Skip to main content

Launchpad is a library that simplifies writing distributed programs and seamlessly launching them on a range of supported platforms.

Project description

Launchpad

PyPI - Python Version PyPI version

Launchpad is a library that simplifies writing distributed programs by seamlessly launching them on a variety of different platforms. Switching between local and distributed execution requires only a flag change.

Launchpad introduces a programming model that represents a distributed system as a graph data structure (a Program) describing the system’s topology. Each node in the program graph represents a service in the distributed system, i.e. the fundamental unit of computation that we are interested in running. As nodes are added to this graph, Launchpad constructs a handle for each of them. A handle ultimately represents a client to the yet-to-be-constructed service. A directed edge in the program graph, representing communication between two services, is created when the handle associated with one node is given to another at construction time. This edge originates from the receiving node, indicating that the receiving node will be the one initiating communication. This process allows Launchpad to define cross-service communication simply by passing handles to nodes. The open-sourced version of Launchpad currently provides following types of nodes (you can implement your own types as needed):

  • PyNode - a simple node executing provided Python code upon entry. It is similar to a main function, but with the distinction that each node may be running in separate processes and on different machines.
  • CourierNode - it enables cross-node communication. CourierNodes can communicate by calling public methods on each other either synchronously or asynchronously via futures. The underlying remote procedure calls are handled transparently by Launchpad.
  • ReverbNode - it exposes functionality of Reverb, an easy-to-use data storage and transport system primarily used by RL algorithms as an experience replay. You can read more about Reverb here.
  • MultiThreadingColocation - allows to colocate multiple other nodes in a single process.

Using Launchpad involves implementing nodes and defining the topology of your distributed program by passing to each node references of the other nodes that it can communicate with. The core data structure dealing with this is called a Launchpad program, which can then be used for local debugging runs, tests, distributed executions, etc.

Table of Contents

Installation

Please keep in mind that Launchpad is not hardened for production use, and while we do our best to keep things in working order, things may break or segfault.

:warning: Launchpad currently only supports Linux based OSes.

The recommended way to install Launchpad is with pip. We also provide instructions to build from source using the same docker images we use for releases.

TensorFlow can be installed separately or as part of the pip install. Installing TensorFlow as part of the install ensures compatibility.

$ pip install dm-launchpad[tensorflow]

# Without Tensorflow install and version dependency check.
$ pip install dm-launchpad

Nightly builds

PyPI version

$ pip install dm-launchpad-nightly[tensorflow]

# Without Tensorflow install and version dependency check.
$ pip install dm-launchpad-nightly

Similarily, Reverb can be installed ensuring compatibility:

$ pip install dm-launchpad[reverb]

Develop Launchpad inside a docker container

The most convenient way to develop Launchpad is with Docker. This way you can compile and test Launchpad inside a container without having to install anything on your host machine, while you can still use your editor of choice for making code changes. The steps are as follows.

Checkout Launchpad's source code from GitHub.

$ git checkout https://github.com/deepmind/launchpad.git
$ cd launchpad

Build the Docker container to be used for compiling and testing Launchpad. You can specify tensorflow_pip parameter to set the version of Tensorflow to build against. You can also specify which version(s) of Python container should support. The command below enables support for Python 3.6, 3.7, 3.8 and 3.9.

$ docker build --tag launchpad:devel \
  --build-arg tensorflow_pip=tensorflow==2.3.0 \
  --build-arg python_version="3.6 3.7 3.8 3.9" - < docker/build.dockerfile

The next step is to enter the built Docker image, binding checked out Launchpad's sources to /tmp/launchpad within the container.

$ docker run --rm --mount "type=bind,src=$PWD,dst=/tmp/launchpad" \
  -it launchpad:devel bash

At this point you can build and install Launchpad within the container by executing:

$ /tmp/launchpad/oss_build.sh

By default it builds Python 3.8 version, you can change that with --python flag.

$ /tmp/launchpad/oss_build.sh --python 3.6

To make sure installation was successful and Launchpad works as expected, you can run some examples provided:

$ python3.6 -m launchpad.examples.hello_world.launch
$ python3.6 -m launchpad.examples.consumer_producers.launch --lp_launch_type=local_mp

To make changes to Launchpad codebase, edit sources checked out from GitHub directly on your host machine (outside of the Docker container). All changes are visible inside the Docker container. To recompile just run the oss_build.sh script again from the Docker container. In order to reduce compilation time of the consecutive runs, make sure to not exit the Docker container.

Quick Start

The complete implementation can be found here.

Implement example nodes

In this producer-consumer example, we have one consumer and multiple producers. The consumer sends work to the producers, which perform some time-intensive task before returning the result to the consumer. Finally, the consumer summarizes the work done by all of the producers.

consumer-producer-topology

The producer in this example has just one method which performs some work (for you to implement) in a given context provided by the caller. Any method of the class can be exposed for other nodes to call by wrapping a node with a CourierNode. In a typical setup, all nodes live in separate processes or on distinct machines, while the communication between the nodes is taken care of transparently by Launchpad. Some care has to be taken though. For example, the work() method may be called from multiple threads within the same process, so if the producer were to have any shared state then access to it must be made thread-safe. In this case, the producer is stateless so it is not a concern.

class Producer:
  def work(self, context):
    return context

The consumer defines an initializer and a run() method. The initializer takes a list of handles to the producers (CourierNodes).

Any Launchpad PyClassNode with a run() method will have that method called automatically upon program entry. Here the run() method simply calls work() on each producer and collects the results. At the end, it calls launchpad.stop() to terminate all nodes running within a program.

class Consumer:
  def __init__(self, producers):
    self._producers = producers

  def run(self):
    results = [producer.work(context)
               for context, producer in enumerate(self._producers)]
    logging.info('Results: %s', results)
    lp.stop()

In the example above, work() methods are called sequentially, so there is no benefit in running this program distributed. Launchpad, however, allows for asynchronous calls as well through the use of futures. In the example below all producers will perform their work in parallel while consumer waits on all of their results when calling future.result().

class Consumer:
  def __init__(self, producers):
    self._producers = producers

  def run(self):
    futures = [producer.futures.work(context)
               for context, producer in enumerate(self._producers)]
    results = [future.result() for future in futures]
    logging.info('Results: %s', results)
    launchpad.stop()

Define the topology

The next step is to instantiate nodes for the consumer and producers and then connect them so that the consumer can call methods on the producers. The connections between nodes define the topology of the distributed program.

Launchpad uses an lp.Program class to hold all the nodes. There are several different types of nodes but here lp.CourierNode is used since it is the simplest type which supports communication between nodes. The parameters to lp.CourierNode are the name of the class and the parameters of its initializer. Connecting the consumer node to the producer nodes is as simple as passing in handles to all producers in the initializer of the consumer. The handles themselves are returned by lp.Program.add_node().

def make_program(num_producers):
  program = lp.Program('consumer_producers')
  with program.group('producer'):
    producers = [
        program.add_node(lp.CourierNode(Producer)) for _ in range(num_producers)
    ]
  node = lp.CourierNode(
      Consumer,
      producers=producers)
  program.add_node(node, label='consumer')
  return program

With the above function defining the topology all that remains is to implement main() for Launchpad:

def main(_):
  program = make_program(num_producers=FLAGS.num_producers)
  lp.launch(program)

if __name__ == '__main__':
  app.run(main)

Launch the program

To launch the program (assuming it is called launch.py), simply run:

python3 -m launch --lp_launch_type=local_mp

The --lp_launch_type controls how the program is launched. In the above case it is launched locally with each node executed in a separate process. List of supported execution modes can be found here.

Add a test

Here are some points to keep in mind when creating a test for a Launchpad program.

  • The easiest way to add a test for your program is to reuse the same topology for an integration test (i.e. call make_program() from above in this example).
  • Launch a test by calling lp.launch() just like in main() in the above example, but explicitly specify launch_type='test_mt' (multithreaded tests) as a parameter.
  • It is possible to disable automatic execution of a node's run() method before launching. Do so by calling disable_run() on the node in question.
  • In order to call methods to test on a Courier node you will need to explicitly dereference the handle of the node first. Do so by calling create_handle() followed by dereference() on the node in question.

Below is an incomplete example illustrating the above concepts. A complete example can be found here.

import launchpad as lp
from launchpad.examples.consumer_producers import launch
from absl.testing import absltest

class LaunchTest(absltest.TestCase):
  def test_consumer(self):
    program = launch.make_program(num_producers=2)
    (consumer_node,) = program.groups['consumer']
    consumer_node.disable_run()
    lp.launch(program, launch_type='test_mt')
    consumer = consumer_node.create_handle().dereference()
    # Perform actual test here by calling methods on `consumer` ...

Citing Launchpad

If you use Launchpad in your work, please cite the accompanying technical report:

@article{yang2021launchpad,
    title={Launchpad: A Programming Model for Distributed Machine Learning
           Research},
    author={Fan Yang and Gabriel Barth-Maron and Piotr Stańczyk and Matthew
            Hoffman and Siqi Liu and Manuel Kroiss and Aedan Pope and Alban
            Rrustemi},
    year={2021},
    journal={arXiv preprint arXiv:2106.04516},
    url={https://arxiv.org/abs/2106.04516},
}

Acknowledgements

We greatly appreciate all the help from Reverb and TF-Agents teams in setting up building and testing setup for Launchpad.

Project details


Release history Release notifications | RSS feed

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distributions

No source distribution files available for this release.See tutorial on generating distribution archives.

Built Distributions

dm_launchpad_nightly-0.3.0.dev20210622-cp39-cp39-manylinux2010_x86_64.whl (4.1 MB view details)

Uploaded CPython 3.9 manylinux: glibc 2.12+ x86-64

dm_launchpad_nightly-0.3.0.dev20210622-cp38-cp38-manylinux2010_x86_64.whl (4.1 MB view details)

Uploaded CPython 3.8 manylinux: glibc 2.12+ x86-64

dm_launchpad_nightly-0.3.0.dev20210622-cp37-cp37m-manylinux2010_x86_64.whl (4.1 MB view details)

Uploaded CPython 3.7m manylinux: glibc 2.12+ x86-64

dm_launchpad_nightly-0.3.0.dev20210622-cp36-cp36m-manylinux2010_x86_64.whl (4.1 MB view details)

Uploaded CPython 3.6m manylinux: glibc 2.12+ x86-64

File details

Details for the file dm_launchpad_nightly-0.3.0.dev20210622-cp39-cp39-manylinux2010_x86_64.whl.

File metadata

File hashes

Hashes for dm_launchpad_nightly-0.3.0.dev20210622-cp39-cp39-manylinux2010_x86_64.whl
Algorithm Hash digest
SHA256 c8fd7ea2c7a70e0a796bedd9fbaa602f32b1ba70d0979c2088798e6d9fc861e5
MD5 99e53aed14e3104a82614bdc092e2ed2
BLAKE2b-256 e4bea6bb88b455b47d6b4974776512e9b5f937427b8597fd07812166c90d5a42

See more details on using hashes here.

File details

Details for the file dm_launchpad_nightly-0.3.0.dev20210622-cp38-cp38-manylinux2010_x86_64.whl.

File metadata

File hashes

Hashes for dm_launchpad_nightly-0.3.0.dev20210622-cp38-cp38-manylinux2010_x86_64.whl
Algorithm Hash digest
SHA256 b5760cc69c8fdf64dd82d40fd89b1d0a87669152b868d3e500a5d8335f19ecf8
MD5 321027d74ec9e492aeeeda7b07b39ad4
BLAKE2b-256 b6bf862c9716c043d6c0dbcc513f644d5876ac15646ca97d17ae298c36c3bf43

See more details on using hashes here.

File details

Details for the file dm_launchpad_nightly-0.3.0.dev20210622-cp37-cp37m-manylinux2010_x86_64.whl.

File metadata

File hashes

Hashes for dm_launchpad_nightly-0.3.0.dev20210622-cp37-cp37m-manylinux2010_x86_64.whl
Algorithm Hash digest
SHA256 161e06b22042b3ce0bdaba01f4ea239d87d3d95228ed64ca455ce905814d8853
MD5 3cf809afe8a85894a5349fe62ef8e44c
BLAKE2b-256 e4de2e90fdd9864f2032d786bd8a0d8eb8eda9018ee88115d2a23e72759af9b6

See more details on using hashes here.

File details

Details for the file dm_launchpad_nightly-0.3.0.dev20210622-cp36-cp36m-manylinux2010_x86_64.whl.

File metadata

File hashes

Hashes for dm_launchpad_nightly-0.3.0.dev20210622-cp36-cp36m-manylinux2010_x86_64.whl
Algorithm Hash digest
SHA256 363a3f9c7154f741b87e58b03dbaeccb681c8c0bf5d279f1d78d196b3c9e7159
MD5 782faebc6ebb64fe0ede4237dc27b100
BLAKE2b-256 ec8b27b4d055ce3d46b2613febe6f850bf890ed423e7ea4be6a226817d45159d

See more details on using hashes here.

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page