Frequently Asked Questions

If you have any additional questions, feel free to contact us through Support.

Common Errors

“Manager only - must be at least one worker (2 MPI tasks)” when running with multiprocessing and multiple workers specified.

If your calling script code was recently switched from MPI to multiprocessing, make sure that libE_specs is populated with "comms": "local" and "nworkers": [int].

“AssertionError: alloc_f did not return any work, although all workers are idle.”

This error occurs when the manager is waiting although all workers are idle. Note that a worker can be in a persistent state but is marked as idle when it has returned data to the manager and is ready to receive work.

Some possible causes of this error are:

  • An MPI libEnsemble run was initiated with only one process, resulting in one manager but no workers. Similarly, the error may arise when running with only two processes when using a persistent generator. The generator will occupy one worker, leaving none to run simulation functions.

  • An error in the allocation function. For example, perhaps the allocation waiting for all requested evaluations to be returned (e.g, before starting a new generator), but this condition is not returning True even though all scheduled evaluations have returned. This can be due to incorrect implementation (e.g., it has not considered points that are cancelled or paused or in some other state that prevents the allocation function from sending them out to workers).

  • A persistent worker (usually a generator) has sent a message back to the manager but is still performing work and may return further points. In this case, consider starting the generator in active_recv mode. This can be specified in the allocation function and will cause the worker to maintain its active status.

  • A persistent worker has requested resources that prevents any simulations from taking place. By default, persistent workers hold onto resources even when not active. This may require the worker to return from persistent mode.

  • When returning points to a persistent generator (often the top code block in allocation functions). For example, support.avail_worker_ids(persistent=EVAL_GEN_TAG) Make sure that the EVAL_GEN_TAG is specified and not just persistent=True.

I keep getting: “Not enough processors per worker to honor arguments.” when using the Executor. Can I submit tasks to allocated processors anyway?

You may have set enforce_worker_core_bounds to True when setting up the Executor. Also, the resource manager can be completely disabled with:

libE_specs["disable_resource_manager"] = True

Note that the Executor submit() method has a parameter hyperthreads which will attempt to use all hyperthreads/SMT threads available if set to True.

FileExistsError: [Errno 17] File exists: “./ensemble”

This can happen when libEnsemble tries to create ensemble or simulation directories that already exist from previous runs. To avoid this, ensure the ensemble directory paths are unique by appending some unique value to libE_specs["ensemble_dir_path"]

PETSc and MPI errors with “[unset]: write_line error; fd=-1 buf=:cmd=abort exitcode=59”

with python [test with PETSc].py --comms local --nworkers 4

This error occurs on some platforms when using PETSc with libEnsemble in local (multiprocessing) mode. We believe this is due to PETSc initializing MPI before libEnsemble forks processes using multiprocessing. The recommended solution is running libEnsemble in MPI mode. An alternative solution may be using a serial build of PETSc.

Note

This error may depend on how multiprocessing handles an existing MPI communicator in a particular platform.

HPC Errors and Questions

Why does libEnsemble hang on certain systems when running with MPI?

Another symptom may be the manager only communicating with Worker 1. This issue may occur if matching probes, which mpi4py uses by default, are not supported by the communications fabric, like Intel’s Truescale (TMI) fabric. This can be solved by switching fabrics or disabling matching probes before the MPI module is first imported.

Add these two lines BEFORE from mpi4py import MPI:

import mpi4py
mpi4py.rc.recv_mprobe = False

Also see https://software.intel.com/en-us/articles/python-mpi4py-on-intel-true-scale-and-omni-path-clusters.

can’t open hfi unit: -1 (err=23) [13] MPI startup(): tmi fabric is not available and fallback fabric is not enabled

This may occur on TMI when libEnsemble Python processes have been launched to a node and these, in turn, execute tasks on the node; creating too many processes for the available contexts. Note that while processes can share contexts, the system is confused by the fact that there are two phases: first libEnsemble processes and then subprocesses to run user tasks. The solution is to either reduce the number of processes running or to specify a fallback fabric through environment variables:

unset I_MPI_FABRICS
export I_MPI_FABRICS_LIST=tmi,tcp
export I_MPI_FALLBACK=1

Alternatively, libEnsemble can be run in central mode where all workers run on dedicated nodes while launching all tasks onto other nodes. To do this add a node for libEnsemble, and add libE_specs["dedicated_mode"] = True to your calling script.

What does “_pickle.UnpicklingError: invalid load key, “x00”.” indicate?

This has been observed with the OFA fabric when using mpi4py and usually indicates MPI messages aren’t being received correctly. The solution is to either switch fabric or turn off matching probes. See the answer for “Why does libEnsemble hang on certain systems when running with MPI?”

For more information see https://bitbucket.org/mpi4py/mpi4py/issues/102/unpicklingerror-on-commrecv-after-iprobe.

srun: Job ****** step creation temporarily disabled, retrying (Requested nodes are busy)

You may also see: srun: Job ****** step creation still disabled, retrying (Requested nodes are busy)

When running on a SLURM system, this implies that you are trying to run on a resource that is already dedicated to another task. The reason can vary, some reasons are:

  • All the contexts are in use. This has occurred when using TMI fabric on clusters. See question can’t open hfi unit: -1 (err=23) for more info.

  • All the memory is assigned to the first job-step (srun application), due to a default exclusive mode scheduling policy. This has been observed on Perlmutter and SDF.

    In some cases using these environment variables will stop the issue:

    export SLURM_EXACT=1
export SLURM_MEM_PER_NODE=0

    Alternatively, this can be resolved by limiting the memory and other resources given to each task using the --exact option to srun along with other relevant options. For example:

    srun --exact -n 4 -c 1 --mem-per-cpu=4G

    would ensure that one CPU and 4 Gigabytes of memory are assigned to each MPI process. The amount of memory should be determined by the memory on the node divided by the number of CPUs. In the executor, this can be expressed via the extra_args option.

    If libEnsemble is sharing nodes with submitted tasks (user applications launched by workers), then you may need to do this for your launch of libEnsemble also, ensuring there are enough resources for both the libEnsemble manager and workers and the launched tasks. If this is complicated, we recommended using a dedicated node for libEnsemble.

libEnsemble Help

How can I debug specific libEnsemble processes?

This is most easily addressed when running libEnsemble locally. Try

mpiexec -np [num processes] xterm -e "python [calling script].py"

to launch an xterm terminal window specific to each process. Mac users will need to install xQuartz.

If running in local mode, try using one of the ForkablePdb routines in libensemble.tools to set breakpoints and debug similarly to pdb. How well this works varies by system.

from libensemble.tools import ForkablePdb
ForkablePdb().set_trace()

Can I use the MPI Executor when running libEnsemble with multiprocessing?

Yes. The Executor type determines only how libEnsemble workers execute and interact with user applications and is independent of comms chosen for manager/worker communications.

How can I disable libEnsemble’s output files?

To disable libe_stats.txt and ensemble.log, which libEnsemble typically always creates, set libE_specs["disable_log_files"] to True.

If libEnsemble aborts on an exception, the History array and persis_info dictionaries will be dumped. This can be suppressed by setting libE_specs["save_H_and_persis_on_abort"] to False.

See here for more information about these files.

How can I silence libEnsemble or prevent printed warnings?

Some logger messages at or above the MANAGER_WARNING level are mirrored to stderr automatically. To disable this, set the minimum stderr displaying level to CRITICAL via the following:

from libensemble import logger
logger.set_stderr_level("CRITICAL")

This effectively puts libEnsemble in silent mode.

See the Logger Configuration docs for more information.

macOS and Windows Errors

Can I run libEnsemble on Windows

Although we have run many libEnsemble workflows successfully on Windows using both MPI and local comms, Windows is not rigorously supported. We highly recommend Unix-like systems. Windows tends to produce more platform-specific issues that are difficult to reproduce and troubleshoot.

Feel free to check our Github Actions page to see what tests we run regularly on Windows.

Windows - How can I run libEnsemble with MPI comms?

We have run Windows workflows with MPI comms. However, as most MPI distributions have either dropped Windows support (MPICH and Open MPI) or are no longer being maintained (msmpi), we cannot guarantee success.

If users wish to try, we recommend experimenting with the many Unix-like emulators, containers, virtual machines, and other such systems. The Installing PETSc On Microsoft Windows documentation contains valuable information.

Otherwise, install msmpi and mpi4py from conda and experiment, or use local comms.

Windows - “A required privilege is not held by the client”

Assuming you were trying to use the sim_dir_symlink_files or gen_dir_symlink_files options, this indicates that to allow libEnsemble to create symlinks, you need to run your current cmd shell as administrator.

“RuntimeError: An attempt has been made to start a new process… this probably means that you are not using fork… ” if __name__ == “__main__”: freeze_support() …

You need to place your main calling script code underneath an if __name__ == "__main__": block.

Explanation: Python chooses one of three methods to start new processes when using multiprocessing (--comms local with libEnsemble). These are "fork", "spawn", and "forkserver". "fork" is the default on Unix, and in our experience is quicker and more reliable, but "spawn" is the default on Windows and macOS (See the Python multiprocessing docs).

Prior to libEnsemble v0.9.2, if libEnsemble detected macOS, it would automatically switch the multiprocessing method to "fork". We decided to stop doing this to avoid overriding defaults and compatibility issues with some libraries.

If you’d prefer to use "fork" or not reformat your code, you can set the multiprocessing start method by placing the following near the top of your calling script:

import multiprocessing
multiprocessing.set_start_method("fork", force=True)

“macOS - Fatal error in MPI_Init_thread: Other MPI error, error stack: … gethostbyname failed”

Resolve this by appending 127.0.0.1   [your hostname] to /etc/hosts. Unfortunately, 127.0.0.1   localhost isn’t satisfactory for preventing this error.

macOS - How do I stop the Firewall Security popups when running with the Executor?

There are several ways to address this nuisance, but all involve trial and error. An easy (but insecure) solution is temporarily disabling the firewall through System Preferences -> Security & Privacy -> Firewall -> Turn Off Firewall. Alternatively, adding a firewall “Allow incoming connections” rule can be attempted for the offending executable. We’ve had limited success running sudo codesign --force --deep --sign - /path/to/application.app on our executables, then confirming the next alerts for the executable and mpiexec.hydra.

Frozen PETSc installation following a failed wheel build with pip install petsc petsc4py

Following a failed wheel build for PETSc, the installation process may freeze when attempting to configure PETSc with the local Fortran compiler if it doesn’t exist. Run the above command again after disabling Fortran configuring with export PETSC_CONFIGURE_OPTIONS="--with-fc=0". The wheel build will still fail, but PETSc and petsc4py should still install successfully via setup.py after some time.