

And this spec will depend on the following libraries:

  • libvirt >= 1.3.3

Testing

  • Add unit test case to verify guest XML has been updated correctly.

Documentation Impact

  • Add explanation of new added libvirt configuration.

References

History

Revisions

Release Name

Description

Newton

introduced.