Completes fix for LP #928910 - libvirt performance

This patch adds the remainder of the recommended fixes
from the original bug report:

* Modifies methods in the compute manager that relied on
  the DB power state to be in sync with the virt driver to
  instead just query the power state of the instance from the
  virt driver. This enables us to set the periodic tick to 10
  for the problematic compute.manager.Manager._sync_power_states()
  method.
* Modifies the _sync_power_states method in the following ways:
 ** Replace the call to driver.list_instances_detail() to a new,
    driver-overrideable get_num_instances() call
 ** For each instance known by the database, call driver.get_info()
    separately inside the loop instead of calling the expensive
    list_instances_detail() method that can take a very long time
    to complete on hosts with lots of instances
 ** Call greenthread.sleep(0) before each call to update the
    database power state, enabling other periodic tasks to do work

Once again, I left an inefficient default implementation of the
new driver.get_num_instances() method in the base driver class. I
need help from folks who understand the Xen/VMWare drivers to do
an override for get_num_instances() in those drivers that calls
the underlying XenAPI or VMWare API.

Change-Id: I88002689cdda32124423da320f8c542e286be51b

0 files changed, 0 insertions, 0 deletions