Cells functionality allows you to scale an OpenStack Compute cloud in a more distributed fashion without having to use complicated technologies like database and message queue clustering. It is intended to support very large deployments. When this functionality is enabled, the hosts in an OpenStack Compute cloud are partitioned into groups called cells. Cells are configured as a tree. The top-level cell should have a host that runs a nova-api service, but no nova-compute services. Each child cell should run all of the typical nova-* services in a regular Compute cloud except for nova-api. You can think of cells as a normal Compute deployment in that each cell has its own database server and message queue broker. The nova-cells service handles communication between cells and selects cells for new instances. This service is required for every cell. Communication between cells is pluggable, and currently the only option is communication through RPC. Cells scheduling is separate from host scheduling. nova-cells first picks a cell (now randomly, but future releases plan to add filtering/weighing functionality, and decisions will be based on broadcasts of capacity/capabilities). Once a cell is selected and the new build request reaches its nova-cells service, it is sent over to the host scheduler in that cell and the build proceeds as it would have without cells. Cell functionality is currently considered experimental.

Cells are disabled by default. All cell-related configuration options go under a [cells] section in nova.conf. The following cell-related options are currently supported: enable Set this is True to turn on cell functionality, which is off by default. name Name of the current cell. This must be unique for each cell. capabilities List of arbitrary key=value pairs defining capabilities of the current cell. Values include hypervisor=xenserver;kvm,os=linux;windows. call_timeout How long in seconds to wait for replies from calls between cells. scheduler_filter_classes Filter classes that the cells scheduler should use. By default, uses "nova.cells.filters.all_filters" to map to all cells filters included with Compute. scheduler_weight_classes Weight classes the cells scheduler should use. By default, uses "nova.cells.weights.all_weighers" to map to all cells weight algorithms (weighers) included with Compute. ram_weight_multiplier Multiplier used for weighing ram. Negative numbers mean you want Compute to stack VMs on one host instead of spreading out new VMs to more hosts in the cell. Default value is 10.0.

The compute API class must be changed in the API cell so that requests can be proxied through nova-cells down to the correct cell properly. Add the following to nova.conf in the API cell:[DEFAULT] compute_api_class=nova.compute.cells_api.ComputeCellsAPI ... [cells] enable=True name=api

Add the following to nova.conf in the child cells, replacing cell1 with the name of each cell:[DEFAULT] # Disable quota checking in child cells. Let API cell do it exclusively. quota_driver=nova.quota.NoopQuotaDriver [cells] enable=True name=cell1

Before bringing the services online, the database in each cell needs to be configured with information about related cells. In particular, the API cell needs to know about its immediate children, and the child cells must know about their immediate agents. The information needed is the RabbitMQ server credentials for the particular cell. Use the nova-manage cell create command to add this information to the database in each cell:# nova-manage cell create -h Options: -h, --help show this help message and exit --name=<name> Name for the new cell --cell_type=<parent|child> Whether the cell is a parent or child --username=<username> Username for the message broker in this cell --password=<password> Password for the message broker in this cell --hostname=<hostname> Address of the message broker in this cell --port=<number> Port number of the message broker in this cell --virtual_host=<virtual_host> The virtual host of the message broker in this cell --woffset=<float> (weight offset) It might be used by some cell scheduling code in the future --wscale=<float> (weight scale) It might be used by some cell scheduling code in the future As an example, assume we have an API cell named api and a child cell named cell1. Within the api cell, we have the following RabbitMQ server info:rabbit_host=10.0.0.10 rabbit_port=5672 rabbit_username=api_user rabbit_password=api_passwd rabbit_virtual_host=api_vhost And in the child cell named cell1 we have the following RabbitMQ server info:rabbit_host=10.0.1.10 rabbit_port=5673 rabbit_username=cell1_user rabbit_password=cell1_passwd rabbit_virtual_host=cell1_vhost We would run this in the API cell, as root.# nova-manage cell create --name=cell1 --cell_type=child --username=cell1_user --password=cell1_passwd --hostname=10.0.1.10 --port=5673 --virtual_host=cell1_vhost --woffset=1.0 --wscale=1.0 Repeat the above for all child cells. In the child cell, we would run the following, as root:# nova-manage cell create --name=api --cell_type=parent --username=api1_user --password=api1_passwd --hostname=10.0.0.10 --port=5672 --virtual_host=api_vhost --woffset=1.0 --wscale=1.0 To customize the Compute cells, use the configuration option settings documented in .

To determine the best cell for launching a new instance, Compute uses a set of filters and weights configured in /etc/nova/nova.conf. The following options are available to prioritize cells for scheduling: scheduler_filter_classes - Specifies the list of filter classes. By default nova.cells.weights.all_filters is specified, which maps to all cells filters included with Compute (see . scheduler_weight_classes - Specifies the list of weight classes. By default nova.cells.weights.all_weighers is specified, which maps to all cell weight algorithms (weighers) included with Compute. The following modules are available: mute_child: Downgrades the likelihood of child cells being chosen for scheduling requests, which haven't sent capacity or capability updates in a while. Options include mute_weight_multiplier (multiplier for mute children; value should be negative) and mute_weight_value (assigned to mute children; should be a positive value). ram_by_instance_type: Select cells with the most RAM capacity for the instance type being requested. Because higher weights win, Compute returns the number of available units for the instance type requested. The ram_weight_multiplier option defaults to 10.0 that adds to the weight by a factor of 10. Use a negative number to stack VMs on one host instead of spreading out new VMs to more hosts in the cell. weight_offset: Allows modifying the database to weight a particular cell. You can use this when you want to disable a cell (for example, '0'), or to set a default cell by making its weight_offset very high (for example, '999999999999999'). The highest weight will be the first cell to be scheduled for launching an instance. Additionally, the following options are available for the cell scheduler: scheduler_retries - Specifies how many times the scheduler tries to launch a new instance when no cells are available (default=10). scheduler_retry_delay - Specifies the delay (in seconds) between retries (default=2). As an admin user, you can also add a filter that directs builds to a particular cell. The policy.json file must have a line with "cells_scheduler_filter:TargetCellFilter" : "is_admin:True" to let an admin user specify a scheduler hint to direct a build to a particular cell.

Cells currently keeps all inter-cell communication data, including user names and passwords, in the database. This is undesirable and unnecessary since cells data isn't updated very frequently. Instead, create a JSON file to input cells data specified via a [cells]cells_config option. When specified, the database is no longer consulted when reloading cells data. The file will need the columns present in the Cell model (excluding common database fields and the id column). The queue connection information must be specified through a transport_url field, instead of username, password, and so on. The transport_url has the following form: rabbit://<username>:<password>@<hostname>:<port>/<virtual_host> The scheme can be either qpid or rabbit, as shown previously. The following sample shows this optional configuration: