aiida.scheduler documentation

We report here the generic AiiDA scheduler implementation.

Generic scheduler class

class aiida.scheduler.Scheduler[source]

Base class for all schedulers.

classmethod create_job_resource(**kwargs)[source]

Create a suitable job resource from the kwargs specified

getJobs(jobs=None, user=None, as_dict=False)[source]

Get the list of jobs and return it.

Typically, this function does not need to be modified by the plugins.

Parameters:
  • jobs (list) – a list of jobs to check; only these are checked
  • user (str) – a string with a user: only jobs of this user are checked
  • as_dict (list) – if False (default), a list of JobInfo objects is returned. If True, a dictionary is returned, having as key the job_id and as value the JobInfo object.

Note: typically, only either jobs or user can be specified. See also comments in _get_joblist_command.

get_detailed_jobinfo(jobid)[source]

Return a string with the output of the detailed_jobinfo command.

At the moment, the output text is just retrieved and stored for logging purposes, but no parsing is performed.

classmethod get_short_doc()[source]

Return the first non-empty line of the class docstring, if available

get_submit_script(job_tmpl)[source]

Return the submit script as a string. :parameter job_tmpl: a aiida.scheduler.datastrutures.JobTemplate object.

The plugin returns something like

#!/bin/bash <- this shebang line could be configurable in the future scheduler_dependent stuff to choose numnodes, numcores, walltime, … prepend_computer [also from calcinfo, joined with the following?] prepend_code [from calcinfo] output of _get_script_main_content postpend_code postpend_computer

kill(jobid)[source]

Kill a remote job, and try to parse the output message of the scheduler to check if the scheduler accepted the command.

..note:: On some schedulers, even if the command is accepted, it may take some seconds for the job to actually disappear from the queue.

Parameters:jobid (str) – the job id to be killed
Returns:True if everything seems ok, False otherwise.
logger

Return the internal logger.

set_transport(transport)[source]

Set the transport to be used to query the machine or to submit scripts. This class assumes that the transport is open and active.

submit_from_script(working_directory, submit_script)[source]

Goes in the working directory and submits the submit_script.

Return a string with the JobID in a valid format to be used for querying.

Typically, this function does not need to be modified by the plugins.

transport

Return the transport set for this scheduler.

aiida.scheduler.SchedulerFactory(module)[source]

Used to load a suitable Scheduler subclass.

Parameters:module (str) – a string with the module name
Returns:the scheduler subclass contained in module ‘module’

Scheduler datastructures

This module defines the main data structures used by the Scheduler.

In particular, there is the definition of possible job states (job_states), the data structure to be filled for job submission (JobTemplate), and the data structure that is returned when querying for jobs in the scheduler (JobInfo).

class aiida.scheduler.datastructures.JobInfo(init=None)[source]

Contains properties for a job in the queue. Most of the fields are taken from DRMAA v.2.

Note that default fields may be undefined. This is an expected behavior and the application must cope with this case. An example for instance is the exit_status for jobs that have not finished yet; or features not supported by the given scheduler.

Fields:

  • job_id: the job ID on the scheduler
  • title: the job title, as known by the scheduler
  • exit_status: the exit status of the job as reported by the operating system on the execution host
  • terminating_signal: the UNIX signal that was responsible for the end of the job.
  • annotation: human-readable description of the reason for the job being in the current state or substate.
  • job_state: the job state (one of those defined in aiida.scheduler.datastructures.job_states)
  • job_substate: a string with the implementation-specific sub-state
  • allocated_machines: a list of machines used for the current job. This is a list of MachineInfo objects.
  • job_owner: the job owner as reported by the scheduler
  • num_mpiprocs: the total number of requested MPI procs
  • num_cpus: the total number of requested CPUs (cores) [may be undefined]
  • num_machines: the number of machines (i.e., nodes), required by the job. If allocated_machines is not None, this number must be equal to len(allocated_machines). Otherwise, for schedulers not supporting the retrieval of the full list of allocated machines, this attribute can be used to know at least the number of machines.
  • queue_name: The name of the queue in which the job is queued or running.
  • wallclock_time_seconds: the accumulated wallclock time, in seconds
  • requested_wallclock_time_seconds: the requested wallclock time, in seconds
  • cpu_time: the accumulated cpu time, in seconds
  • submission_time: the absolute time at which the job was submitted, of type datetime.datetime
  • dispatch_time: the absolute time at which the job first entered the ‘started’ state, of type datetime.datetime
  • finish_time: the absolute time at which the job first entered the ‘finished’ state, of type datetime.datetime
class aiida.scheduler.datastructures.JobResource(init=None)[source]

A class to store the job resources. It must be inherited and redefined by the specific plugin, that should contain a _job_resource_class attribute pointing to the correct JobResource subclass.

It should at least define the get_tot_num_mpiprocs() method, plus an __init__ to accept its set of variables.

Typical attributes are:

  • num_machines
  • num_mpiprocs_per_machine

or (e.g. for SGE)

  • tot_num_mpiprocs
  • parallel_env

The __init__ should take care of checking the values. The init should raise only ValueError or TypeError on invalid parameters.

classmethod accepts_default_mpiprocs_per_machine()[source]

Return True if this JobResource accepts a ‘default_mpiprocs_per_machine’ key, False otherwise.

Should be implemented in each subclass.

get_tot_num_mpiprocs()[source]

Return the total number of cpus of this job resource.

classmethod get_valid_keys()[source]

Return a list of valid keys to be passed to the __init__

class aiida.scheduler.datastructures.JobTemplate(init=None)[source]

A template for submitting jobs. This contains all required information to create the job header.

The required fields are: working_directory, job_name, num_machines,
num_mpiprocs_per_machine, argv.

Fields:

  • submit_as_hold: if set, the job will be in a ‘hold’ status right after the submission

  • rerunnable: if the job is rerunnable (boolean)

  • job_environment: a dictionary with environment variables to set before the execution of the code.

  • working_directory: the working directory for this job. During submission, the transport will first do a ‘chdir’ to this directory, and then possibly set a scheduler parameter, if this is supported by the scheduler.

  • email: an email address for sending emails on job events.

  • email_on_started: if True, ask the scheduler to send an email when the job starts.

  • email_on_terminated: if True, ask the scheduler to send an email when the job ends. This should also send emails on job failure, when possible.

  • job_name: the name of this job. The actual name of the job can be different from the one specified here, e.g. if there are unsupported characters, or the name is too long.

  • sched_output_path: a (relative) file name for the stdout of this job

  • sched_error_path: a (relative) file name for the stdout of this job

  • sched_join_files: if True, write both stdout and stderr on the same file (the one specified for stdout)

  • queue_name: the name of the scheduler queue (sometimes also called partition), on which the job will be submitted.

  • job_resource: a suitable JobResource subclass with information on how many nodes and cpus it should use. It must be an instance of the aiida.scheduler.Scheduler._job_resource_class class. Use the Scheduler.create_job_resource method to create it.

  • num_machines: how many machines (or nodes) should be used

  • num_mpiprocs_per_machine: how many MPI procs should be used on each machine (or node).

  • priority: a priority for this job. Should be in the format accepted by the specific scheduler.

  • max_memory_kb: The maximum amount of memory the job is allowed to allocate ON EACH NODE, in kilobytes

  • max_wallclock_seconds: The maximum wall clock time that all processes of a job are allowed to exist, in seconds

  • custom_scheduler_commands: a string that will be inserted right after the last scheduler command, and before any other non-scheduler command; useful if some specific flag needs to be added and is not supported by the plugin

  • prepend_text: a (possibly multi-line) string to be inserted in the scheduler script before the main execution line

  • append_text: a (possibly multi-line) string to be inserted in the scheduler script after the main execution line

  • import_sys_environment: import the system environment variables

  • codes_info: a list of aiida.common.datastructures.CalcInfo objects. Each contains the information necessary to run a single code. At the moment, it can contain:

    • cmdline_parameters: a list of strings with the command line arguments of the program to run. This is the main program to be executed. NOTE: The first one is the executable name. For MPI runs, this will probably be “mpirun” or a similar program; this has to be chosen at a upper level.
    • stdin_name: the (relative) file name to be used as stdin for the program specified with argv.
    • stdout_name: the (relative) file name to be used as stdout for the program specified with argv.
    • stderr_name: the (relative) file name to be used as stderr for the program specified with argv.
    • join_files: if True, stderr is redirected on the same file specified for stdout.
  • codes_run_mode: sets the run_mode with which the (multiple) codes have to be executed. For example, parallel execution:

    mpirun -np 8 a.x &
    mpirun -np 8 b.x &
    wait
    

    The serial execution would be without the &’s. Values are given by aiida.common.datastructures.code_run_modes.

class aiida.scheduler.datastructures.MachineInfo(init=None)[source]

Similarly to what is defined in the DRMAA v.2 as SlotInfo; this identifies each machine (also called ‘node’ on some schedulers) on which a job is running, and how many CPUs are being used. (Some of them could be undefined)

  • name: name of the machine
  • num_cpus: number of cores used by the job on this machine
  • num_mpiprocs: number of MPI processes used by the job on this machine
class aiida.scheduler.datastructures.NodeNumberJobResource(**kwargs)[source]

An implementation of JobResource for schedulers that support the specification of a number of nodes and a number of cpus per node

classmethod accepts_default_mpiprocs_per_machine()[source]

Return True if this JobResource accepts a ‘default_mpiprocs_per_machine’ key, False otherwise.

get_tot_num_mpiprocs()[source]

Return the total number of cpus of this job resource.

classmethod get_valid_keys()[source]

Return a list of valid keys to be passed to the __init__

class aiida.scheduler.datastructures.ParEnvJobResource(**kwargs)[source]

An implementation of JobResource for schedulers that support the specification of a parallel environment (a string) + the total number of nodes

classmethod accepts_default_mpiprocs_per_machine()[source]

Return True if this JobResource accepts a ‘default_mpiprocs_per_machine’ key, False otherwise.

get_tot_num_mpiprocs()[source]

Return the total number of cpus of this job resource.