Ensemble with an MPI Application

This tutorial highlights libEnsemble’s capability to portably execute and monitor external scripts or user applications within simulation or generator functions using the executor.

The 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 the result.

This tutorial uses libEnsemble’s MPI Executor, which automatically detects available MPI runners and resources.

This example also uses a persistent generator. This generator runs on a worker throughout the ensemble, producing new simulation parameters as requested.

Getting Started

The simulation source code forces.c can be obtained directly from the libEnsemble repository in the forces_app directory.

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

Alternative build lines for different platforms can be found in the build_forces.sh file in the same directory.

Calling Script

Complete scripts for this example can be found in the forces_simple directory.

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

import os
import sys

import numpy as np
from forces_simf import run_forces  # Sim func from current dir

from libensemble import Ensemble
from libensemble.alloc_funcs.start_only_persistent import only_persistent_gens as alloc_f
from libensemble.executors import MPIExecutor
from libensemble.gen_funcs.persistent_sampling import persistent_uniform as gen_f
from libensemble.specs import AllocSpecs, ExitCriteria, GenSpecs, LibeSpecs, SimSpecs

if __name__ == "__main__":
    # Initialize MPI Executor
    exctr = MPIExecutor()

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

    if not os.path.isfile(sim_app):
        sys.exit("forces.x not found - please build first in ../forces_app dir")

    exctr.register_app(full_path=sim_app, app_name="forces")

    # Parse number of workers, comms type, etc. from arguments
    ensemble = Ensemble(parse_args=True, executor=exctr)

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

The last line in the above codeblock initializes the ensemble. The parse_args parameter is used to read comms and nworkers from the command line. This sets the respective libE_specs options.

Next, we will add basic configuration for the ensemble. As one worker will run a persistent generator, we calculate the number of workers that need resources to run simulations. We also set sim_dirs_make so that a directory is created for each simulation. This helps organize output and also helps prevent workers from overwriting previous results.

    nsim_workers = ensemble.nworkers - 1  # One worker is for persistent generator

    # Persistent gen does not need resources
    ensemble.libE_specs = LibeSpecs(
        num_resource_sets=nsim_workers, sim_dirs_make=True, ensemble_dir_path="./test_executor_forces_tutorial"
    )

Next we define the sim_specs and gen_specs. Recall that these are used to specify to libEnsemble what user functions and input/output fields to expect, and also to parameterize user functions:

    ensemble.sim_specs = SimSpecs(
        sim_f=run_forces,
        inputs=["x"],
        outputs=[("energy", float)],
    )

    ensemble.gen_specs = GenSpecs(
        gen_f=gen_f,
        inputs=[],  # No input when starting persistent generator
        persis_in=["sim_id"],  # Return sim_ids of evaluated points to generator
        outputs=[("x", float, (1,))],
        user={
            "initial_batch_size": nsim_workers,
            "lb": np.array([1000]),  # min particles
            "ub": np.array([3000]),  # max particles
        },
    )  # gen_specs_end_tag

Next, configure an allocation function, which starts the one persistent generator and farms out the simulations. We also tell it to wait for all simulations to return their results, before generating more parameters.

    ensemble.alloc_specs = AllocSpecs(
        alloc_f=alloc_f,
        user={
            "async_return": False,  # False causes batch returns
        },
    )

Now we set exit_criteria to exit after running eight simulations.

We also give each worker a seeded random stream, via the persis_info option. These can be used for random number generation if required.

Finally we run the ensemble.

    # Instruct libEnsemble to exit after this many simulations
    ensemble.exit_criteria = ExitCriteria(sim_max=8)

    # Seed random streams for each worker, particularly for gen_f
    ensemble.add_random_streams()

    # Run ensemble
    ensemble.run()

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:

Set libEnsemble’s logger to print debug messages.

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

Tell the allocation function to return results to the generator asynchronously.

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

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

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


def run_forces(H, persis_info, sim_specs, libE_info):
    """Runs the forces MPI application"""

    calc_status = 0

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

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

    # Retrieve our MPI Executor
    exctr = libE_info["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.

We then retrieve our previously instantiated Executor. libEnsemble will use the MPI runner detected (or provided by platform options). As num_procs (or similar) is not specified, libEnsemble will assign the processors available to this worker.

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 timestep 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:

    # Try loading final energy reading, set the sim's status
    statfile = "forces.stat"
    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
    output = np.zeros(1, dtype=sim_specs["out"])
    output["energy"] = final_energy

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

Running the example

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

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

where nworkers is one more than the number of concurrent simulations.

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.

Example run / output

For example, after running:

python run_libe_forces.py --comms local --nworkers 3

my libE_stats.txt resembled:

Manager     : Starting ensemble at: 2023-09-12 18:12:08.517
Worker     2: sim_id     0: sim Time: 0.205 Start: ... End: ... Status: Completed
Worker     3: sim_id     1: sim Time: 0.284 Start: ... End: ... Status: Completed
Worker     2: sim_id     2: sim Time: 0.117 Start: ... End: ... Status: Completed
Worker     3: sim_id     3: sim Time: 0.294 Start: ... End: ... Status: Completed
Worker     2: sim_id     4: sim Time: 0.124 Start: ... End: ... Status: Completed
Worker     3: sim_id     5: sim Time: 0.174 Start: ... End: ... Status: Completed
Worker     3: sim_id     7: sim Time: 0.135 Start: ... End: ... Status: Completed
Worker     2: sim_id     6: sim Time: 0.275 Start: ... End: ... Status: Completed
Worker     1: Gen no     1: gen Time: 1.038 Start: ... End: ... Status: Persis gen finished
Manager     : Exiting ensemble at: 2023-09-12 18:12:09.565 Time Taken: 1.048

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

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

[0]  ... libensemble.libE (INFO): Logger initializing: [workerID] precedes each line. [0] = Manager
[0]  ... libensemble.libE (INFO): libE version v0.10.2+dev
[0]  ... libensemble.manager (INFO): Manager initiated on node shuds
[0]  ... libensemble.manager (INFO): Manager exit_criteria: {'sim_max': 8}
[2]  ... libensemble.worker (INFO): Worker 2 initiated on node shuds
[3]  ... libensemble.worker (INFO): Worker 3 initiated on node shuds
[1]  ... libensemble.worker (INFO): Worker 1 initiated on node shuds
[2]  ... libensemble.executors.mpi_executor (INFO): Launching task libe_task_forces_worker2_0: mpirun -hosts shuds -np 2 --ppn 2 /home/.../forces_app/forces.x 2023 10 2023
[3]  ... libensemble.executors.mpi_executor (INFO): Launching task libe_task_forces_worker3_0: mpirun -hosts shuds -np 2 --ppn 2 /home/.../forces_app/forces.x 2900 10 2900
[2]  ... libensemble.executors.executor (INFO): Task libe_task_forces_worker2_0 finished with errcode 0 (FINISHED)
[3]  ... libensemble.executors.executor (INFO): Task libe_task_forces_worker3_0 finished with errcode 0 (FINISHED)
[2]  ... libensemble.executors.mpi_executor (INFO): Launching task libe_task_forces_worker2_1: mpirun -hosts shuds -np 2 --ppn 2 /home/.../forces_app/forces.x 1288 10 1288
[3]  ... libensemble.executors.mpi_executor (INFO): Launching task libe_task_forces_worker3_1: mpirun -hosts shuds -np 2 --ppn 2 /home/.../forces_app/forces.x 2897 10 2897
[2]  ... libensemble.executors.executor (INFO): Task libe_task_forces_worker2_1 finished with errcode 0 (FINISHED)
[3]  ... libensemble.executors.executor (INFO): Task libe_task_forces_worker3_1 finished with errcode 0 (FINISHED)
[2]  ... libensemble.executors.mpi_executor (INFO): Launching task libe_task_forces_worker2_2: mpirun -hosts shuds -np 2 --ppn 2 /home/.../forces_app/forces.x 1623 10 1623
[3]  ... libensemble.executors.mpi_executor (INFO): Launching task libe_task_forces_worker3_2: mpirun -hosts shuds -np 2 --ppn 2 /home/.../forces_app/forces.x 1846 10 1846
[2]  ... libensemble.executors.executor (INFO): Task libe_task_forces_worker2_2 finished with errcode 0 (FINISHED)
[3]  ... libensemble.executors.executor (INFO): Task libe_task_forces_worker3_2 finished with errcode 0 (FINISHED)
[2]  ... libensemble.executors.mpi_executor (INFO): Launching task libe_task_forces_worker2_3: mpirun -hosts shuds -np 2 --ppn 2 /home/.../forces_app/forces.x 2655 10 2655
[3]  ... libensemble.executors.mpi_executor (INFO): Launching task libe_task_forces_worker3_3: mpirun -hosts shuds -np 2 --ppn 2 /home/.../forces_app/forces.x 1818 10 1818
[3]  ... libensemble.executors.executor (INFO): Task libe_task_forces_worker3_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: 1.043

Note again that the four cores were divided equally among two workers that run simulations.

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. See HPC platform guides for more information.

See the forces_gpu tutorial for a similar workflow including GPUs. That tutorial also shows how to dynamically assign resources to each simulation.

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.

Adjust submit() to launch with four processes.

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

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

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.

Running the generator on the manager

As of version 1.3.0, the generator can be run on a thread on the manager, using the libE_specs option gen_on_manager.

Change the libE_specs as follows.

nsim_workers = ensemble.nworkers

# Persistent gen does not need resources
ensemble.libE_specs = LibeSpecs(
    gen_on_manager=True,
    sim_dirs_make=True,
    ensemble_dir_path="./test_executor_forces_tutorial",
)

When running set nworkers to the number of workers desired for running simulations. E.g., Instead of:

python run_libe_forces.py --comms local --nworkers 5

use:

python run_libe_forces.py --comms local --nworkers 4

Note that as the generator random number seed will be zero instead of one, the checksum will change.

For more information see Running generator on the manager.

Running forces application with input file

Many applications read an input file instead of being given parameters directly on the run line.

forces_simple_with_input_file directory contains a variant of this example, where a templated input file is parameterized for each evaluation.

This requires jinja2 to be installed:

pip install jinja2

The file forces_input contains the following (remember we are using particles as seed also for simplicity):

num_particles = {{particles}}
num_steps = 10
rand_seed = {{particles}}

libEnsemble will copy this input file to each simulation directory.

The sim_f uses the following function to customize the input file with the parameters for the current simulation.

def set_input_file_params(H, sim_specs, ints=False):
    """
    This is a general function to parameterize the input file with any inputs
    from sim_specs["in"]

    Often sim_specs_in["x"] may be multi-dimensional, where each dimension
    corresponds to a different input name in sim_specs["user"]["input_names"]).
    Effectively an unpacking of "x"
    """
    input_file = sim_specs["user"]["input_filename"]
    input_values = {}
    for i, name in enumerate(sim_specs["user"]["input_names"]):
        value = int(H["x"][0][i]) if ints else H["x"][0][i]
        input_values[name] = value
    with open(input_file, "r") as f:
        template = jinja2.Template(f.read())
    with open(input_file, "w") as f:
        f.write(template.render(input_values))

This is called in the simulation function as follows.

def run_forces(H, persis_info, sim_specs, libE_info):
    """Runs the forces MPI application reading input from file"""

    calc_status = 0

    input_file = sim_specs["user"]["input_filename"]
    set_input_file_params(H, sim_specs, ints=True)

    # Retrieve our MPI Executor
    exctr = libE_info["executor"]

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

    # Block until the task finishes
    task.wait(timeout=60)

Notice that we convert the parameters to integers in this example.

The calling script then specifies the templated input file as follows.

input_file = "forces_input"

# Persistent gen does not need resources
ensemble.libE_specs = LibeSpecs(
    num_resource_sets=nsim_workers,
    sim_dirs_make=True,
    sim_dir_copy_files=[input_file],
)

ensemble.sim_specs = SimSpecs(
    sim_f=run_forces,
    inputs=["x"],
    outputs=[("energy", float)],
    user={"input_filename": input_file, "input_names": ["particles"]},
)

Line 36 tells the templated input file to be copied to each simulation directory.

An alternative is to use sim_input_dir, which gives the name of a directory that may contain multiple files and will be used as the base of each simulation directory.

Line 43 gives the input file name and the name of each parameter to the simulation function.

Multiple parameters

In our case, the only parameter name is x. However, in some cases, x (as defined by sim_specs["in"]) may be multi-dimensional, where each component has a different parameter name (e.g., “x”, “y”). For example, if the input file were:

num_particles = {{particles}}
num_steps = {{nsteps}}
rand_seed = {{seed}}

then line 43 would be:

user = {"input_filename": input_file, "input_names": ["particles", "nsteps", "seed"]}

and gen_specs would contain something similar to:

ensemble.gen_specs = GenSpecs(
    gen_f=gen_f,
    inputs=[],
    persis_in=["sim_id"],
    outputs=[("x", float, 3)],
    ...,
)

libEnsemble uses a convention of a multi-dimensional x in generator functions. However, these parameters can also be specified as different variables with corresponding modification to generator and simulator functions.