Job Managers¶
By default the Pulsar will maintain its own queue of jobs. Under Linux however, Pulsar can be configured to maintain multiple such queues with different properties or to delegate to external job queues (via DRMAA, qsub/qstat CLI commands, or Condor).
To configure job managers, uncomment the managers
section of app.yml
and modify it as needed. For instance, the default job manager corresponds to
a configuration of
managers:
_default_:
type: queued_python
num_concurrent_jobs: 1
The type
of queued_python
is indicating that the jobs are queued but
that the queue is managed locally by Pulsar. Other possible values for
type
include queued_drmaa
, queued_condor
, queued_cli
,
queued_external_drmaa
(examples of each follow).
Named Managers¶
The managers
section can contain any number of named managers. For example:
managers:
_default_:
type: queued_python
num_concurrent_jobs: 1
example:
type: queued_python
num_concurrent_jobs: "*"
In this instance, Pulsar creates a second named queued (example
) that will
run as many concurrent jobs as the server has cores. The Galaxy Pulsar url
should have /managers/example
appended to it to use a named manager such
as this.
DRMAA¶
The queued_python
manager type is easy to configure but has serious
limitations - for instance jobs running when Pulsar is restarted will be lost.
For these reasons it is best to configure a real external job manager when
possible.
Likely the cleanest way to interface with an external queueing system is going
to be DRMAA. This method will likely work with Slurm, PBS Torque, LSF,
etc…. In this case, one should likely setup a local_env.sh
file and
update it to set DRMAA_LIBRARY_PATH
to point to the correct
libdrmaa.so
file. Also, the Python drmaa
module must be installed
(e.g. via pip install drmaa
).:
managers:
_default_:
type: queued_drmaa
native_specification: "-P bignodes -R y -pe threads 8"
Here the optional native_specification
is going to depend on the underlying job
manager.
In addition to the default dependencies described in the installation
documentation, a DRMAA library will need to be installed and the python
dependency drmaa
will need to be installed as well to use the
queued_drmaa
manager. This can be done by activating Pulsar’s virtual environment and running:
pip install drmaa
If you are using DRMAA, be sure to define DRMAA_LIBRARY_PATH
in Pulsar’s
local_env.sh
file.
Condor¶
Condor can also be used as a backend.
managers:
_default_:
type: queued_condor
# Optional attributes...
submit_universe: vanilla
submit_request_memory: 32
submit_requirements: 'OpSys == "LINUX" && Arch =="INTEL"'
submit_rank: "Memory >= 64"
This would set universe, request_memory, requirements, and rank in the condor submission file to the specified values. For more information on condor submission files see the HTCondor quickstart for more information.
CLI¶
Pulsar can manage jobs via command-line execution of qsub
, qdel
,
stat
on the local machine.
managers:
_default_:
type: queued_cli
job_plugin: Torque
job_plugin
can also be slurm
(to use srun
, etc…) or
slurm_torque
(to use the Slurm varaint of qsub
, etc…).
Pulsar can also login into a remote host before executing these commands if the job manager is not accessible from the Pulsar host.
managers:
_default_:
type: queued_cli
job_plugin: Torque
shell_plugin: SecureShell
shell_hostname: queuemanager
shell_username: queueuser
This will login to queuemanager
as user queueuser
to submit jobs. Be
sure keyless SSH between Pulsar and the remote host is configured in this
case.
Run-As-Real User DRMAA¶
All of the proceeding will run jobs as the same operating system user that
Pulsar is running as. The queued_external_drmaa
manager type will actually
run DRMAA jobs via the user requested by the client (e.g. the Galaxy user).
managers:
_default_:
type: queued_external_drmaa
production: true
# Following are optional - should leave as defaults in most cases.
#chown_working_directory_script: scripts/chown_working_directory.bash
#drmaa_kill_script: scripts/drmaa_kill.bash
#drmaa_launch_script: scripts/drmaa_launch.bash
For more information on running jobs as the real user, check out this discussion from the Galaxy mailing list.
More Options¶
Any manager can override the staging_directory
used by setting this
property in its configuration section.
The min_polling_interval: 0.5
option can be set on any manager to control
how frequently Pulsar will poll the resource manager for job updates.
For staging actions initiated by Pulsar (e.g. when driving Pulsar by message queue) - the following parameters can be set to control retrying these actions (if they) fail. (XXX_max_retries=-1 => no retry, XXX_max_retries=0 => retry forever - this may be a bit counter-intuitive but is consistent with Kombu.
preprocess_action_max_retries: -1
preprocess_action_interval_start: 2
preprocess_action_interval_step: 2
preprocess_action_interval_max: 30
postprocess_action_max_retries: -1
postprocess_action_interval_start: 2
postprocess_action_interval_step: 2
postprocess_action_interval_max: 30