Tuning Job execution#

Jobs with time consuming calculations require to properly configure the environment and the resources used to execute them. This section focuses on which options can be tuned and the ways available in jobflow-remote to change them.

Tuning options#

Worker#

A worker is a computational unit that will actually execute the function inside a Job. The list of workers is given in the Workers project configuration.

Workers are set by the name used to define them in the project and a worker should always be defined for each Job when adding a Flow to the database.

Note

A single worker should not necessarily be identified with a computation resource as a whole. Different workers referring to the same for example to the same HPC center, but with different configurations can be created. The Runner will still open a single connection if the host is the same.

Execution configuration#

An execution configuration, represented as an ExecutionConfig object in the code, contains information to run additional commands before and after the execution of a Job.

These can be typically used to define the modules to load on an HPC center, specific python environment to load or setting the PATH for some executable needed by the Job.

They can be usually given as a string referring to the setting defined in the project configuration file, or as an instance of ExecutionConfig.

Resources#

If the worker executing the Job runs under the control of a queueing system (e.g. SLURM, PBS), it is also important to specify which resources need to be allocated when running a Job.

Since the all the operations involving the queueing system are handled with qtoolkit, jobflow-remote supports the same functionalities. In particular it is either possible to pass a dictionary containing the keywords specific to the selected queuing system or to pass an instance of a QResources, a generic object defining resources for standard use cases. These will be used to fill in a template and generate a suitable submission script.

Note

There are SLURM and PBS specific keywords that can be passed to submit_flow to use the respective queueing system commands.

How to tune#

Different ways of setting the worker, execution configuration and resources for each Job are available. A combination of them can be used to ease the configuration for all the Jobs.

Note

If not defined otherwise, Jobs generated dynamically will inherit the configuration of the Job that generated them.

Submission#

The first entry point to customize the execution of the Jobs in a Flow is to use the arguments in the submit_flow function.

resource = {"nodes": 1, "ntasks": 4, "partition": "batch"}
submit_flow(
    flow, worker="local_shell", exec_config="somecode_v.x.y", resources=resources
)

This will set the passed values for all the Jobs for which have not been set in the Job previously.

Warning

Once the flow has been submitted to the database, any further change to the Flow object will not be taken into account.

JobConfig#

Each jobflow’s Job has a JobConfig attribute. This can be used to store a manager_config dictionary with configuration specific to that Job.

This can be done with the set_run_config function, that targets Jobs based on their name or on the callable they are wrapping. Consider the following example

from jobflow_remote.utils.examples import add, value
from jobflow_remote import submit_flow, set_run_config
from jobflow import Flow

job1 = value(5)
job2 = add(job1.output, 2)

flow = Flow([job1, job2])

flow = set_run_config(
    flow, name_filter="add", worker="secondw", exec_config="anotherconfig"
)

resource = {"nodes": 1, "ntasks": 4, "partition": "batch"}
submit_flow(flow, worker="firstw", exec_config="somecode_v.x.y", resources=resources)

After being submitted to the database the value Job will be executed on the firstw worker, while the add Job will be executed on the secondw worker. On the other hand, since resources is not set explicitly when set_run_config is called the same resource dictionary is applied to all the Jobs in the Flow.

Note

If the values in a Job should not be overridden by those passed in submit_flow, but no specific value is required, empty objects need to be passed. An empty dict or an empty QResources() to set an empty resources and an empty ExecutionConfig() to set and empty exec_config

Warning

If set_run_config is used to set a worker with scheduler_type: shell always set resources to an empty dictionary (or QResources) as well. Otherwise the resources value passed to submit_flow will be used and it will likely be incompatible, with the shell worker.

In addition, since set_run_config makes use of jobflow’s update_config method, these updates will also automatically be applied to any new Job automatically generated in the Flow.

Warning

The name_filter matches any name containing the string passed. So using a name_filter=add will match both a job named add and one named add more.

CLI#

After a Job has been added to the database, it is still possible to change its settings. This can be achieved with the jf job set CLI command. For example running:

jf job set worker -did 8 example_worker

sets the worker for Job with DB id 8 to example_worker. Similarly, the jf job set resources and jf job set exec-config can be used to set the values of the resources and execution configurations.

Note

In order for this to be meaningful only a subset of Jobs states are acceptable. This commands can be applied only to Jobs in the READY, WAITING, COMPLETED, FAILED, PAUSED and REMOTE_ERROR states.

On this page