The Openstack Load Leveller is a utility that will move VMs around hypervisors until the load (as defined in the configuration) in terms of memory and CPU usage reaches equilibrium.
The utility depends on having a Prometheus setup that can provide load metrics.
git clone https://github.com/manahl/openstack_load_levellercd openstack_load_leveller; pip install .
Modify the configuration file (and optionally rename it) to reflect your environment. It is recommended that you make the config file readable only by the user you will be running the load leveller as. Also, if you prefer, you can set the config options as environment variables instead.
The leveller needs to be able to connect to the Openstack API. For this to work, the following variables need to be defined:
OS_AUTH_URL:
The Openstack API URL
OS_USERNAME, OS_DOMAIN_NAME, OS_PASSWORD:
Credentials for an Admin user in Openstack
Openstack itself places no limitations on live migrations into or out of availability zones. This means that if you have more than one availability zone, you probably want to set the below parameter, otherwise you will probably find that VMs from one availability zone have migrated to other AZs after some time.
OS_LIMIT_TO_AZ:
Only consider hypervisors in a particular availability zone when evaluating
what to move.
The leveller needs a way of figuring out how loaded hosts are. The way this is currently done is by querying Prometheus to get load stats (I.e. you will need a Prometheus instance running in your environment).
The Prometheus options are described below:
PROMETHEUS_QUERY_URL:
Base URL for prometheus queries
PROMETHEUS_MEM_USED:
This query should return a normalised (0 to 1) number describing how much
memory is used on the box. 0 = no memory used, 1 = all memory used.
PROMETHEUS_CPU_USED:
Normalised CPU usage query (0 = no CPU used, 1 = all CPU used)
PROMETHEUS_METRIC_WEIGHT_MEM and PROMETHEUS_METRIC_WEIGHT_CPU:
This is the relative weights assigned to the memory/CPU components when
calculating the synthetic score.
E.g if with PROMETHEUS_METRIC_WEIGHT_MEM=2, and
PROMETHEUS_METRIC_WEIGHT_CPU=1, memory would be twice as significant in
the synthetic metric calculation.
PROMETHEUS_MIN_LOAD_BEFORE_MOVE:
Minimum load on the busiest hypervisor before a move is possible (i.e. if
the Openstack cluster is lightly loaded, do nothing).
PROMETHEUS_MIN_LOAD_DIFF_BEFORE_MOVE:
The minimum load difference betwen the highest and the lowest loaded
hypervisor before a move is initiated, measured as a ratio between 0 and 1.
This metric is calculated as: most_loaded_score / least_loaded_score - 1.0
If you installed the tool with pip, you can simply run openstacklb from
the command line:
$ openstacklb --help
Usage: openstacklb [OPTIONS]
Openstack load leveller script
This script attempts to "load balance" openstack, by live-migrating
instances from busy hypervisors to less busy hypervisors.
The busyness is calculated from the memory and cpu utilization (as
reported by prometheus).
The openstack credentials etc are stored in /etc/loadleveller-secrets.conf
(by default).
Exit codes:
EX_SOFTWARE (70) if Prometheus can't be contacted, VMs are lost etc.
EX_IOERR (74) for IO errors, specifically when reading in the config.
EX_NOPERM (77) If openstack denies an action based on permissions.
EX_CONFIG (78) For config errors, including requesting openstack
resources that do not exist.
Options:
--dryrun Run in dryrun mode - output what hypervisors would have a
host moved
--configfile TEXT Config file in 'dotenv' format.
--debug Dump debug messages to screen as well as log file
--help Show this message and exit.
Example:
$ openstacklb --configfile /etc/openstack-leveller.conf --dryrun --debug