Skip to main content

A *simpler* and *cheaper* way to distribute python (training) code on machines of your choice in the (AWS) cloud.

Project description

Simple Sagemaker

A simpler and cheaper way to distribute python (training) code on machines of your choice in the (AWS) cloud.

Note: this (initial) work is still in progress. Only PyTorch and Tensorflow frameworks are currently supported (but these images can be used to distribute any python code, if no special additional needs are set).

Requirements

  1. Python 3.6+
  2. AWS account credentials configured for boto3, as explained on the Boto3 docs

Getting started

To install Simple Sagemaker

pip install simple-sagemaker

Then, to run the following worker1.py on 2 ml.p3.2xlarge spot instances

import torch

for i in range(torch.cuda.device_count()):
    print(f"-***- Device {i}: {torch.cuda.get_device_properties(i)}")

Just run the below command:

ssm -p simple-sagemaker-example-cli -t task1 -e worker1.py -o ./output/example1 --it ml.p3.2xlarge --ic 2

The output logs are saved to ./output/example1/logs. The relevant part from the log files (./output/example1/logs/logs0 and ./output/example1/logs/logs1) is:

...
-***- Device 0: _CudaDeviceProperties(name='Tesla V100-SXM2-16GB', major=7, minor=0, total_memory=16160MB, multi_processor_count=80)
...

More examples (below)

CLI based examples:

API only example:

Background

Simple Sagemaker is a thin warpper around SageMaker's training jobs, that makes distribution of python code on any supported instance type very simple.

The solutions is composed of two parts, one on each side: a runner on the client machine, and a worker which is the distributed code on AWS.

  • The runner is the main part of this package, can mostly be controlled by using the ssm command line interface (CLI), or be fully customized by using the python API.
  • The worker is basically the code you're trying to distribute, with possible minimal code changes that should use a small task_tollkit library which is injected to it, for extracting the environment configuration, i.e. input/output/state paths and running parameters.

The runner is used to configure the tasks and projects:

  • A task is a logical step that runs on a defined input and provide output. It's defined by providing a local code path, entrypoint, and a list of additional local dependencies
  • A SageMaker job is a task instance, i.e. a single job is created each time a task is executed
    • State is maintained between consecutive execution of the same task
    • Task can be markd as completed, to avoid re-running it next time (unlesss eforced otherwise)
  • A prjoect is a series of related tasks, with possible depencencies
    • The output of one task can be consumed by a consequetive task

Main features

  1. Simpler - Except for holding an AWS account credentials, no other pre-configuration nor knowledge is assumed (well, almost :). Behind the scenes you get:
    • Jobs IAM role creation, including policies for accesing needed S3 buckets
    • Building and uploading a customized docker image to AWS (ECS service)
    • Synchronizing local source code / input data to a S3 bucket
    • Downloading the results from S3
    • ...
  2. Cheaper - "pay only for what you use", and save up to 90% of the cost with spot instances, which got used by default!
  3. Abstraction of how data is maintianed on AWS (S3 service)
    • No need to mess with S3 paths, the data is automatically
    • State is automaticall maintained between consequetive execution of jobs that belongs to the same task
  4. A simple way to define how data flows between tasks of the same project, e.g. how the first task's outputs is used as an input for a second task
  5. (Almost) no code changes are to youe existing code - the API is mostly wrapped by a command line interface (named ssm) to control the execution (a.k.a implement the runner, see below)
    • In most cases it's only about 2 line for getting the environment configuration (e.g. input/output/state paths and running parameters) and passing it on to the original code
  6. Easy customization of the docker image (based on a pre-built one)
  7. The rest of the SageMaker advantages, which (mostly) behaves "normally" as defined by AWS, e.g.
    • (Amazon SageMaker Developer Guide)[https://docs.aws.amazon.com/sagemaker/latest/dg/whatis.html]
    • (Amazon SageMaker Python SDK @ Read the Docs)[https://sagemaker.readthedocs.io/en/stable/index.html]

High level flow diagram

High level flow diagram

Data maintainance on S3

All data, including input, code, state and output, is maintained on S3. The bucket to use can be defined, or the default one is used. The files and directories structure is as follows:

[Bucket name]/[Project name]/[Task name]
|-- state
|-- input
|-- [Job name]
|   |-- output
|   |   |-- model.tar.gz
|   |   `-- output.tar.gz
|   `-- source/sourcedir.tar.gz
|-- [Job name 2]
|        ...
  • state - the task state, shared between all jobs, i.e. task executions
  • input - the task input, shared as well
  • [Job name] - a per job specific folder
    • model.tar.gz - model output data, merged from all instances
    • output.tar.gz - the main instance output data (other outputs are ignored)
    • sourcedir.tar.gz - source code and dependencies
  • [Job name 2] - another execution of the same task

A fully featured advanced example

And now to a real advanced and fully featured version, yet simple to implement. In order to examplify most of the possible features, the following files are used in CLI Example 6_1:

.
|-- Dockerfile
|-- code
|   |-- internal_dependency
|   |   `-- lib2.py
|   |-- requirements.txt
|   `-- worker6.py
|-- data
|   |-- sample_data1.txt
|   `-- sample_data2.txt
`-- external_dependency
    `-- lib1.py
  • Dockerfile - the dockerfile specifying how to extend the pre-built image
    # __BASE_IMAGE__ is automatically replaced with the correct base image
    FROM __BASE_IMAGE__ 
    RUN pip3 install pandas==0.25.3 scikit-learn==0.21.3
    
  • code - the source code folder
    • internal_dependency - a dependency that is part of the source code folder
    • requirements.txt - pip requirements file lists needed packages to be installed before running the worker
      transformers==3.0.2
      
    • worker6.py - the worker code
  • data - input data files
  • external_dependency - additional code dependency

The code is then launced a few time by run.sh, to demonstrate different features:

# Example 6_1 - a complete example part 1. 
#   - Uses local data folder as input, that is distributed among instances (--i, ShardedByS3Key)
#   - Uses a public s3 bucket as an additional input (--iis)
#   - Builds a custom docker image (--df, --repo_name, --aws_repo)
#   - Hyperparameter task_type
#   - 2 instance (--ic)
#   - Use an on-demand instance (--no_spot)
ssm -p simple-sagemaker-example-cli -t task6-1 -s $BASEDIR/example6/code -e worker6.py \
    -i $BASEDIR/example6/data ShardedByS3Key --iis persons s3://awsglue-datasets/examples/us-legislators/all/persons.json \
    --df $BASEDIR/example6 --repo_name "task6_repo" --aws_repo "task6_repo" \
    --ic 2 --task_type 1 -o $1/example6_1

# Example 6_2 - a complete example part 2.
#   - Uses outputs from part 1 (--iit)
#   - Uses additional local code dependencies (-d)
#   - Uses the tensorflow framework as pre-built image (-f)
#   - Tags the jobs (--tag)
#   - Defines sagemaker metrics (-m, --md)
ssm -p simple-sagemaker-example-cli -t task6-2 -s $BASEDIR/example6/code -e worker6.py \
    -d $BASEDIR/example6/external_dependency --iit task_6_1_model task6-1 model --iit task_6_1_state task6-1 state ShardedByS3Key \
    -f tensorflow -m --md "Score" "Score=(.*?);" --tag "MyTag" "MyValue" \
    --ic 2 --task_type 2 -o $1/example6_2 &

# Running task6_1 again
#   - A completed task isn't exsecuted again, but the current output is used instead. 
#       --ks (keep state, the default) is used to keep the current state
ssm -p simple-sagemaker-example-cli -t task6-1 -s $BASEDIR/example6/code -e worker6.py \
    -i $BASEDIR/example6/data ShardedByS3Key --iis persons s3://awsglue-datasets/examples/us-legislators/all/persons.json \
    --df $BASEDIR/example6 --repo_name "task6_repo" --aws_repo "task6_repo" \
    --ic 2 --task_type 1 -o $1/example6_1 > $1/example6_1_2_stdout --ks &


wait # wait for all processes

The metrics graphs can be viewed on the AWS console:

High level flow diagram

More information can be found here.

Feel free to dive more into the files of this example. Specifically, note how the same worker code is used for the two parts, and the task_type hyperparameter is used to distinguish between the two.

More examples

CLI based examples:

API only example:

Passing command line arguments

Any extra argument passed to the command line in assumed to be an hypermarameter. To get access to all environment arguments, call task_lib.parseArgs(). For example, see the following worker code worker2.py:

from task_toolkit import task_lib

args = task_lib.parseArgs()
print(args.hps["msg"])

Running command:

ssm -p simple-sagemaker-example-cli -t task2 -e worker2.py --msg "Hello, world!" -o ./output/example2

Output from the log file

Invoking script with the following command:

/opt/conda/bin/python worker2.py --msg Hello, world!

Hello, world!

Task state and output

State

State is maintained between executions of the same task, i.e. between execution jobs that belongs to the same task. The local path is available in args.state. When running multiple instances, the data is merged into a single directory, so in order to avoid collisions, task_lib.initMultiWorkersState(args) initializes a per instance sub directory. On top of that, task_lib provides an additional important API to mark the task as completed: task_lib.markCompleted(args). If all instances of the job mark it as completed, the task is assumed to be completed by that job, which allows:

  1. To skip it next time (unlesss eforced otherwise)
  2. To use its output as input for other tasks (see below: "Chaining tasks")

Output

There're 3 main output mechanisms:

  1. Logs - any output writen to standard output
  2. Output data - args.output_data_dir is compressed into a tar.gz file, only the main instance data is kept
  3. Model - args.model_dir is compressed into a tar.gz file, data from all instance is merged, so be carful with collisions.

A complete example can be seen in worker3.py:

import os

from task_toolkit import task_lib

args = task_lib.parseArgs()

open(os.path.join(args.output_data_dir, "output_data_dir"), "wt").write(
    "output_data_dir file"
)
open(os.path.join(args.model_dir, "model_dir"), "wt").write("model_dir file")
open(os.path.join(args.state, "state_dir"), "wt").write("state_dir file")

# Mark the tasks as completed, to allow other tasks using its output, and to avoid re-running it (unless enforced)
task_lib.markCompleted(args)

Running command:

ssm -p simple-sagemaker-example-cli -t task3 -e worker3.py -o ./output/example3

Output from the log file

Invoking script with the following command:

/opt/conda/bin/python worker2.py --msg Hello, world!

Hello, world!

Providing input data

Job can be configured to get a few data sources:

  • A single local path can be used with the -i argument. This path is synchronized to the task directory on the S3 bucket before running the task. On the worker side the data is accesible in args.input_data
  • Additional S3 paths (many) can be set as well. Each input source is provided with --iis [name] [S3 URI], and is accesible by the worker with args.input_[name] when [name] is the same one as was provided on the command line.
  • Setting an output of a another task on the same project, see below "Chaining tasks"

Assuming a local data folder containtin a single sample_data.txt file, a complete example can be seen in worker4.py:

import logging
import subprocess
import sys

from task_toolkit import task_lib

logger = logging.getLogger(__name__)


def listDir(path):
    logger.info(f"*** START listing files in {path}")
    logger.info(
        subprocess.run(
            ["ls", "-la", "-R", path], stdout=subprocess.PIPE, universal_newlines=True
        ).stdout
    )
    logger.info(f"*** END file listing {path}")


if __name__ == "__main__":
    logging.basicConfig(stream=sys.stdout)
    task_lib.setDebugLevel()
    args = task_lib.parseArgs()
    listDir(args.input_data)
    listDir(args.input_bucket)

Running command:

ssm -p simple-sagemaker-example-cli -t task4 -e worker4.py -i ./data --iis bucket s3://awsglue-datasets/examples/us-legislators/all/persons.json -o ./output/example4

Output from the log file

...
INFO:__main__:*** START listing files in /opt/ml/input/data/data
INFO:__main__:/opt/ml/input/data/data:
total 12
drwxr-xr-x 2 root root 4096 Sep 14 21:51 .
drwxr-xr-x 4 root root 4096 Sep 14 21:51 ..
-rw-r--r-- 1 root root   19 Sep 14 21:51 sample_data.txt

INFO:__main__:*** END file listing /opt/ml/input/data/data
INFO:__main__:*** START listing files in /opt/ml/input/data/bucket
INFO:__main__:/opt/ml/input/data/bucket:
total 7796
drwxr-xr-x 2 root root    4096 Sep 14 21:51 .
drwxr-xr-x 4 root root    4096 Sep 14 21:51 ..
-rw-r--r-- 1 root root 7973806 Sep 14 21:51 persons.json

INFO:__main__:*** END file listing /opt/ml/input/data/bucket
...

Chaining tasks

The output of a completed task on the same project can be used as an input to another task, by using the --iit [name] [task name] [output type] command line parameter, where:

  • [name] - is the name of the input source, caccesible by the worker with args.input_[name]
  • [task name] - the name of the task whose output is used as input
  • [output type] - the task output type, one of "model", "output", "state"

Using the output of task3 and the same worker4.py code, we can now run:

ssm -p simple-sagemaker-example-cli -t task5 -e worker4.py --iit bucket task3 model -o ./output/example5

And get the following output from in the log file:

INFO:__main__:*** START listing files in 
INFO:__main__:
INFO:__main__:*** END file listing 
INFO:__main__:*** START listing files in /opt/ml/input/data/bucket
INFO:__main__:/opt/ml/input/data/bucket:
total 12
drwxr-xr-x 2 root root 4096 Sep 14 21:55 .
drwxr-xr-x 3 root root 4096 Sep 14 21:55 ..
-rw-r--r-- 1 root root  128 Sep 14 21:55 model.tar.gz

INFO:__main__:*** END file listing /opt/ml/input/data/bucket

Configuring the docker image

The image used to run a task can either be selected from a pre-built ones or extended with additional Dockerfile commands. The framework, framework_version and python_version CLI parameters are used to define the pre-built image, then if a path to a directory containing the Dockerfile is given by docker_file_path, it used along with aws_repo, repo_name and image_tag to build and push an image to ECS, and then set it as the used image. The base image should be set to __BASE_IMAGE__ within the Dockerfile, and is automatically replaced with the correct base image (according to the provided parameters above) before building it. The API parameter for the Dockerfile path is named docker_file_path_or_content and allows to provide the content of the Dockerfile, e.g.

dockerFileContent = """
# __BASE_IMAGE__ is automatically replaced with the correct base image
FROM __BASE_IMAGE__
RUN pip3 install pandas==1.1 scikit-learn==0.21.3
"""

Sample usages:

  1. CLI Example 6_1- a CLI example launched by run.sh
  2. single file example - API with Dockerfile content
  3. single task example - API with Dockerfile path

Defining code dependencies

Additional local code dependencies can be specified with the dependencies CLI/API parameters. These dependencies are packed along with the source code, and are extracted on the root code folder in run time.

Sample usages:

  1. CLI Example 6_2- a CLI example launched by run.sh
  2. single task example - API

Single file example

A single file example can be found in the examples directory. First, define the runner:

dockerFileContent = """
# __BASE_IMAGE__ is automatically replaced with the correct base image
FROM __BASE_IMAGE__
RUN pip3 install pandas==1.1 scikit-learn==0.21.3
"""
file_path = Path(__file__).parent


def runner(project_name="simple-sagemaker-sf", prefix="", postfix="", output_path=None):
    from simple_sagemaker.sm_project import SageMakerProject

    sm_project = SageMakerProject(project_name=project_name)
    # define the code parameters
    sm_project.setDefaultCodeParams(source_dir=None, entry_point=__file__, dependencies=[])
    # define the instance parameters
    sm_project.setDefaultInstanceParams(instance_count=2)
    # docker image
    sm_project.setDefaultImageParams(
        aws_repo_name="task_repo",
        repo_name="task_repo",
        img_tag="latest",
        docker_file_path_or_content=dockerFileContent,
    )
    image_uri = sm_project.buildOrGetImage(
        instance_type=sm_project.defaultInstanceParams.instance_type
    )

    # *** Task 1 - process input data
    task1_name = "task1"
    # set the input data
    input_data_path = file_path.parent / "data"
    # run the task
    sm_project.runTask(
        task1_name,
        image_uri,
        distribution="ShardedByS3Key",  # distribute the input files among the workers
        hyperparameters={"worker": 1, "arg": "hello world!", "task": 1},
        input_data_path=str(input_data_path) if input_data_path.is_dir() else None,
        clean_state=True,  # clean the current state, also forces re-running
    )
    # download the results
    if not output_path:
        output_path = file_path.parent / "output"
    shutil.rmtree(output_path, ignore_errors=True)
    sm_project.downloadResults(task1_name, Path(output_path) / "output1")

An additional task that depends on the previous one can now be scheduled as well:

    # *** Task 2 - process the results of Task 1
    task2_name = "task2"
    # set the input
    additional_inputs = {
        "task2_data": sm_project.getInputConfig(task1_name, "model"),
        "task2_data_dist": sm_project.getInputConfig(
            task1_name, "model", distribution="ShardedByS3Key"
        ),
    }
    # run the task
    sm_project.runTask(
        task2_name,
        image_uri,
        hyperparameters={"worker": 1, "arg": "hello world!", "task": 2},
        clean_state=True,  # clean the current state, also forces re-running
        additional_inputs=additional_inputs,
    )
    # download the results
    sm_project.downloadResults(task1_name, Path(output_path) / "output2")

    return sm_project

Then, the worker code (note: the same function is used for the two different tasks, depending on the task hyperparameter):

def worker():
    from task_toolkit import task_lib

    task_lib.setDebugLevel()

    logger.info("Starting worker...")
    # parse the arguments
    args = task_lib.parseArgs()

    state_dir = task_lib.initMultiWorkersState(args)

    logger.info(f"Hyperparams: {args.hps}")
    logger.info(f"Input data files: {list(Path(args.input_data).rglob('*'))}")
    logger.info(f"State files: { list(Path(args.state).rglob('*'))}")

    if int(args.hps["task"]) == 1:
        # update the state per running instance
        open(f"{state_dir}/state_{args.current_host}", "wt").write("state")
        # write to the model output directory
        for file in Path(args.input_data).rglob("*"):
            relp = file.relative_to(args.input_data)
            path = Path(args.model_dir) / (str(relp) + "_proc_by_" + args.current_host)
            path.write_text(file.read_text() + " processed by " + args.current_host)
        open(f"{args.model_dir}/output_{args.current_host}", "wt").write("output")
    elif int(args.hps["task"]) == 2:
        logger.info(f"Input task2_data: {list(Path(args.input_task2_data).rglob('*'))}")
        logger.info(
            f"Input task2_data_dist: {list(Path(args.input_task2_data_dist).rglob('*'))}"
        )

    # mark the task as completed
    task_lib.markCompleted(args)
    logger.info("finished!")

To pack everything in a single file, we use the command line argumen --worker (as defined in the runner function) to distinguish between runner and worker runs

import logging
import shutil
import sys
from pathlib import Path

logger = logging.getLogger(__name__)

...

def main():
    logging.basicConfig(stream=sys.stdout, level=logging.INFO)
    if "--worker" in sys.argv:
        worker()
    else:
        runner()


if __name__ == "__main__":
    main()

Running the file, with a sibling directory named data with a sample file as on the example, prduces the following outputs for Task 1:

INFO:__main__:Hyperparams: {'arg': 'hello world!', 'task': 1, 'worker': 1}
INFO:__main__:Input data files: [PosixPath('/opt/ml/input/data/data/sample_data1.txt')]
INFO:__main__:State files: [PosixPath('/state/algo-1')]
INFO:task_toolkit.task_lib:Marking instance algo-1 completion
INFO:task_toolkit.task_lib:Creating instance specific state dir
INFO:__main__:finished!
INFO:__main__:Hyperparams: {'arg': 'hello world!', 'task': 1, 'worker': 1}
INFO:__main__:Input data files: [PosixPath('/opt/ml/input/data/data/sample_data2.txt')]
INFO:__main__:State files: [PosixPath('/state/algo-2')]
INFO:task_toolkit.task_lib:Marking instance algo-2 completion
INFO:task_toolkit.task_lib:Creating instance specific state dir
INFO:__main__:finished!

And the following for Task 2:

INFO:__main__:Hyperparams: {'arg': 'hello world!', 'task': 2, 'worker': 1}
INFO:__main__:Input data files: [PosixPath('task_toolkit'), PosixPath('example.py'), PosixPath('task_toolkit/task_lib.py'), PosixPath('task_toolkit/__pycache__'), PosixPath('task_toolkit/__init__.py'), PosixPath('task_toolkit/__pycache__/__init__.cpython-38.pyc'), PosixPath('task_toolkit/__pycache__/task_lib.cpython-38.pyc')]
INFO:__main__:State files: [PosixPath('/state/algo-1')]
INFO:__main__:Input task2_data: [PosixPath('/opt/ml/input/data/task2_data/model.tar.gz')]
INFO:__main__:Input task2_data_dist: [PosixPath('/opt/ml/input/data/task2_data_dist/model.tar.gz')]
INFO:task_toolkit.task_lib:Marking instance algo-1 completion
INFO:task_toolkit.task_lib:Creating instance specific state dir
INFO:__main__:Hyperparams: {'arg': 'hello world!', 'task': 1, 'worker': 1}
INFO:__main__:Input data files: [PosixPath('/opt/ml/input/data/data/sample_data2.txt')]
INFO:__main__:State files: [PosixPath('/state/algo-2')]
INFO:task_toolkit.task_lib:Marking instance algo-2 completion
INFO:task_toolkit.task_lib:Creating instance specific state dir
INFO:__main__:finished!

As mentioned, the complete code can be found in this directory,

Development

Pushing a code change

  1. Develop ...
  2. Format & lint
tox -e cf
tox -e lint
  1. Cleanup
tox -e clean
  1. Test
tox
  1. Generate & test coverage
tox -e report
  1. [Optionally] - bump the version string on /src/simple_sagemaker/init to allow the release of a new version
  2. Push your code to a development branch
    • Every push is tested for linting + some
  3. Create a pull request to the master branch
    • Every master push is fully tested
  4. If the tests succeed, the new version is publihed to PyPi

Open issue

  1. S3_sync doesn't delete remote files if deleted locally + optimization
  2. Handling spot instance termination / signals

Project details


Download files

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

Source Distribution

simple_sagemaker-0.9.13.tar.gz (98.6 kB view hashes)

Uploaded Source

Built Distribution

simple_sagemaker-0.9.13-py3-none-any.whl (33.9 kB view hashes)

Uploaded Python 3

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