nova/doc/source/user/placement.rst
Matt Riedemann 4b7a150584 Update nova-status and docs for nova-compute requiring placement 1.14
With change I2f367b06e683ed7c815dd9e0536a46e5f0a27e6c, nova-compute
now unconditionally requires Placement 1.14 to be available (the
client side code doesn't check to see if 1.14 is available before
trying to use it).

This change updates the nova-status check for the minimum required
version of Placement and also starts the "Queens" section of the
Placement upgrade notes docs.

Change-Id: I37415e384d375bc9b548a0223f787a9236286bb0
2017-12-13 15:18:27 -05:00

15 KiB

Placement API

Overview

Nova introduced the placement API service in the 14.0.0 Newton release. This is a separate REST API stack and data model used to track resource provider inventories and usages, along with different classes of resources. For example, a resource provider can be a compute node, a shared storage pool, or an IP allocation pool. The placement service tracks the inventory and usage of each provider. For example, an instance created on a compute node may be a consumer of resources such as RAM and CPU from a compute node resource provider, disk from an external shared storage pool resource provider and IP addresses from an external IP pool resource provider.

The types of resources consumed are tracked as classes. The service provides a set of standard resource classes (for example DISK_GB, MEMORY_MB, and VCPU) and provides the ability to define custom resource classes as needed.

Each resource provider may also have a set of traits which describe qualitative aspects of the resource provider. Traits describe an aspect of a resource provider that cannot itself be consumed but a workload may wish to specify. For example, available disk may be solid state drives (SSD).

References

The following specifications represent the stages of design and development of resource providers and the Placement service. Implementation details may have changed or be partially complete at this time.

Deployment

The placement-api service must be deployed at some point after you have upgraded to the 14.0.0 Newton release but before you can upgrade to the 15.0.0 Ocata release. This is so that the resource tracker in the nova-compute service can populate resource provider (compute node) inventory and allocation information which will be used by the nova-scheduler service in Ocata.

Steps

1. Deploy the API service

At this time the placement API code is still in Nova alongside the compute REST API code (nova-api). So once you have upgraded nova-api to Newton you already have the placement API code, you just need to install the service. Nova provides a nova-placement-api WSGI script for running the service with Apache, nginx or other WSGI-capable web servers. Depending on what packaging solution is used to deploy OpenStack, the WSGI script may be in /usr/bin or /usr/local/bin.

Note

The placement API service is currently developed within Nova but it is designed to be as separate as possible from the existing code so that it can eventually be split into a separate project.

nova-placement-api, as a standard WSGI script, provides a module level application attribute that most WSGI servers expect to find. This means it is possible to run it with lots of different servers, providing flexibility in the face of different deployment scenarios. Common scenarios include:

In all of these scenarios the host, port and mounting path (or prefix) of the application is controlled in the web server's configuration, not in the configuration (nova.conf) of the placement application.

When placement was first added to DevStack it used the mod_wsgi style. Later it was updated to use mod_proxy_uwsgi. Looking at those changes can be useful for understanding the relevant options.

DevStack is configured to host placement at /placement on either the default port for http or for https (80 or 443) depending on whether TLS is being used. Using a default port is desirable.

By default, the placement application will get its configuration for settings such as the database connection URL from /etc/nova/nova.conf. The directory the configuration file will be found in can be changed by setting OS_PLACEMENT_CONFIG_DIR in the environment of the process that starts the application.

Note

When using uwsgi with a front end (e.g., apache2 or nginx) something needs to ensure that the uwsgi process is running. In DevStack this is done with systemd. This is one of many different ways to manage uwsgi.

This document refrains from declaring a set of installation instructions for the placement service. This is because a major point of having a WSGI application is to make the deployment as flexible as possible. Because the placement API service is itself stateless (all state is in the database), it is possible to deploy as many servers as desired behind a load balancing solution for robust and simple scaling. If you familiarize yourself with installing generic WSGI applications (using the links in the common scenarios list, above), those techniques will be applicable here.

2. Synchronize the database

In the Newton release the Nova api database is the only deployment option for the placement API service and the resources it manages. After upgrading the nova-api service for Newton and running the nova-manage api_db sync command the placement tables will be created.

Note

There are plans to add the ability to run the placement service with a separate placement database that would just have the tables necessary for that service and not everything else that goes into the Nova api database.

3. Create accounts and update the service catalog

Create a placement service user with an admin role in Keystone.

The placement API is a separate service and thus should be registered under a placement service type in the service catalog as that is what the resource tracker in the nova-compute node will use to look up the endpoint.

Devstack sets up the placement service on the default HTTP port (80) with a /placement prefix instead of using an independent port.

4. Configure and restart nova-compute services

The 14.0.0 Newton nova-compute service code will begin reporting resource provider inventory and usage information as soon as the placement API service is in place and can respond to requests via the endpoint registered in the service catalog.

nova.conf on the compute nodes must be updated in the [placement] group to contain credentials for making requests from nova-compute to the placement-api service.

Note

After upgrading nova-compute code to Newton and restarting the service, the nova-compute service will attempt to make a connection to the placement API and if that is not yet available a warning will be logged. The nova-compute service will keep attempting to connect to the placement API, warning periodically on error until it is successful. Keep in mind that Placement is optional in Newton, but required in Ocata, so the placement service should be enabled before upgrading to Ocata. nova.conf on the compute nodes will need to be updated in the [placement] group for credentials to make requests from nova-compute to the placement-api service.

Upgrade Notes

The follow sub-sections provide notes on upgrading to a given target release.

Note

As a reminder, the nova-status upgrade check </cli/nova-status> tool can be used to help determine the status of your deployment and how ready it is to perform an upgrade.

Ocata (15.0.0)

  • The nova-compute service will fail to start in Ocata unless the [placement] section of nova.conf on the compute is configured. As mentioned in the deployment steps above, the Placement service should be deployed by this point so the computes can register and start reporting inventory and allocation information. If the computes are deployed and configured before the Placement service, they will continue to try and reconnect in a loop so that you do not need to restart the nova-compute process to talk to the Placement service after the compute is properly configured.
  • The nova.scheduler.filter_scheduler.FilterScheduler in Ocata will fallback to not using the Placement service as long as there are older nova-compute services running in the deployment. This allows for rolling upgrades of the computes to not affect scheduling for the FilterScheduler. However, the fallback mechanism will be removed in the 16.0.0 Pike release such that the scheduler will make decisions based on the Placement service and the resource providers (compute nodes) registered there. This means if the computes are not reporting into Placement by Pike, build requests will fail with NoValidHost errors.
  • While the FilterScheduler technically depends on the Placement service in Ocata, if you deploy the Placement service after you upgrade the nova-scheduler service to Ocata and restart it, things will still work. The scheduler will gracefully handle the absence of the Placement service. However, once all computes are upgraded, the scheduler not being able to make requests to Placement will result in NoValidHost errors.
  • It is currently possible to exclude the CoreFilter, RamFilter and DiskFilter from the list of enabled FilterScheduler filters such that scheduling decisions are not based on CPU, RAM or disk usage. Once all computes are reporting into the Placement service, however, and the FilterScheduler starts to use the Placement service for decisions, those excluded filters are ignored and the scheduler will make requests based on VCPU, MEMORY_MB and DISK_GB inventory. If you wish to effectively ignore that type of resource for placement decisions, you will need to adjust the corresponding cpu_allocation_ratio, ram_allocation_ratio, and/or disk_allocation_ratio configuration options to be very high values, e.g. 9999.0.
  • Users of CellsV1 will need to deploy a placement per cell, matching the scope and cardinality of the regular nova-scheduler process.

Pike (16.0.0)

  • The nova.scheduler.filter_scheduler.FilterScheduler in Pike will no longer fall back to not using the Placement Service, even if older computes are running in the deployment.

  • The FilterScheduler now requests allocation candidates from the Placement service during scheduling. The allocation candidates information was introduced in the Placement API 1.10 microversion, so you should upgrade the placement service before the Nova scheduler service so that the scheduler can take advantage of the allocation candidate information.

    The scheduler gets the allocation candidates from the placement API and uses those to get the compute nodes, which come from the cell(s). The compute nodes are passed through the enabled scheduler filters and weighers. The scheduler then iterates over this filtered and weighed list of hosts and attempts to claim resources in the placement API for each instance in the request. Claiming resources involves finding an allocation candidate that contains an allocation against the selected host's UUID and asking the placement API to allocate the requested instance resources. We continue performing this claim request until success or we run out of allocation candidates, resulting in a NoValidHost error.

    For a move operation, such as migration, allocations are made in Placement against both the source and destination compute node. Once the move operation is complete, the resource tracker in the nova-compute service will adjust the allocations in Placement appropriately.

    For a resize to the same host, allocations are summed on the single compute node. This could pose a problem if the compute node has limited capacity. Since resizing to the same host is disabled by default, and generally only used in testing, this is mentioned for completeness but should not be a concern for production deployments.

Queens (17.0.0)

  • The minimum Placement API microversion required by the nova-compute service is 1.14 in order to support Nested Resource Providers. This means you must upgrade the Placement service before upgrading any nova-compute services to Queens.

REST API

The placement API service has its own REST API and data model. One can get a sample of the REST API via the functional test gabbits.

Microversions

The placement API uses microversions for making incremental changes to the API which client requests must opt into.

It is especially important to keep in mind that nova-compute is a client of the placement REST API and based on how Nova supports rolling upgrades the nova-compute service could be Newton level code making requests to an Ocata placement API, and vice-versa, an Ocata compute service in a cells v2 cell could be making requests to a Newton placement API.