Job Submission and Management
For the requirements a command, script, or program, must fulfil in order to function as a Cylc task, see Task Implementation.
When a task is ready Cylc generates a job script (see Job Scripts). The job script is submitted to be run by the job runner from the platform chosen for the task. Like other runtime properties, you can set a workflow default platform and override it for specific tasks or families:
[runtime] [[root]] # workflow defaults platform = platform_with_loadleveler [[foo]] # just task foo platform = platform_with_at
Supported Job Submission Methods
Cylc provided built-in support for the following batch submission systems:
Cylc provides support for the following job runner handlers.
Submits job scripts to the rudimentary Unix
Runs job scripts as Unix background processes.
Submits job scripts to loadleveler by the
Submits job scripts to IBM Platform LSF by the
Submits job scripts to the Moab workload manager with
Submits job scripts to PBS (or Torque) by the
Submits job scripts to Sun/Oracle Grid Engine with
Submits job scripts to Simple Linux Utility for Resource Management.
SLURM job submission and manipulation.
See Custom Job Submission Methods for how to add new job submission methods.
Default Directives Provided
For job runners that use job file directives (PBS, Loadleveler, etc.) default directives are provided to set the job name, stdout and stderr file paths, and the execution time limit (if specified).
Cylc constructs the job name string using a combination of the task ID and the
workflow name. PBS fails a job submit if the job name in
-N name is
too long. For version 12 or below, this is 15 characters. For version 13, this
is 236 characters. The default setting will truncate the job name string to 236
characters. If you have PBS 12 or older at your site, you should modify your
site’s global configuration file to allow the job name to be truncated at 15
global.cylc[platforms][<platform name>]job name length maximum.
Directives Section Quirks (PBS, SGE, …)
To specify an option with no argument, such as
-V in PBS or
-cwd in SGE you must give a null string as the directive value in
The left hand side of a setting (i.e. the string before the first equal sign)
must be unique. To specify multiple values using an option such as
-l option in PBS, SGE, etc., either specify all items in a single
(Left hand side is
-l. A second
-l=... line will override the first.)
Or separate the items:
-l select=28 -l ncpus=36 -l mpiprocs=18 -l ompthreads=2 -l walltime=12:00:00
There is no equal sign after
(Left hand sides are now
-l ncpus, etc.)
Task stdout and stderr Logs
When a task is ready to run Cylc generates a filename root to be used for the job script and log files. The file path contains the task name, cycle point, and a submit number that increments if the same task is re-triggered multiple times:
# job script: ~/cylc-run/my-workflow/basic/log/job/1/hello/01/job # stdout: ~/cylc-run/my-workflow/basic/log/job/1/hello/01/job.out # stderr: ~/cylc-run/my-workflow/basic/log/job/1/hello/01/job.err
How the stdout and stderr streams are directed into these files depends on the
job runner. The
background method just
uses appropriate output redirection on the command line, as shown above. The
loadleveler method writes
appropriate directives to the job script that is submitted to loadleveler.
Cylc obviously has no control over the stdout and stderr output from tasks that do their own internal output management (e.g. tasks that submit internal jobs and direct the associated output to other files). For less internally complex tasks, however, the files referred to here will be complete job logs.
Some job runners, such as
redirect a job’s stdout
and stderr streams to a separate cache area while the job is running. The
contents are only copied to the normal locations when the job completes. This
cylc cat-log will be unable to find the
job’s stdout and stderr streams while the job is running. Some sites with these
job runners are known to provide commands for viewing and/or
tail-follow a job’s stdout and stderr streams that are redirected to these
cache areas. If this is the case at your site, you can configure Cylc to make
use of the provided commands by adding some settings to the global site/user
[hosts] [[HOST]] # <= replace this with a real host name [[[batch systems]]] [[[[pbs]]]] err tailer = qcat -f -e \%(job_id)s out tailer = qcat -f -o \%(job_id)s err viewer = qcat -e \%(job_id)s out viewer = qcat -o \%(job_id)s
Overriding The Job Submission Command
To change the form of the actual command used to submit a job you
need to define a new
global.cylc[platforms][<platform name>]job runner command template.
[platform] [[my_custom_platform]] hosts = host1, host2 job runner = loadleveler # Use '-s' to stop llsubmit returning # until all job steps have completed: job runner command template = llsubmit -s %(job)s
%(job)s will be substituted by the job file path.
For supported job runners, jobs can be polled to
check that their true state matches what the scheduler expects based on received
job status messages or previous polls. The scheduler executes a process
on the task host, by non-interactive ssh, to interrogate the job runner, and to
job.status file of the task which is automatically generated by the
job script as it runs.
Polling may be required to update the workflow state correctly after unusual events such as
a job gets killed by the untrappable SIGKILL signal (e.g.
kill -9 PID)
a machine being rebooted with tasks running on it
network problems prevent task messages from getting back to the workflow host
the scheduler itself was down when active jobs finished
Tasks can be polled on demand by using the
cylc poll command.
cylc poll --help
Tasks are polled automatically, once, if they timeout while queueing in a
job runner and submission timeout is set.
for how to configure timeouts).
Tasks are polled multiple times, where necessary, when they exceed their
execution time limits. These are normally set with some initial delays to allow
the job runners to kill the jobs.
execution time limit polling intervals
for how to configure the polling intervals).
Any tasks recorded in the submitted or running states at workflow restart are automatically polled to determine what happened to them while the workflow was down.
By default, regular polling also takes place every 15 minutes whilst a job is submitted or running. The default polling intervals can be overridden in the global configuration:
The polling intervals can also be configured for individual tasks:
Polling can be used as the sole method of determining task status on hosts that do not allow task messages to be routed back to the workflow host. See Polling to Track Job Status.
For polling to work correctly, the job runner queueing system must have a
job listing command for listing your jobs, and that the job listing must
display job IDs as they are returned by the job runner queueing system submit
command. For example, for pbs, moab and sge, the
should list jobs with their IDs displayed in exactly the same format as they
are returned by the
Tasks can be killed on demand by using the
cylc kill command.
Execution Time Limit
You can specify an
execution time limit for all supported job
submission methods. E.g.:
[runtime] [[task-x]] execution time limit = PT1H
execution time limit setting will also inform the workflow when a
a job should complete by. If a job has not reported completing within
the specified time, the workflow will poll the job. (The default
PT1M, PT2M, PT7M. The accumulated times for these intervals will be
roughly 1 minute, 1 + 2 = 3 minutes and 1 + 2 + 7 = 10 minutes after a job
execution time limit.)
Execution Time Limit and Execution Timeout
If you specify an
execution time limit the
execution timeout event handler will only be called if the job has
not completed after the final poll (by default, 10 min after the time limit).
This should only happen if the submission method you are using is not enforcing
wallclock limits (unlikely) or you are unable to contact the machine to confirm
the job status.
If you specify an
[events]execution timeout and not an
execution time limit then the
execution timeout event handler will be called as soon as the
specified time is reached. The job will also be polled to check its latest
status (possibly resulting in an update in its status and the calling of the
relevant event handler). This behaviour is deprecated, which users should avoid