A typical libEnsemble workflow will include launching tasks from a sim_f (or gen_f) running on a worker. We use “task” to represent an application submission by libEnsemble to the system, may be the compute nodes of a supercomputer, cluster, or other compute resource.
The task could be launched via a subprocess call to
mpirun or an alternative
launcher such as
sim_f may then monitor this task,
check output, and possibly kill the task.
An Executor interface is provided by libEnsemble to remove the burden of
system interaction from the user and ease the writing of portable user scripts that
launch applications. The Executor provides the key functions:
kill(). Task attributes can be queried to determine
the status following each of these commands. Functions are also provided to access
and interrogate files in the task’s working directory.
The main Executor class is an abstract class and is inherited by the MPIExecutor, for direct running of MPI applications. Another Executor is the BalsamMPIExecutor, which submits an MPI run request from a worker running on a compute node to a Balsam process running on a launch node (suitable for systems that do not allow running MPI applications directly from compute nodes).
In a calling script, an Executor object is created, and the executable generator or simulation applications are registered to it for submission. If an alternative Executor like Balsam is used, then the applications can be registered as in the example below. Once in the user-side worker code (sim/gen func), an MPI-based Executor can be retrieved without any need to specify the type.
Example usage (code runnable with or without a Balsam backend):
In calling function:
sim_app = '/path/to/my/exe' USE_BALSAM = False if USE_BALSAM: from libensemble.executors.balsam_executor import BalsamMPIExecutor exctr = BalsamMPIExecutor() else: from libensemble.executors.mpi_executor import MPIExecutor exctr = MPIExecutor() exctr.register_calc(full_path=sim_app, calc_type='sim')
The Executor set up in the calling script is stored as a class attribute and does not have to be passed to libE. It is extracted via Executor.executor in the sim function (regardless of type).
In user sim func:
import time from libensemble.executors.executor import Executor # Will return Executor (whether MPI or inherited such as Balsam). exctr = Executor.executor task = exctr.submit(calc_type='sim', num_procs=8, app_args='input.txt', stdout='out.txt', stderr='err.txt') timeout_sec = 600 poll_delay_sec = 1 while(not task.finished): # Has manager sent a finish signal exctr.manager_poll() if exctr.manager_signal == 'finish': task.kill() my_cleanup() # Check output file for error and kill task elif task.stdout_exists(): if 'Error' in task.read_stdout(): task.kill() elif task.runtime > timeout_sec: task.kill() # Timeout else: time.sleep(poll_delay_sec) task.poll() print(task.state) # state may be finished/failed/killed
See the executor interface for API.
For a more realistic example see
the Electrostatic Forces example,
which launches the
forces.x application as an MPI task.
Applications or tasks submitted via the Balsam Executor are referred to as “jobs” within Balsam, including within Balsam’s database and when describing the state of a completed submission.
Note that applications can also be registered to the Executor using a name. The equivalent lines in the above example would be:
User sim func:
task = exctr.submit(app_name='forces_app', num_procs=8, app_args='input.txt', stdout='out.txt', stderr='err.txt')
app_name can be any identfier, while
full_path is the application to
be run. This approach allows multiple applications to be registered.
The MPIExecutor autodetects system criteria such as the appropriate MPI launcher
and mechanisms to poll and kill tasks. It will also partition resources amongst
workers, ensuring that runs utilise different resources (e.g. nodes). The
zero_resource_workers list option specifies workers that will not need
resources (e.g. a persistent generator might run on worker 1).
Furthermore, the MPIExecutor offers resilience via the feature of re-launching
tasks that fail because of system factors.
Various back-end mechanisms may be used by the Executor to best interact with each system, including proxy launchers or task management systems such as Balsam. Currently, these Executors launch at the application level within an existing resource pool. However, submissions to a batch scheduler may be supported in future Executors.
See Running on HPC Systems to see, with
diagrams, how common Executor options such as
central_mode affect the
run configuration on clusters and supercomputers.