This service provides a REST-API to retrieve information about OpenStack flavors (total and available). It is written for de.NBI cloud OpenStack setups especially in Bielefeld, but can easily be adapted to any OpenStack installation providing GPU (or other) resources.
The current implementation is not yet fully compatible with
OpenStack Unified Resources, when
using special resources like GPUs.
The OpenStackFlavorExporter is the successor to the OpenstackGPUService and, for compatibility reasons, adopts its API.
Without special resources (such as GPUs), the Flavor Exporter should run with any reasonably current OpenStack implementation (tested with OpenStack Caracal - 2024.1).
Provided that GPU-supported flavor or the specified GPUs are taken into account, it needs some additional configuration of aggregates and flavors to work properly.
GPU hypervisors have to be grouped into different aggregates based on their GPU type (gpu_type)
and on the number of GPUs they provide (gpu_count). Every aggregate has to make some extra
metadata available:
- gpu='' Whether this aggregate provides GPU resources.
- gpu_count='' Number of GPUs provided.
- gpu_type='' Type (alias name) of GPU provided.
$ openstack aggregate show 'Delta GPU V100'
+-------------------+--------------------------------------------------------------------------------+
| Field | Value |
+-------------------+--------------------------------------------------------------------------------+
| availability_zone | default |
| created_at | 2018-10-13T19:08:59.000000 |
| deleted_at | None |
| hosts | host1,host2,... |
| id | 10 |
| is_deleted | False |
| name | Delta GPU V100 |
| properties | cores='28', gpu='True', gpu_count='2', gpu_type='V100', hyperthreading='false' |
| updated_at | 2022-03-23T14:31:00.000000 |
| uuid | xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx |
+-------------------+--------------------------------------------------------------------------------+
Currently GPU are supported using the PCI-passthrough feature of Openstack. Flavors must be configured setting the pci_passthrough:alias property=':. The service matches the alias name with aggregates gpu_type (aggregates metadata) to calculate the total count and current available number of each flavor.
There is of course more configuration work to do to support GPU(PCI) devices properly.
Python 3.10+
It recommend to install dependencies in a separate virtual environment.
$ python3 -m venv venv
$ source venv/bin/activate
(venv)$ pip3 install -r requirements.txt
When connecting to a "real" Openstack Setup, you have to source your RC file beforehand. The Openstack user should has at least the capability to list all servers, flavors, hypervisor and aggregates.
(venv)$ source my_project-openrc.sh
For performance reasons (and better user feelings) the server returns only cached data. The cache (only memcached is supported) is updated continuously.
- A user/client requests for data.
- The server ask the cache if the requested resources are available.
- Depending on the cache status, the result is returned immediately or calculated (long running task!).
- Independent of the REST API, the cache is updated continuously (every 150 seconds by default).
To run the server, please execute the following from the root directory:
(venv)$ python3 OpenStackFlavorExporterServer.py
When running OpenstackGPUServer without further arguments a memcached service listening on 127.0.0.1:11211 is expected and REST is bind to http://127.0.0.1:8080.
The swagger documentation is available at:
http://localhost:8080/docs/
Your OpenAPI definition lives here:
http://localhost:8080/openapi.json
curl -X 'GET' 'http://localhost:8080/gpus/flavors/' -H 'accept: application/json'
curl -X 'GET' 'http://localhost:8080/v2/flavors/' -H 'accept: application/json'
The project comes with a full set of units test, covering the internal and REST (FastAPI) functions. All external dependencies concerning the Openstack API are mocked using unittest.mock functionality. Run pytest from project root dir to run the tests.
To launch the integration tests, use tox from the root directory:
(venv)$ pip install tox
(venv)$ tox
To run the server inside a Docker container, please execute the following from the root directory:
# building the image
docker build -t denbi/openstackflavorexporter .
# start up a container in the host network
docker run --env-file ./env.file denbi/openstackflavorexporterwhere env.file contains the Openstack environment :
OS_REGION_NAME=Bielefeld
OS_PROJECT_DOMAIN_ID=default
OS_INTERFACE=public
OS_AUTH_URL=https://openstack.cebitec.uni-bielefeld.de:5000/v3/
OS_USERNAME=XXXX
OS_PROJECT_ID=XXXXX
OS_USER_DOMAIN_NAME=Default
OS_PROJECT_NAME=XXXXX
OS_PASSWORD=XXXXXX
Run a OpenStackFlavorExporter together with a memcached service container with docker compose is also supported.
The OpenStackFlavorExporter is based on FastAPI and therefore does not implement any authentication or authorization for protection. It is strongly recommended to run the OpenStackFlavorExporter behind a reverse proxy.
The project is licensed under Apache License Version 2.0.