The nodes must all be able to communicate with the OpenQuake Engine DbServer and a RabbitMQ server. Both services run on a single "master" node. It is not necessary and not recommended to run RabbitMQ on a worker node and Celery on master node.
Running OpenQuake on an MPI cluster is currently not supported. See the FAQ for more information.
Have read Installing on RedHat and derivatives or Installing on Ubuntu (depending on the operating system been used).
The python3-oq-engine-master package must be installed on the master node. It provides extra functionalities like RabbitMQ.
On RHEL/CentOS EPEL repository must be configured and enabled in the system.
On worker nodes python3-oq-engine-worker must be installed instead; it adds celery support on top of the standard python3-oq-engine package.
Note: you have to restart every celery node after a configuration change.
In all the nodes, the following file should be modified to enable the Celery support:
/etc/openquake/openquake.cfg:
[distribution]
# enable celery only if you have a cluster
oq_distribute = celery
On all worker nodes, the /etc/openquake/openquake.cfg file should be also modified to set the DbServer and RabbitMQ daemons IP address:
[amqp]
host = w.x.y.z
port = 5672
user = openquake
password = openquake
vhost = openquake
# This is where tasks will be enqueued.
celery_queue = celery
[dbserver]
# enable multi_user if you have a multiple user installation
multi_user = true
file = /var/lib/openquake/db.sqlite3
log = /var/lib/openquake/dbserver.log
host = w.x.y.z
port = 1907
authkey = changeme
The required daemons are:
- RabbitMQ
- OpenQuake Engine DbServer
- OpenQuake Engine WebUI (optional)
A user and vhost are automatically added to the RabbitMQ configuration during the installation on python3-oq-engine.
You can verify that the openquake user and vhost exist running:
# Check the user
sudo rabbitmqctl list_users | grep openquake
# Check the vhost
sudo rabbitmqctl list_vhosts | grep openquake If, for any reason, the openquake user and vhost are missing they can be set up running:
# Add the user
sudo rabbitmqctl add_user openquake openquake
# Add the vhost
sudo rabbitmqctl add_vhost openquake
sudo rabbitmqctl set_permissions -p openquake openquake ".*" ".*" ".*"For more information please refer to https://www.rabbitmq.com/man/rabbitmqctl.1.man.html.
- Celery
celery must run all of the worker nodes. It can be started with
## RHEL
sudo service openquake-celery start
## Ubuntu
sudo supervisorctl start openquake-celeryThe Celery daemon is not started at boot by default on the workers node and the DbServer, WebUI can be disabled on the workers. Have a look at the documentation for Ubuntu or RedHat to see how to enable or disable services.
oq celery status can be used to check the status of the worker nodes, the task distribution and the cluster occupation. An output like this is produced:
==========
Host: celery@marley
Status: Online
Worker processes: 64
Active tasks: 0
==========
Host: celery@mercury
Status: Online
Worker processes: 64
Active tasks: 0
==========
Host: celery@dylan
Status: Online
Worker processes: 64
Active tasks: 0
==========
Host: celery@cobain
Status: Online
Worker processes: 64
Active tasks: 0
==========
Total workers: 256
Active tasks: 0
Cluster utilization: 0.00%
OpenQuake 2.4 introduces the concept of shared directory (aka shared_dir). This shared dir allows the workers to read directly from the master's filesystem, thus increasing scalability and performance; this feature is optional: the old behaviour, transmitting data via rabbitmq, will be used when shared_dir isn't set.
The shared directory must be exported from the master node to the workers via a POSIX compliant filesystem (like NFS). The export may be (and should be) exported and/or mounted as read-only by the workers.
As soon as the shared export is in place, the shared_dir config parameter in openquake.cfg must be configured to point to the path of the exported dir on the master node and to where the export is mounted on each worker node.
[directory]
# the base directory containing the <user>/oqdata directories;
# if set, it should be on a shared filesystem; for instance in our
# cluster it is /home/openquake; if not set, the oqdata directories
# go into $HOME/oqdata, unless the user sets his own OQ_DATADIR variable
shared_dir = /home/openquake
When shared_dir is set, the oqdata folders will be stored under $shared_dir/<user>/oqdata instead of /home/<user>/oqdata. See the comment in the openquake.cfg for further information.
You need then to give RWX permission to the shared_dir on master to the openquake group (which is usually created by packages) and add all the cluster users to the openquake group. For example:
$ mkdir /home/openquake
$ chown openquake.openquake /home/openquake
# Enable setgid on the tree
$ chmod 2770 /home/openquakeOn the workers the shared_dir should be mounted as the openquake user too, or access must be given to the user running celeryd (which is openquake by default in the official packages).
The worker nodes should be isolated from the external network using either a dedicated internal network or a firewall. Additionally, access to the RabbitMQ, and DbServer ports should be limited (again by internal LAN or firewall) so that external traffic is excluded.
The following ports must be open on the master node:
- 5672 for RabbitMQ
- 1907 for DbServer (or any other port allocated for the DbServer in the
openquake.cfg) - 1912-1920 for ZeroMQ receivers
- 8800 for the API/WebUI (optional)
The worker nodes must be able to connect to the master on port 5672, and port 1907.
Storage requirements depend a lot on the type of calculations you want to run. On a worker node you will need just the space for the operating system, the logs and the OpenQuake installation: less than 20GB are usually enough. Workers can be also diskless (using iSCSI or NFS for example). Starting from OpenQuake 2.3 the software and all its libraries are located in /opt/openquake.
On the master node you will also need space for:
- the users' home directory (usually located under
/home): it contains the calculations datastore (hdf5files located in theoqdatafolder) - the OpenQuake database (located under
/var/lib/openquake): it contains only logs and metadata, the expected size is tens of megabyte - RabbitMQ mnesia dir (usually located under
/var/lib/rabbitmq) - the temporary folder (
/tmp). A different temporary folder can be customized via theopenquake.cfg
On large installations we strongly suggest to create separate partition for /home, RabbitMQ (/var/lib/rabbitmq) and /tmp (or any custom temporary folder set in the openquake.cfg.
Having swap active on resources dedicated to the OpenQuake Engine is strongly discouraged because of the performance penality when it's being used and because how python allocates memory. In most cases (when memory throughput is relevant) is totally useless and it will just increase by many orders of magnitude the time required to complete the job (making the job actually stuck).
Jobs can be submitted through the master node using the oq engine command line interface, the API or the WebUI if active. See the documentation about how to run a calculation or about how to use the WebUI