lxm3 0.4.2

LXM3: XManager launch backend for HPC clusters

LXM3: XManager launch backend for HPC clusters

lxm3 provides an implementation for DeepMind's XManager launch API that aims to provide a similar experience for running experiments on traditional HPC.

Currently, lxm3 provides a local execution backend and support for the SGE and Slurm schedulers.

NOTE: lxm3 is still in early development. The API is not stable and may change in the future. We periodically update tag versions which are considered in good shape. If you use lxm3 in your project, please pin to a specific commit/tag to avoid breaking changes.

Installation

For running on a cluster, you should install Singularity and rsync before using lxm3. It may be possible to run on the cluster without Singularity, but that path was not tested thoroughly.

You can install lxm3 from PyPI by running.

pip install lxm3

You can also install from GitHub for the latest features.

# Consider pinning to a specific commit/tag.
pip install git+https://github.com/ethanluoyc/lxm3

Documentation

At a high level you can launch experiment by creating a launch script called launcher.py that looks like:

with xm_cluster.create_experiment(experiment_title="hello world") as experiment:
    # Launch on a Slurm cluster
    executor = xm_cluster.Slurm()
    # or, if you want to use SGE:
    # executor = xm_cluster.GridEngine()
    # or, if you want to run locally:
    # executor = xm_cluster.Local()

    spec = xm_cluster.PythonPackage(
       path=".",
       entrypoint=xm_cluster.ModuleName("my_package.main"),
    )

    # package your code
    [executable] = experiment.package(
        [xm.Packageable(spec, executor_spec=executor.Spec())]
    )

    # add jobs to your experiment
    experiment.add(
        xm.Job(executable=executable, executor=executor)
    )

and launch the experimet from the command line with

lxm3 launch launcher.py

Many things happen under the hood. Since lxm3 implements the XManager API, you should get familiar with the concepts in the XManager. Once you are familiar with the concepts, checkout the examples/ directory for a quick start guide.

Components

lxm3 provides the following executable specification and executors.

Executable specifications

Name	Description
lxm3.xm_cluster.PythonPackage	A python application packageable with pip
lxm3.xm_cluster.UniversalPackage	A universal package
lxm3.xm_cluster.SingularityContainer	An executable running with Singularity

Executors

Name	Description
lxm3.xm_cluster.Local	Runs a executable locally, mainly used for testing
lxm3.xm_cluster.GridEngine	Runs a executable on SGE cluster
lxm3.xm_cluster.Slurm	Runs a executable on Slurm cluster

Jobs

• Currently, only xm.Job and xm.JobGenerator that generates xm.Job are supported.
• We support HPC array jobs via xm_cluster.ArrayJob. See below.

Implementation Details

Managing Dependencies with Containers

lxm3 uses of Singularity containers for running jobs on HPCs.

lxm3 aims at providing a easy workflow for launching jobs on traditional HPC clusters, which deviates from typical workflows for launching experiments on Cloud platforms.

lxm3 is designed for working with containerized applications using Singularity as the runtime. Singularity is a popular choice for HPC clusters because it allows users to run containers without requiring root privileges, and is supported by many HPC clusters worldwide.

There are many benefits to using containers for running jobs on HPCs compared to traditional isolation via venv or conda.

1. venv and conda are laid out as a directory of files on the environment. For many HPCs, normally these will be installed on a networked filesystem such as NFS. Operations on these virtual environments are slow and inefficient. For example, on our cluster, removing a conda environment with many dependencies can take an hour when these environments are on NFS. There are usually quota put in places not only for the file sizes but also the number of files. For ML projects that uses depends on many (large) packages such as TensorFlow, PyTorch, it is very easy to hit the quota limit. Singularity containers are a single file. This is both easy for deployment and also avoids the file number quota.
2. Containers provide consistent environment for running jobs on different clusters as well as making it easy to use system dependencies not installed on HPC's host environment.

Automated Deployment.

HPC deployments normally use a filesystem that are detached from the filesystems of the user's workstation. Many tutorials for running jobs on HPCs request the users to either clone their repository on the login node or ask the users manually copy files to the cluster. Doing this repeatedly is tedious. lxm3 automates the deployments from your workstation to the HPC cluster so that you can do most of your work locally without having directly login into the cluster.

Unlike Docker or other OCI images that are composed of multiple layers, the Singularity Image Format (SIF) used by Singularity is a single file that contains the entire filesystem of the container. While this is convenient as deployment to a remote cluster can be performed with a single scp/rsync command. The lack of layer caching/sharing makes repeated deployments slow and inefficient. For this reason, unlike typical cloud deployments where the application and dependencies are packaged into a single image, lxm3 uses a two-stage packaging process to separate the application and dependencies. This allows applications with heavy dependencies to be packaged once and reused across multiple experiments by reusing the same singularity container.

For Python applications, we rely on the user to first build a runtime image for all of the dependencies and use standard Python pacakging tools to create a distribution that is deployed separately to the cluster. Concretely, the user is expected to create a simple pyproject.toml which describes how to create a distribution for their applications. This is convenient, as lxm3 does not have to invent a custom packaging format for Python applications. For example, a simple pyproject.toml that uses hatchling as the build backend looks like:

[build-system]
requires = ["hatchling"]
build-backend = "hatchling.build"

[project]
name = "py_package"
authors = [{ name = "Joe" }]
version = "0.1.0"

lxm3 uses pip install --no-deps to create a zip archive that contains all of your application code. In addition to the packaging, lxm3 also allows you to carry extra files for your deployment via xm_cluster.Fileset. This is for example useful if you want to deploy configuration files.

The zip archive is automatically extracted into a temporary directory on the cluster and executed from there. Using a zip archive again minimizes the number of files that are deployed to the cluster so that you are less likely to hit a file number limit.

If you are using a different language or you cannot package your python application easily with standard Python packaging tools, you can use xm_cluster.UniversalPackage to package your application.

Easy Hyperparameter Sweeping.

For many scientific research projects, it's common to run the same experiment with different hyperparameters. lxm3 automatically generates jobs scripts that can be submitted to the cluster's scheduler for running multiple experiments with different hyperparameters passed as differnt command line arguments or environment variables.

For large parameter sweep, launching many separate jobs at once can overwhelm the scheduler. For this reason, HPC schedulers encourage the use of job arrays to submit sweeps. lxm3 provide a ArrayJob xm.JobConfig that allows you to submit multiple jobs with the same executable and job requirements but different hyperparameters as a single job array.

For example:

from lxm3 import xm
from lxm3 import xm_cluster
with xm_cluster.create_experiment() as experiment:
    executable = ...
    executor = ...
    parameters = [{"seed": seed} for seed in range(5)]
    experiment.add(
        xm_cluster.ArrayJob(executable=executable, executor=executor, args=parameters)
    )

This will be translated as passing --seed {0..4} to your executable. We also support customzing environment variables, which is convenient for example if you use Weights and Biases where you can configure run names and groups from environment variables (TODO(yl): migrate examples for configuring wandb).

There is a lot of flexibility on how to create the args for each job. For example, you can use itertools.product to create a cartesian product of all the hyperparameters. You can create arbitrary sweeps in pure python without resorting to a DSL.

Under the hood, lxm3 automatically generates job scripts that map from the array job index to command line arguments and environment variables.

Separate experiment launching from your application.

Similar to the design of XManager, lxm3 separates the launching of experiments from your application so that you are not bound to a specific experiment framework. You can develop your project with your favorite framework without having your application code be aware of the existence of lxm3. In fact, we recommend that you install lxm3 as a development dependency that are not bundled with your dependencies used at runtime. You can also install lxm3 globally or in its own virtual environment via pex or pipx.

Notes for existing Xmanager users

1. We vendored a copy of xmanager core API (v0.4.0) into lxm3 with light modification to support Python 3.9. This also allows us to just use the launch API without xm_local's dependencies. Thus, you should import the core API as from lxm3 import xm instead of from xmanager import xm. Our executables specs are defined in xm_cluster instead xm as we do not support the executable specs from the core API.