Skip to content

d-krupke/slurminade

Repository files navigation

slurminade - A decorator-based slurm runner for Python-code.

PyPI version CI status License

slurminade makes using the workload manager slurm with Python beautiful. It is based on simple_slurm, but instead of just allowing to comfortably execute shell commands in slurm, it allows to directly distribute Python-functions. A function decorated with @slurminade.slurmify(partition="alg") will automatically be executed by a node of the partition alg by just calling .distribute(yes_also_args_are_allowed). The general idea is that the corresponding Python-code exists on both machines, thus, the slurm-node can also call the functions of the original code if you tell if which one and what arguments to use. This is similar to celery but you do not need to install anything, just make sure the same Python-environment is available on the nodes (usually the case in a proper slurm setup).

Please check the documentation of simple_slurm to get to know more about the possible parameters. You can also call simple_slurm directly by srun and sbatch (automatically with the configuration specified with slurminade).

slurminade has two design goals:

  1. Pythonic slurm: Allowing to use slurm in a Pythonic-way, without any shell commands etc.
  2. Compatibility: Scripts can also run without slurm. You can share a script and also people without slurm can execute it without any changes.

We use it to empirically evaluate optimization algorithms for research papers on hundreds of instances that can require 15min each to solve. With slurminade, we can distribute the workload by just changing a few lines of code in our local Python scripts (those that you use for probing and development before running big experiments). An example of such a usage can be found here: Example of an empirical algorithm performance study for graph coloring heuristics using slurminade and AlgBench. You will find the original runner and the slurmified runner, showing the simplicity of distributing your experiments with slurminade.

A simple script could look like this:

import slurminade

slurminade.update_default_configuration(
    partition="alg", exclusive=True
)  # global options for slurm

# If no slurm environment is found, the functions are called directly to make scripts
# compatible with any environment.
# You can enforce slurm with `slurminade.set_dispatcher(slurminade.SlurmDispatcher())`


@slurminade.node_setup
def setup():
    print("I will run automatically on every slurm node at the beginning!")


# use this decorator to make a function distributable with slurm
@slurminade.slurmify(
    constraint="alggen02"
)  # function specific options can be specified
def prepare():
    print("Prepare")


@slurminade.slurmify()
def f(foobar):
    print(f"f({foobar})")


@slurminade.slurmify()
def clean_up():
    print("Clean up")


if __name__ == "__main__":
    prepare.distribute()
    slurminade.join()  # make sure that no job runs before prepare has finished
    with slurminade.JobBundling(max_size=20):  # automatically bundles up to 20 tasks
        # run 100x f after `prepare` has finished
        for i in range(100):
            f.distribute(i)

    slurminade.join()  # make sure that the clean up jobs runs after all f-jobs have finished
    clean_up.distribute()

If slurm is not available, distribute results in a local function call. Analogous for srun and sbatch (giving some extra value on top of just forwarding to simple_slurm).

Warning

Always use JobBundling when distributing many small tasks to few nodes. Slurm jobs have a certain overhead and you do not want to spam your infrastructure with too many jobs. However, function calls joined by JobBundling are considered as a single job by slurm, thus, not shared across nodes.

What are the limitations of slurminade? Slurminade reconstructs the environment by basically loading the code on the slurm node (without the __main__-part) and then calling the slurmified function with parameters serialized as JSONSs. This means that the code must be written in a common .py-file and all (distributed) function arguments must be JSON-serializable. Also, the function must not use any global state (e.g., global variables, file or database connections) initialized in the __main__-part. Additionally, the Python-environment must be available under the same path on the slurm node as slurminade will use the same paths on the slurm node to reconstruct the environment (allowing to use virtual environments).

Does slurminade work with Python 2? No, it is a Python 3 project. We tested it with Python 3.7 and higher.

Does slurminade work with Windows? Probably not, but I never saw a slurm cluster running on Windows. The (automatic) slurm-less mode should work on Windows. So your code will run, but all function calls will be local.

Are multi-file projects supported? Yes, as long as the files are available on the slurm node.

Does slurminade work with virtual environments? Yes. We recommend to use slurminade with conda. We have not tested it with other virtual environments.

Can I run my slurmified code outside a slurm environment? Yes, if you do not have slurm, the distributed functions are run as normal Python function calls. This means that you can share the same code with people that do not have slurm. It was important to us that the experimental evaluations we run on our slurm cluster can also be run in a common Python environment by reviewers without any changes.

Can I receive the return value of a slurmified function? No, the return value is not transmitted back to the caller. Note that the distribute-calls are non-blocking, i.e., the function returns immediately. Return values could be implemented via a Promise-object like for other distributed computing frameworks, but we did not see the need for it yet. We are usually saving the results in a database or files, e.g., using AlgBench.

Can I use command line arguments ``sys.argv`` in my scripts? Yes, but only in your __main__-part. The arguments are not transmitted to the slurm nodes as they are not part of the function call. You can add these as normal function arguments to your slurmified functions if needed. It is important that your global objects to not rely on these arguments for initialization, as the __main__-part is not executed on the slurm node. It is theoretically possible to transmit the arguments to the slurm node, but we did not see the need for it. Let us know if you need it and we may implement it.

The code is super simple and open source, don’t be afraid to create a fork that fits your own needs.

Note

Talk with you system administrator or supervisor to get the proper slurm configuration.

Installation

You can install slurminade with pip install slurminade.

Usage

You can set task specific slurm arguments within the decorator, e.g., @slurminade.slurmify(constraint="alggen03"). These arguments are directly passed to simple_slurm, such that all its arguments are supported.

In order for slurminade to work, the code needs to be in a Python file/project shared by all slurm-nodes. Otherwise, slurminade will not find the corresponding function. The slurmified functions also must be importable, i.e., on the top level. Currently, all function names must be unique as slurminade will only transmit the function’s name.

Don’t do:

Bad: Non blocking system calls

import slurminade
import os
import subprocess


@slurminade.slurmify()
def run_shell_command():
    # non-blocking system call
    subprocess.Popen("complex call")
    # BAD! The system call will run outside of slurm! The slurm task directly terminates.

instead use

import slurminade

if __name__ == "__main__":
    slurminade.sbatch(
        "complex call"
    )  # forwards your call to simple_slurm that is better used for such things.

Bad: Global variables in the __main__ part

import slurminade

FLAG = True


@slurminade.slurmify()
def bad_global(args):
    if FLAG:  # BAD! Will be True because the __main__ Part is not executed on the node.
        pass
    else:
        pass


if __name__ == "__main__":
    FLAG = False
    bad_global.distribute("args")

instead do

import slurminade


@slurminade.slurmify()
def bad_global(
    args, FLAG
):  # Now the flag is passed correctly as an argument. Note that only json-compatible arguments are possible.
    if FLAG:
        pass
    else:
        pass


# Without the `if`, the node would also execute this part (*slurminade* will abort automatically)
if __name__ == "__main__":
    FLAG = False
    bad_global.distribute("args", FLAG)

Warning

The same is true for any global state such as file or database connections. You can use global variables, but be wary of side effects.

Error: Complex objects as arguments

import slurminade


@slurminade.slurmify()
def sec_order_func(func):
    func()


def f():
    print("hello")


def g():
    print("world!")


if __name__ == "__main__":
    sec_order_func.distribute(f)  # will throw an exception
    sec_order_func.distribute(g)

Instead, create individual slurmified functions for each call or pass a simple identifier that lets the function deduce, what to do, e.g., a switch-case. If you really need to pass complex objects, you could also pickle the object and only pass the file name.

Default configuration

You can set up a default configuration in ~/.slurminade_default.json. This should simply be a dictionary of arguments for simple_slurm. For example

{
  "partition": "alg"
}

The current version checks the following files and overwrites values in the following order:

  1. ~/.slurminade_default.json
  2. ~/$XDG_CONFIG_HOME/slurminade/.slurminade_default.json
  3. ./.slurminade_default.json

Debugging

You can use

import slurminade

slurminade.set_dispatcher(slurminade.TestDispatcher())

to see the serialization or

import slurminade

slurminade.set_dispatcher(slurminade.SubprocessDispatcher())

to distribute the tasks without slurm using subprocesses.

If there is a bug, you will directly see it in the output (at least for most bugs).

Project structure

The project is reasonably easy:

  • bundling.py: Contains code for bundling tasks, so we don’t spam slurm with too many.
  • conf.py: Contains code for managing the configuration of slurm.
  • dispatcher.py: Contains code for actually dispatching tasks to slurm.
  • execute.py: Contains code to execute the task on the slurm node.
  • function.py: Contains the code for making a function slurm-compatible.
  • function_map.py: Saves all the slurmified functions.
  • guard.py: Contains code to prevent you accidentally DDoSing your infrastructure.
  • options.py: Contains a simple data structure to save slurm options.

Changes

  • 1.1.2: Fixing some return types for job bundling. Still not perfect. Be aware of potentially breaking changes if you have been using the job ids.
  • 1.1.1: Fixing bug when there is some output to stdout when loading the code, such as deprecation warnings.
  • 1.1.0: Slurminade can now be called from iPython, too! exec has been renamed shell to prevent confusion with the Python call exec which will evaluate a string as Python code.
  • 1.0.1: Dispatcher now return jobs references instead of job ids. This allows to do some fancier stuff in the future, when the jobs infos are only available a short time after the job has been submitted.
  • 0.10.1: FIX: Listing functions will no longer execute setup functions.
  • 0.10.0: Batch is now named JobBundling. There is a method join for easier synchronization. exec allows to executed commands just like srun and sbatch, but uniform syntax with other slurmified functions. Functions can now also be called with distribute_and_wait. If you call python3 -m slurminade.check --partition YOUR_PARTITION --constraint YOUR_CONSTRAINT you can check if your slurm configuration is running correctly.
  • 0.9.0: Lots of improvements.
  • 0.8.1: Bugfix and automatic detection of wrong usage when using Batch with wait_for.
  • 0.8.0: Added extensive logging and improved typing.
  • 0.7.0: Warning if a Batch is flushed multiple times, as we noticed this to be a common indentation error.
  • 0.6.2: Fixes recursive distribution guard, which seemed to be broken.
  • 0.6.1: Bugfixes in naming
  • 0.6.0: Autmatic naming of tasks.
  • 0.5.5: Fixing bug guard bug in subprocess dispatcher.
  • 0.5.4: Dispatched function calls that are too long for the command line now use a temporary file instead.
  • 0.5.3: Fixed a bug that caused the dispatch limit to have no effect.
  • 0.5.2: Added pyproject.toml for PEP compliance
  • 0.5.1: Batch will now flush on delete, in case you forgot.
  • 0.5.0:
    • Functions now have a wait_for-option and return job ids.
    • Braking changes: Batches have a new API.
      • add is no longer needed.
      • AutoBatch is now called Batch.
    • Fundamental code changes under the hood.
  • <0.5.0:
    • Lots of experiments on finding the right interface.

Contributors

This project is developed at the Algorithms Group at TU Braunschweig, Germany. The lead developer is Dominik Krupke. Further contributors are Matthias Konitzny and Patrick Blumenberg.

Similar Projects

  • This project is greatly inspired by Celery, but does not require any additional infrastructure except for slurm.
  • If you want a more powerful library to, e.g., also distribute lambdas or functions with complex arguments, check out submitit. It is a great project, and we may use it as a backend in the future. However, it does not support the slurm-less mode and can easily hide non-deterministic errors. Slurminade on the other hand is restricted on purpose to write reproducible scripts that can also be run without slurm.