Defining the workflow

One of the most important step that you have to do when planning to use autosubmit for an experiment is the definition of the workflow the experiment will use. In this section you will learn about the workflow definition syntax so you will be able to exploit autosubmit’s full potential

Warning

This section is NOT intended to show how to define your jobs. Please go to Tutorial section for a comprehensive list of job options.

Simple workflow

The simplest workflow that can be defined it is a sequence of two jobs, with the second one triggering at the end of the first. To define it, we define the two jobs and then add a DEPENDECIES attribute on the second job referring to the first one.

It is important to remember when defining workflows that DEPENDENCIES on autosubmit always refer to jobs that should be finished before launching the job that has the DEPENDENCIES attribute.

[One]
FILE = one.sh

[Two]
FILE = two.sh
DEPENDENCIES = One

The resulting workflow can be seen in Figure 25

simple workflow plot

25 Example showing a simple workflow with two sequential jobs

Running jobs once per startdate, member or chunk

Autosubmit is capable of running ensembles made of various startdates and members. It also has the capability to divide member execution on different chunks.

To set at what level a job has to run you have to use the RUNNING attribute. It has four possible values: once, date, member and chunk corresponding to running once, once per startdate, once per member or once per chunk respectively.

[once]
FILE = Once.sh

[date]
FILE = date.sh
DEPENDENCIES = once
RUNNING = date

[member]
FILE = Member.sh
DEPENDENCIES = date
RUNNING = member

[chunk]
FILE = Chunk.sh
DEPENDENCIES = member
RUNNING = chunk

The resulting workflow can be seen in Figure 26 for a experiment with 2 startdates, 2 members and 2 chunks.

simple workflow plot

26 Example showing how to run jobs once per startdate, member or chunk.

Dependencies

Dependencies on autosubmit were introduced on the first example, but in this section you will learn about some special cases that will be very useful on your workflows.

Dependencies with previous jobs

Autosubmit can manage dependencies between jobs that are part of different chunks, members or startdates. The next example will show how to make a simulation job wait for the previous chunk of the simulation. To do that, we add sim-1 on the DEPENDENCIES attribute. As you can see, you can add as much dependencies as you like separated by spaces

[ini]
FILE = ini.sh
RUNNING = member

[sim]
FILE = sim.sh
DEPENDENCIES = ini sim-1
RUNNING = chunk

[postprocess]
FILE = postprocess.sh
DEPENDENCIES = sim
RUNNING = chunk

The resulting workflow can be seen in Figure 27

Warning

Autosubmit simplifies the dependencies, so the final graph usually does not show all the lines that you may expect to see. In this example you can see that there are no lines between the ini and the sim jobs for chunks 2 to 5 because that dependency is redundant with the one on the previous sim

simple workflow plot

27 Example showing dependencies between sim jobs on different chunks.

Dependencies between running levels

On the previous examples we have seen that when a job depends on a job on a higher level (a running chunk job depending on a member running job) all jobs wait for the higher running level job to be finished. That is the case on the ini sim dependency on the next example.

In the other case, a job depending on a lower running level job, the higher level job will wait for ALL the lower level jobs to be finished. That is the case of the postprocess combine dependency on the next example.

[ini]
FILE = ini.sh
RUNNING = member

[sim]
FILE = sim.sh
DEPENDENCIES = ini sim-1
RUNNING = chunk

[postprocess]
FILE = postprocess.sh
DEPENDENCIES = sim
RUNNING = chunk

[combine]
FILE = combine.sh
DEPENDENCIES = postprocess
RUNNING = member

The resulting workflow can be seen in Figure 28

simple workflow plot

28 Example showing dependencies between jobs running at different levels.

Job frequency

Some times you just don’t need a job to be run on every chunk or member. For example, you may want to launch the postprocessing job after various chunks have completed. This behaviour can be achieved using the FREQUENCY attribute. You can specify an integer I for this attribute and the job will run only once for each I iterations on the running level.

Hint

You don’t need to adjust the frequency to be a divisor of the total jobs. A job will always execute at the last iteration of its running level

[ini]
FILE = ini.sh
RUNNING = member

[sim]
FILE = sim.sh
DEPENDENCIES = ini sim-1
RUNNING = chunk

[postprocess]
FILE = postprocess.sh
DEPENDENCIES = sim
RUNNING = chunk
FREQUENCY = 3

[combine]
FILE = combine.sh
DEPENDENCIES = postprocess
RUNNING = member

The resulting workflow can be seen in Figure 29

simple workflow plot

29 Example showing dependencies between jobs running at different frequencies.

Job synchronize

For jobs running at chunk level, and this job has dependencies, you could want not to run a job for each experiment chunk, but to run once for all member/date dependencies, maintaining the chunk granularity. In this cases you can use the SYNCHRONIZE job parameter to determine which kind of synchronization do you want. See the below examples with and without this parameter.

Hint

This job parameter works with jobs with RUNNING parameter equals to ‘chunk’.

[ini]
FILE = ini.sh
RUNNING = member

[sim]
FILE = sim.sh
DEPENDENCIES = INI SIM-1
RUNNING = chunk

[ASIM]
FILE = asim.sh
DEPENDENCIES = SIM
RUNNING = chunk

The resulting workflow can be seen in Figure 30

simple workflow plot

30 Example showing dependencies between chunk jobs running without synchronize.

[ASIM]
SYNCHRONIZE = member

The resulting workflow of setting SYNCHRONIZE parameter to ‘member’ can be seen in Figure 31

simple workflow plot

31 Example showing dependencies between chunk jobs running with member synchronize.

[ASIM]
SYNCHRONIZE = date

The resulting workflow of setting SYNCHRONIZE parameter to ‘date’ can be seen in Figure 32

simple workflow plot

32 Example showing dependencies between chunk jobs running with date synchronize.

Job split

For jobs running at chunk level, it may be useful to split each chunk into different parts. This behaviour can be achieved using the SPLITS attribute to specify the number of parts. It is possible to define dependencies to specific splits within [], as well as to a list/range of splits, in the format [1:3,7,10] or [1,2,3]

Hint

This job parameter works with jobs with RUNNING parameter equals to ‘chunk’.

[ini]
FILE = ini.sh
RUNNING = member

[sim]
FILE = sim.sh
DEPENDENCIES = ini sim-1
RUNNING = chunk

[asim]
FILE = asim.sh
DEPENDENCIES = sim
RUNNING = chunk
SPLITS = 3

[post]
FILE = post.sh
RUNNING = chunk
DEPENDENCIES = asim[1] asim[1]+1

The resulting workflow can be seen in Figure 33

simple workflow plot

33 Example showing the job ASIM divided into 3 parts for each chunk.

Job delay

Some times you need a job to be run after a certain number of chunks. For example, you may want to launch the asim job after various chunks have completed. This behaviour can be achieved using the DELAY attribute. You can specify an integer N for this attribute and the job will run only after N chunks.

Hint

This job parameter works with jobs with RUNNING parameter equals to ‘chunk’.

[ini]
FILE = ini.sh
RUNNING = member

[sim]
FILE = sim.sh
DEPENDENCIES = ini sim-1
RUNNING = chunk

[asim]
FILE = asim.sh
DEPENDENCIES = sim asim-1
RUNNING = chunk
DELAY = 2

[post]
FILE = post.sh
DEPENDENCIES = sim asim
RUNNING = chunk

The resulting workflow can be seen in Figure 34

simple workflow with delay option

34 Example showing the asim job starting only from chunk 3.

Rerun dependencies

Autosubmit has the possibility to rerun some chunks of the experiment without affecting everything else. In this case, autosubmit will automatically rerun all jobs of that chunk. If some of this jobs need another one on the workflow you have to add the RERUN_DEPENDENCIES attribute and specify which jobs to rerun.

It is also usual that you will have some code that it is needed only in the case of a rerun. You can add this jobs to the workflow as usual and set the attribute RERUN_ONLY to true. This jobs will be omitted from the workflow in the normal case, but will appear on the reruns.

[prepare_rerun]
FILE = prepare_rerun.sh
RERUN_ONLY = true
RUNNING = member

[ini]
FILE = ini.sh
RUNNING = member

[sim]
FILE = sim.sh
DEPENDENCIES = ini combine prepare_rerun
RERUN_DEPENDENCIES = combine prepare_rerun
RUNNING = chunk

[postprocess]
FILE = postprocess.sh
DEPENDENCIES = sim
RUNNING = chunk

[combine]
FILE = combine.sh
DEPENDENCIES = postprocess
RUNNING = member

The resulting workflow can be seen in Figure 35 for a rerun of chunks 2 and 3 of member 2.

simple workflow plot

35 Example showing a rerun workflow for chunks 2 and 3.