FAQ

Here is a list of questions we have either been asked by users or potential pitfalls we hope to help users avoid:

Q: It doesn’t work, what now?

A: Check the log-files that are created and look for an error message. If you suspect a bug in adaptive_scheduler check out run_manager.task_status() and if that doesn’t reveal anything, open an issue on GitHub.

Q: What if I have more learners than cores?

A: You can distribute all learners in a certain amount of adaptive.BalancingLearners. Like so

from functools import partial

import adaptive
import numpy as np
from adaptive_scheduler.utils import split_in_balancing_learners, shuffle_list

def jacobi(x, n, alpha, beta):
    from scipy.special import eval_jacobi
    return eval_jacobi(n, alpha, beta, x)

combos = adaptive.utils.named_product(
        n=[1, 2, 4, 8],
    alpha=np.linspace(0, 2, 21),
    beta=np.linspace(0, 1, 21),
)

learners = [adaptive.Learner1D(partial(jacobi, **combo), bounds=(0, 1)) for combo in combos]
fnames = [f"data/jacobi/{combo}" for combo in combos]

# shuffle the learners (and fnames in the same order) because
# some learners might be slower than others (not in this example).
unshuffled = learners, fnames  # to have a handle to the unshuffled list
learners, fnames = shuffle_list(*unshuffled)

# split in many new BalancingLearners
# `learners` will be a list of BalancingLeaners
# `fnames` will be a list of lists with fnames
learners, fnames = split_in_balancing_learners(
        learners,
    fnames,
    n_parts=100,  # split into 100 BalancingLeaners
    strategy="npoints"
)

Q: Why aren’t my jobs dying when I cancel the job manager?

A: The job manager just starts the jobs and you want the job to keep running in case the job manager somehow dies. So you still need to scancel or qdel them in case you want to really cancel the jobs or call adaptive_scheduler.cancel_jobs with job_names from your Python environment.

Q: How do I set extra SBATCH/PBS arguments or environment variables in my job script?

A: You can change this in the scheduler object. For example modifying a job script for SLURM:

from adaptive_scheduler.scheduler import SLURM
scheduler = SLURM(
        cores=10,
    extra_scheduler=["--exclusive=user", "--time=1"],
    extra_env_vars=["TMPDIR='/scratch'", "PYTHONPATH='my_dir:$PYTHONPATH'"],
    mpiexec_executable="srun --mpi=pmi2",
)  # pass this to `server_support.start_job_manager` or `RunManager`

# see the job script with
print(scheduler.job_script('this_will_be_the_job_name'))

Q: My code uses MPI so the ~mpi4py.futures.MPIPoolExecutor won’t work for me, I want to use ipyparallel, how?

A: You just have to pass executor_type="ipyparallel" to ~adaptive_scheduler.scheduler.SLURM or ~adaptive_scheduler.scheduler.PBS. For example:

from adaptive_scheduler.scheduler import SLURM

scheduler = SLURM(
    cores=48,
    executor_type="ipyparallel",
)

run_manager = adaptive_scheduler.server_support.RunManager(
        scheduler=scheduler,
    learners=learners,
    fnames=fnames,

)
run_manager.start()

Q: ipyparallel doesn’t work for me, I want to use process-pool, how?

A: Sometimes ipyparallel doesn’t import modules correctly on its workers. In this case you can use process-pool. You just have to pass executor_type="process-pool" to ~adaptive_scheduler.scheduler.SLURM or ~adaptive_scheduler.scheduler.PBS. Note the process-pool uses Python’s ~concurrent.futures.ProcessPoolExecutor for parallelism and cannot be used beyond a single machine (for one learner).

Q: Cool! What else should I check out?

A: There are a bunch of things that are not present in the example notebook, I recommend to take a look at:

• adaptive_scheduler.utils.combo_to_fname

• adaptive_scheduler.server_support.cleanup

• adaptive_scheduler.server_support.parse_log_files

• adaptive_scheduler.utils.load_parallel and adaptive_scheduler.utils.save_parallel