nova-specs/specs/stein/approved/support-hpet-on-guest.rst
Eric Fried 3906a4f7c6 High Precision Event Timer (HPET) on x86 guests
Support HPET for x86 guests in the libvirt driver.

blueprint: support-hpet-on-guest
Co-Authored-By: Chris Friesen <chris.friesen@windriver.com>
Change-Id: I26c5f5386b99f88dcc7f6ea4629a61f93084b3e7
Signed-off-by: Jack Ding <jack.ding@windriver.com>
2018-11-01 16:27:33 -04:00

5.9 KiB

Support High Precision Event Timer (HPET) on x86 guests

https://blueprints.launchpad.net/nova/+spec/support-hpet-on-guest

Problem description

Use Cases

As an end user looking to migrate an existing appliance to run in a cloud environment I would like to be able to request a guest with HPET so that I can share common code between my virtualized and physical products.

As an operator I would like to support onboarding legacy VNFs for my telco customers where a guest image cannot be modified to work without a HPET.

Proposed change

End users can indicate their desire to have HPET in the guest by specifying a image property hw_time_hpet=True.

Setting the new image property to "True" would only be guaranteed to be valid in combination with hypervisor_type=qemu and either architecture=i686 or architecture=x86_64.

Note

A corresponding flavor extra spec will not be introduced since enabling HPET is really a per-image concern rather than a resource concern for capacity planning.

A few options to use Traits were considered as described in the next section, but we end up choosing the simpler approach due to the following reasons:

  1. HPET is provided by qemu via emulation, so there are no security implications as there are already better clock sources available.
  2. The HPET was turned off by default purely because of issues with time drifting on Windows guests. (See nova commit ba3fd16605.)
  3. The emulated HPET device is unconditionally available on all versions of libvirt/qemu supported by OpenStack.
  4. The HPET device is only supported for x86 architectures, so in a cloud with a mix of architectures the image would have to be specific to ensure the instance is scheduled on an x86 host.
  5. Initially we would only support enabling HPET on qemu. Specifying the hypervisor type will ensure the instance is scheduled on a host using the qemu hypervisor. It would be possible to extend this to other hypervisors as well if applicable (vmware supports the ability to enable/disable HPET, I think), and which ones are supported could be documented in the "useful image properties" documentation.

Alternatives

The following options to use Trait were considered, but ultimatedly we chose a simpler approach without using Trait.

Explicit Trait, Implicit Config

Operators can indicate their desire to have HPET in the guest by specifying a placement trait trait:COMPUTE_TIME_HPET=required in the flavor extra-specs.

End users can indicate their desire to have HPET in the guest by uploading their own images with the same trait.

Existing nova scheduler code picks up the trait and passes it to GET /allocation_candidates.

Once scheduled to a compute node, the virt driver looks for trait:COMPUTE_TIME_HPET=required in the flavor/image or trait*:COMPUTE_TIME_HPET=required for numbered request group in flavor and uses that as its cue to enable HPET on the guest.

If we do get down to the virt driver and the trait is set, and the driver for whatever reason (e.g. value(s) wrong in the flavor; wind up on a host that doesn't support HPET etc.) determines it's not capable of flipping the switch, it should fail.1

CON: We're using a trait to effect guest configuration.

Explicit Config, Implicit Trait

  • Operator specifies extra spec hw:hpet=True in the flavor.
  • Nova recognizes this as a known special case and adds required=COMPUTE_TIME_HPET to the GET /allocation_candidates query.
  • The driver uses the hw:hpet=True extra spec as its cue to enable HPET on the guest.

CON: The implicit transformation of a special extra spec into placement-isms is arcane. This wouldn't be the only instance of this; we would need to organize the "special" extra specs in the code for maintainability, and document them thoroughly.

Explicit Config, Explicit Trait

  • Operator specifies both extra specs, hw:hpet=True and trait:COMPUTE_TIME_HPET=required, in the flavor.
  • Existing nova scheduler code picks up the latter and passes it to GET /allocation_candidates.
  • The driver uses the hw:hpet=True extra spec as its cue to enable HPET on the guest.

CON: The operator has to remember to set both extra specs, which is kind of gross UX. (If she forgets hw:hpet=True, she ends up with HPET off; if she forgets trait:COMPUTE_TIME_HPET=required, she can end up with late-failing NoValidHosts.)

Data model impact

None

REST API impact

None

Security impact

None

Notifications impact

None

Other end user impact

None

Performance Impact

Negligible.

Other deployer impact

None

Developer impact

None

Upgrade impact

The new image property will only work reliably after all nodes have been upgraded.

Implementation

Assignee(s)

Primary assignee:

jackding

Other contributors:

jaypipes, efried

Work Items

  • libvirt driver changes to support HPET

Dependencies

None

Testing

Will add unit tests.

Documentation Impact

Update User Documentation for image properties2.

References

History

Revisions
Release Name Description
Stein Introduced

  1. http://lists.openstack.org/pipermail/openstack-dev/2018-October/135446.html↩︎

  2. https://docs.openstack.org/glance/latest/admin/useful-image-properties.html↩︎