Overhead option to differ hypervisor process on a global set of pCPUs
WHAT?
Introducing 'overhead_pin_set' option for compute nodes. Some
workloads require to have the hypervisor overhead processes
running on on an isolated place which do not create lateny spikes
for vCPUs threads running.
For libvirt driver configuring the 'overhead_pin_set' and
configuring the flavor with hw:emulator_
result that all the threads to be pinned on the isolated pCPUs.
WHY?
The currenlty implemented emulator threads placement policy
provide an isolated option which reserved an additionnal host CPU
per guest to handle hypervisor overhead. For some workloads that
is consuming too much host resources.
Blueprint information
- Status:
- Complete
- Approver:
- Sylvain Bauza
- Priority:
- Low
- Drafter:
- Sahid Orentino
- Direction:
- Approved
- Assignee:
- Sahid Orentino
- Definition:
- Approved
- Series goal:
- Accepted for rocky
- Implementation:
- Implemented
- Milestone target:
- rocky-2
- Started by
- melanie witt
- Completed by
- melanie witt
Related branches
Related bugs
Sprints
Whiteboard
Gerrit topic: https:/
Addressed by: https:/
libvirt: set emulator threads on isolated host cpus
Addressed by: https:/
update: introducing isolate emulthreads on host
Addressed by: https:/
libvirt: rework _get_vpcu_total to return online pCPUs
Approved for Rocky. -- mriedem 20180409
It has been indicated in the approved spec that the option introduced will not be 'CONF.overhead_
Also instead of introducing a new policy HOST for hw:emulator_
Addressed by: https:/
compute: introduce cpu_share_set
Addressed by: https:/
libvirt: pass emulator threads policy insteadwhether it's isolated
Addressed by: https:/
libvirt: update emulator threads docstring
Gerrit topic: https:/
Addressed by: https:/
fix blueprint related to the spec overhead-pin-set