Feature pages are design documents that developers have created while collaborating on oVirt.
Most of them are outdated, but provide historical design context.
They are not user documentation and should not be treated as such.
Documentation is available here.
Most of them are outdated, but provide historical design context.
They are not user documentation and should not be treated as such.
Documentation is available here.
Host Power Management Policy
Bug-Url: https://bugzilla.redhat.com/show_bug.cgi?id=1035238
Owner
- Martin Sivak msivak@redhat.com
Status
- Basic implementation in gerrit verified to work with the nice scenarios:
Goals
- shutdown hosts when Powersaving is selected and engine clears the host (migrates all VMs elsewhere)
- wake up a host when the available resources get below configured level
Design
- shutdown methods:
- graceful shutdown - SSH, IPMI (if it supports that)
- standard engine’s fencing methods - IPMI, Drac, …
- wake methods:
- standard fencing methods - IPMI, SSH, … - with fallback to WOL when needed and supported
- use Start/StopVdsCommand internally
Shutdown rules
- decided as part of load balancing thread
- there is enough available hosts in the cluster even without this host
- make sure spm is not killed
Wake Up rule
- decided as part of load balancing thread
- not enough free hosts in the cluster
Shutdown procedure
- qualifying Hosts (powerManagementControlledByPolicy==true) in the Up state (one at a time) will be moved to maintenance and powerManagementControlledByPolicy kept as true by the load balancer
- qualifying Hosts in the Maintenance state that also have powerManagementControlledByPolicy flag set will be shut down
Wake up procedure
- when the load balancer decides to wake up a host or the user clicks the Activate button
- engine will start the host and set powerManagementControlledByPolicy to true flag when the startup is finished
DB changes
- new boolean flag for Host - powerManagementControlledByPolicy - meaning the host can be controlled by the automatic policy, will be cleared by user triggered PM actions and set every time the host goes Up
- new boolean flag for Host - disableAutomaticPowerManagement - that will make this host invisible to the power management policy
new Power saving policy attributes
- Cluster policy needs to allow the user to set:
- the minimum amount of free hosts
- whether host shutdown is allowed
- (how long does a host need to be empty to be considered for shutdown)
UI
- Each host needs a boolean that will override the shutdown (persistent host)
Known issues
- when user tries to start a Vm that is pinned to a powered down host, we will fail the start
- if there is no spare available we might need to fail RunVm and inform the user to try again once we start more hosts
- if we initiate host shutdown and it takes long time, we might try to start it up again - we need a way of distinguishing initiated shutdown and real down state
- if we are powering up a host and it fails, it is possible we will select the same host for the next attempt
- power operations might need to be executed in a separate thread to not block the balancer operations
Possible future enhancements
- Reserve defined in terms of power, memory, …
- useful for reserving space for fast start of some important VM
- bogomips used as the CPU power unit (/proc/cpuinfo)
- Wake On Lan fencing method:
- each host has to report MAC and WOL capability for each NIC
- locality information to know who can send the proper WOL packet (each host will report it’s ARP database of visible MAC addresses for each NIC)