8dd2140c4c
In order to do more sophisticated scheduling (e.g. schedule based on volume
type), filter scheduler is introduced. Several changes are made to make this
possible, some of them are similar to the counterpart in Nova:
- add HostState class to host_manager in order to store volume capabilities
- implement get_volume_stats() method of iSCSIDriver as an example to
demonstrate how volume backend driver reports capabilities as well as status
- add scheduler_options.py and 'scheduler_json_config_location' flag to be
allow loading json configuration file for scheduler at run time
- port common filters/weights from oslo
- add capacity weigher (scheduler/weights/capacity.py) for picking up
target volume backend by weighing free capacity of the host. The default
behavior is to spread volumes across hosts; by changing the
'capacity_weight_multiplier' to negative number, volume placing behavior will
become stacking.
- add capacity filter which filters those hosts have insufficient storage space
to serve the request.
- add 'reserved_percentage' config option to indicate how much space is
reserved. Capacity reservation is needed when volume resize is enabled.
- change default scheduler to 'filter.FilterScheduler'.
- add 'publish_service_capabilities()' method to volume RPC API to allow
scheduler to have volume service report capabilities. This bumps volume RPC
API to version 1.2
- remove 'volume_force_update_capabilities' config option, volume status will be
report to scheduler in every checking.
The implication of this change to storage/backend driver developer:
- implementation of get_volume_stats() of the driver is now a *MUST*, filter
scheduler heavily relies on the status/capabilities reported by backend driver
to makeplacement decision. To ensure Cinder works seamlessly on the storage
system, driver should at least report following capabilities/status:
----------------------+---------------------------+---------------------------
Capability/Status | Description | Example
----------------------+---------------------------+---------------------------
'volume_backend_name'| back-end name, string | 'Example_Storage_Backend'
----------------------+---------------------------+---------------------------
'vendor_name' | vendor name, string | 'OpenStackCinder'
----------------------+---------------------------+---------------------------
'driver_version' | version, string | '1.0a'
----------------------+---------------------------+---------------------------
'storage_protocol' | supported protocols, | 'iSCSI', 'RBD', 'FC', 'NFS'
| string or list of strings | ['iSCSI', 'NFS', 'FC']
----------------------+---------------------------+---------------------------
'total_capacity_gb' | capacity in GB, integer | 102400
----------------------+---------------------------+---------------------------
'free_capacity_gb' | available capacity in GB, | 1000
| integer |
----------------------+---------------------------+---------------------------
'reserved_percentage' | reserved space in | 0, 10
| percentage, integer |
----------------------+---------------------------+---------------------------
The implication of this change to Cinder administrator:
- the default setting for filter scheduler should work well with the benefits
of:
* being able to fully utilize capacity of backends (driver now has to report
actul total space and space utilization and scheduler uses these info) not
limited by the 'max_gigabytes' config option any more;
* being able to choose placement policy between spreading & stacking for
volume creation (by modifying the 'capacity_weight_multiplier' in
CapacityWeigher)
- with filter scheduler, Cinder is now able to utilize the advanced features/
capabilities provided by different storage back-ends via: defining different
volume types with proper extra_specs. Volume types can be considered as sets
of back-end capabilities requirement.
For example, a volume type which has 'storage_protocol':'FC' key/value pair
definition in its extra_spec can only be served by those back-ends who report
they support FiberChannel protocol. Another example is volume type has 'QoS'
requirement can only be served by back-ends support QoS.
Note/TODO:
* Currently scheduler makes its decision based on the status and capabilities
information reported by volume nodes, and these information is stored in memory
of scheduler process. More sophisticated way may be add on table in DB to
record status/capabilities of all volume nodes, like Nova does for compute nodes.
implement bp volume-type-scheduler
DocImpact
Change-Id: I8b5305bce3b24bcc85365037356a8f36d0107133
The Choose Your Own Adventure README for Cinder
You have come across a storage service for an open cloud computing service. It has identified itself as "Cinder." It was abstracted from the Nova project.
To monitor it from a distance: follow @openstack on twitter.
To tame it for use in your own cloud: read http://docs.openstack.org
To study its anatomy: read http://cinder.openstack.org
To dissect it in detail: visit http://github.com/openstack/cinder
To taunt it with its weaknesses: use http://bugs.launchpad.net/cinder
To watch it: http://jenkins.openstack.org
To hack at it: read HACKING
To cry over its pylint problems: http://jenkins.openstack.org/job/cinder-pylint/violations