Executor with Electrostatic Forces

This tutorial highlights libEnsemble’s capability to execute and monitor external scripts or user applications within simulation or generator functions using the executor. In this tutorial, our calling script registers a compiled executable that simulates electrostatic forces between a collection of particles. The simulator function launches instances of this executable and reads output files to determine if the run was successful.

It is possible to use subprocess calls from Python to issue commands such as jsrun or aprun to run applications. Unfortunately, hard-coding such commands within user scripts isn’t portable. Furthermore, many systems like Argonne’s Theta do not allow libEnsemble to submit additional tasks from the compute nodes. On these systems, a proxy launch mechanism (such as Balsam) is required. libEnsemble’s Executors were developed to directly address such issues.

In particular, we’ll be experimenting with libEnsemble’s MPI Executor, since it can automatically detect available MPI runners and resources, and by default divide them equally among workers.

Getting Started

The simulation source code forces.c can be obtained directly from the libEnsemble repository here.

Assuming MPI and its C compiler mpicc are available, compile forces.c into an executable (forces.x) with:

$ mpicc -O3 -o forces.x forces.c -lm

Calling Script

Let’s begin by writing our calling script to parameterize our simulation and generation functions and call libEnsemble. Create a Python file called run_libe_forces.py containing:

#!/usr/bin/env python
import os
import numpy as np
from forces_simf import run_forces  # Sim func from current dir

from libensemble.libE import libE
from libensemble.gen_funcs.sampling import uniform_random_sample
from libensemble.tools import parse_args, add_unique_random_streams
from libensemble.executors import MPIExecutor

# Parse number of workers, comms type, etc. from arguments
nworkers, is_manager, libE_specs, _ = parse_args()

# Initialize MPI Executor instance
exctr = MPIExecutor()

# Register simulation executable with executor
sim_app = os.path.join(os.getcwd(), "../forces_app/forces.x")
exctr.register_app(full_path=sim_app, app_name="forces")

On line 15, we instantiate our MPI Executor class instance, which can optionally be customized by specifying alternative MPI runners. The auto-detected default should be sufficient.

Registering an application is as easy as providing the full file-path and giving it a memorable name. This Executor instance will later be retrieved within our simulation function to launch the registered app.

Next define the sim_specs and gen_specs data structures. Recall that these are used to specify to libEnsemble what user functions and input/output fields to expect, and also to parameterize function instances without hard-coding:

# State the sim_f, inputs, outputs
sim_specs = {
    "sim_f": run_forces,  # sim_f, imported above
    "in": ["x"],  # Name of input for sim_f
    "out": [("energy", float)],  # Name, type of output from sim_f
}

# State the gen_f, inputs, outputs, additional parameters
gen_specs = {
    "gen_f": uniform_random_sample,  # Generator function
    "in": [],  # Generator input
    "out": [("x", float, (1,))],  # Name, type, and size of data from gen_f
    "user": {
        "lb": np.array([1000]),  # User parameters for the gen_f
        "ub": np.array([3000]),
        "gen_batch_size": 8,
    },
}

Our generation function will generate random numbers of particles (between the "lb" and "ub" bounds) for our simulation function to evaluate via our registered application.

The following additional instructs libEnsemble’s workers to each create and work within a separate directory each time they call a simulation function. This helps organize output and also helps prevents workers from overwriting previous results:

# Create and work inside separate per-simulation directories
libE_specs["sim_dirs_make"] = True

After configuring persis_info and exit_criteria, we initialize libEnsemble by calling the primary libE() routine:

# Instruct libEnsemble to exit after this many simulations
exit_criteria = {"sim_max": 8}

# Seed random streams for each worker, particularly for gen_f
persis_info = add_unique_random_streams({}, nworkers + 1)

# Launch libEnsemble
H, persis_info, flag = libE(sim_specs, gen_specs, exit_criteria, persis_info=persis_info, libE_specs=libE_specs)

Exercise

This may take some additional browsing of the docs to complete.

Write an alternative Calling Script similar to above, but with the following differences:

1. Add an additional worker directory so workers operate in /scratch/ensemble instead of the default current working directory.

2. Override the MPIExecutor’s detected MPI runner with "openmpi".

3. Set libEnsemble’s logger to print debug messages.

4. Use the save_libE_output() function to save the History array and persis_info to files after libEnsemble completes.

Click Here for Solution

#!/usr/bin/env python
import os
import numpy as np
from forces_simf import run_forces  # Sim func from current dir

from libensemble import logger
from libensemble.libE import libE
from libensemble.gen_funcs.sampling import uniform_random_sample
from libensemble.tools import parse_args, add_unique_random_streams, save_libE_output
from libensemble.executors import MPIExecutor

# Parse number of workers, comms type, etc. from arguments
nworkers, is_manager, libE_specs, _ = parse_args()

# Adjust logger level
logger.set_level("DEBUG")

# Initialize MPI Executor instance
exctr = MPIExecutor(custom_info={"mpi_runner": "openmpi"})

...

# Instruct workers to operate somewhere else on the filesystem
libE_specs["ensemble_dir_path"] = "/scratch/ensemble"

...

# Launch libEnsemble
H, persis_info, flag = libE(sim_specs, gen_specs, exit_criteria, persis_info=persis_info, libE_specs=libE_specs)

if is_manager:
    save_libE_output(H, persis_info, __file__, nworkers)

Simulation Function

Our simulation function is where we’ll use libEnsemble’s executor to configure and submit our application for execution. We’ll poll this task’s state while it runs, and once we’ve detected it has finished we’ll send any results or exit statuses back to the manager.

Create another Python file named forces_simf.py containing the following for starters:

import numpy as np

# To retrieve our MPI Executor instance
from libensemble.executors.executor import Executor

# Optional status codes to display in libE_stats.txt for each gen or sim
from libensemble.message_numbers import WORKER_DONE, TASK_FAILED


def run_forces(H, _, sim_specs):
    calc_status = 0

    # Parse out num particles, from generator function
    particles = str(int(H["x"][0][0]))

    # num particles, timesteps, also using num particles as seed
    args = particles + " " + str(10) + " " + particles

    # Retrieve our MPI Executor instance
    exctr = Executor.executor

    # Submit our forces app for execution
    task = exctr.submit(app_name="forces", app_args=args)

    # Block until the task finishes
    task.wait()

We retrieve the generated number of particles from H and construct an argument string for our launched application. The particle count doubles up as a random number seed here. Note a fourth argument can be added to forces that gives forces a chance of a “bad run” (a float between 0 and 1), but for now that will default to zero.

We then retrieve our previously instantiated Executor instance from the class definition, where it was automatically stored as an attribute.

After submitting the “forces” app for execution, a Task object is returned that correlates with the launched app. This object is roughly equivalent to a Python future, and can be polled, killed, and evaluated in a variety of helpful ways. For now, we’re satisfied with waiting for the task to complete via task.wait().

We can assume that afterward, any results are now available to parse. Our application produces a forces.stat file that contains either energy computations for every time-step or a “kill” message if particles were lost, which indicates a bad run - this can be ignored for now.

To complete our simulation function, parse the last energy value from the output file into a local output History array, and if successful, set the simulation function’s exit status calc_status to WORKER_DONE. Otherwise, send back NAN and a TASK_FAILED status:

    # Stat file to check for bad runs
    statfile = "forces.stat"

    # Try loading final energy reading, set the sim's status
    try:
        data = np.loadtxt(statfile)
        final_energy = data[-1]
        calc_status = WORKER_DONE
    except Exception:
        final_energy = np.nan
        calc_status = TASK_FAILED

    # Define our output array,  populate with energy reading
    outspecs = sim_specs["out"]
    output = np.zeros(1, dtype=outspecs)
    output["energy"][0] = final_energy

    # Return final information to worker, for reporting to manager
    return output, calc_status

calc_status will be displayed in the libE_stats.txt log file.

That’s it! As can be seen, with libEnsemble, it’s relatively easy to get started with launching applications. Behind the scenes, libEnsemble evaluates default MPI runners and available resources and divides them among the workers.

This completes our calling script and simulation function. Run libEnsemble with:

$ python run_libe_forces.py --comms local --nworkers [nworkers]

This may take up to a minute to complete. Output files—including forces.stat and files containing stdout and stderr content for each task—should appear in the current working directory. Overall workflow information should appear in libE_stats.txt and ensemble.log as usual.

For example, my libE_stats.txt resembled:

Worker     1: Gen no     1: gen Time: 0.001 Start: ... End: ... Status: Not set
Worker     1: sim_id     0: sim Time: 0.227 Start: ... End: ... Status: Completed
Worker     2: sim_id     1: sim Time: 0.426 Start: ... End: ... Status: Completed
Worker     1: sim_id     2: sim Time: 0.627 Start: ... End: ... Status: Completed
Worker     2: sim_id     3: sim Time: 0.225 Start: ... End: ... Status: Completed
Worker     1: sim_id     4: sim Time: 0.224 Start: ... End: ... Status: Completed
Worker     2: sim_id     5: sim Time: 0.625 Start: ... End: ... Status: Completed
Worker     1: sim_id     6: sim Time: 0.225 Start: ... End: ... Status: Completed
Worker     2: sim_id     7: sim Time: 0.626 Start: ... End: ... Status: Completed

Where status is set based on the simulation function’s returned calc_status.

My ensemble.log (on a ten-core laptop) resembled:

[0]  ... libensemble.libE (INFO): Logger initializing: [workerID] precedes each line. [0] = Manager
[0]  ... libensemble.libE (INFO): libE version v0.9.0
[0]  ... libensemble.manager (INFO): Manager initiated on node my_laptop
[0]  ... libensemble.manager (INFO): Manager exit_criteria: {"sim_max": 8}
[1]  ... libensemble.worker (INFO): Worker 1 initiated on node my_laptop
[2]  ... libensemble.worker (INFO): Worker 2 initiated on node my_laptop
[1]  ... libensemble.executors.mpi_executor (INFO): Launching task libe_task_forces_worker1_0: mpirun -hosts my_laptop -np 5 --ppn 5 /Users/.../forces.x 2023 10 2023
[2]  ... libensemble.executors.mpi_executor (INFO): Launching task libe_task_forces_worker2_0: mpirun -hosts my_laptop -np 5 --ppn 5 /Users/.../forces.x 2900 10 2900
[1]  ... libensemble.executors.executor (INFO): Task libe_task_forces_worker1_0 finished with errcode 0 (FINISHED)
[1]  ... libensemble.executors.mpi_executor (INFO): Launching task libe_task_forces_worker1_1: mpirun -hosts my_laptop -np 5 --ppn 5 /Users/.../forces.x 1288 10 1288
[2]  ... libensemble.executors.executor (INFO): Task libe_task_forces_worker2_0 finished with errcode 0 (FINISHED)
[2]  ... libensemble.executors.mpi_executor (INFO): Launching task libe_task_forces_worker2_1: mpirun -hosts my_laptop -np 5 --ppn 5 /Users/.../forces.x 2897 10 2897
[1]  ... libensemble.executors.executor (INFO): Task libe_task_forces_worker1_1 finished with errcode 0 (FINISHED)
[1]  ... libensemble.executors.mpi_executor (INFO): Launching task libe_task_forces_worker1_2: mpirun -hosts my_laptop -np 5 --ppn 5 /Users/.../forces.x 1623 10 1623
[2]  ... libensemble.executors.executor (INFO): Task libe_task_forces_worker2_1 finished with errcode 0 (FINISHED)
[2]  ... libensemble.executors.mpi_executor (INFO): Launching task libe_task_forces_worker2_2: mpirun -hosts my_laptop -np 5 --ppn 5 /Users/.../forces.x 1846 10 1846
[1]  ... libensemble.executors.executor (INFO): Task libe_task_forces_worker1_2 finished with errcode 0 (FINISHED)
[1]  ... libensemble.executors.mpi_executor (INFO): Launching task libe_task_forces_worker1_3: mpirun -hosts my_laptop -np 5 --ppn 5 /Users/.../forces.x 2655 10 2655
[2]  ... libensemble.executors.executor (INFO): Task libe_task_forces_worker2_2 finished with errcode 0 (FINISHED)
[2]  ... libensemble.executors.mpi_executor (INFO): Launching task libe_task_forces_worker2_3: mpirun -hosts my_laptop -np 5 --ppn 5 /Users/.../forces.x 1818 10 1818
[1]  ... libensemble.executors.executor (INFO): Task libe_task_forces_worker1_3 finished with errcode 0 (FINISHED)
[2]  ... libensemble.executors.executor (INFO): Task libe_task_forces_worker2_3 finished with errcode 0 (FINISHED)
[0]  ... libensemble.manager (INFO): Term test tripped: sim_max
[0]  ... libensemble.manager (INFO): Term test tripped: sim_max
[0]  ... libensemble.libE (INFO): Manager total time: 3.939

Note again that the ten cores were divided equally among two workers.

That concludes this tutorial. Each of these example files can be found in the repository in examples/tutorials/forces_with_executor.

For further experimentation, we recommend trying out this libEnsemble tutorial workflow on a cluster or multi-node system, since libEnsemble can also manage those resources and is developed to coordinate computations at huge scales. Please feel free to contact us or open an issue on GitHub if this tutorial workflow doesn’t work properly on your cluster or other compute resource.

Exercises

These may require additional browsing of the documentation to complete.

1. Adjust submit() to launch with four processes.

2. Adjust submit() again so the app’s stdout and stderr are written to stdout.txt and stderr.txt respectively.

3. Add a fourth argument to the args line to make 20% of simulations go bad.

4. Construct a while not task.finished: loop that periodically sleeps for a tenth of a second, calls task.poll(), then reads the output .stat file, and calls task.kill() if the output file contains "kill\n" or if task.runtime exceeds sixty seconds.

Click Here for Solution

import time

...
args = particles + " " + str(10) + " " + particles + " " + str(0.2)
...
statfile = "forces.stat"
task = exctr.submit(
    app_name="forces",
    app_args=args,
    num_procs=4,
    stdout="stdout.txt",
    stderr="stderr.txt",
)

while not task.finished:
    time.sleep(0.1)
    task.poll()

    if task.file_exists_in_workdir(statfile):
        with open(statfile, "r") as f:
            if "kill\n" in f.readlines():
                task.kill()

    if task.runtime > 60:
        task.kill()

...