Skip to content

Latest commit

 

History

History
177 lines (146 loc) · 7.18 KB

DEVELOPMENT.md

File metadata and controls

177 lines (146 loc) · 7.18 KB
Raw

DEVELOPMENT

To develop s3torchconnector, you need to have Python, pip and python-venv installed.

s3torchconnector uses s3torchconnectorclient as the underlying S3 Connector. s3torchconnectorclient is a Python wrapper around MountpointS3Client that uses S3 CRT to optimize performance of S3 read/write . Since MountpointS3Client is implemented in Rust, for development and building from source, you will need to install clang, cmake and rust compiler (as detailed below).

Note: CLI commands for Ubuntu/Debian

Install Python 3.x and pip

sudo apt update
sudo apt install python3
sudo apt install python3-pip

Clone project

  git clone [email protected]:awslabs/s3-connector-for-pytorch.git

Create a Python virtual environment

  cd /path/to/your/project
  python3 -m venv virtual-env
  source virtual-env/bin/activate

Install clang (needed to build the client)

  sudo apt install clang

Install cmake (needed to build the client)

  sudo apt install cmake

Install Rust compiler (needed to build the client)

  curl https://sh.rustup.rs -sSf | sh
  source "$HOME/.cargo/env"

Install project modules in editable mode

  pip install -e s3torchconnectorclient
  pip install -e s3torchconnector

When you make changes to the Rust code, you need to run pip install -e s3torchconnectorclient before changes will be viewable from Python.

Licensing

When developing, ensure to create license headers at the top of each file. This can be automated with Pycharm/Clion with the following configuration:

Go to the settings, and find the 'Copyright profiles' section. Create a new one with the following text:

Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.

// SPDX-License-Identifier: BSD

Then under the 'Copyright' section, create a new scope covering 'all', and assign your new copyright profile.

Making a commit

Our CI uses clippy to lint Rust code changes. Use cargo clippy --all-targets --all-features to lint Rust before pushing new Rust commits.

For Python code changes, run

black --verbose .
flake8 s3torchconnector/ --count --select=E9,F63,F7,F82 --show-source --statistics
flake8 s3torchconnectorclient/python --count --select=E9,F63,F7,F82 --show-source --statistics
mypy s3torchconnector/src
mypy s3torchconnectorclient/python/src

to lint.

To run mypy without lightning installed, run

mypy s3torchconnector/src --exclude s3torchconnector/src/s3torchconnector/lightning
mypy s3torchconnectorclient/python/src

Debugging

Either a Python or GDB style debugger will be useful here.

To use a GDB debugger from Rust, just run the Rust test in question with the debugger enabled.

To use a GDB debugger from Python, you need to create a 'Custom Build Application'. Fill in the path of the Python executable in your virtual environment (venv/bin/python) and fill in the script name as the program argument. Then put a breakpoint in the Rust/C code and try running it.

Enabling Debug Logging

The Python logger handles logging messages from the Python-side of our implementation. For debug purposes, you can also enable the logs for our Rust components, which are off by default. These are handled by tracing_subscriber and can be configured through the following environment variables:

  • S3_TORCH_CONNECTOR_DEBUG_LOGS - Configured similarly to the RUST_LOG variable for filtering logs from our Rust components. This includes finer granularity logs from AWS Common Runtime (CRT). Please note that the AWS CRT logs are very noisy. We recommend to filter them out by appending "awscrt=off" to your S3_TORCH_CONNECTOR_DEBUG_LOGS setup.
  • S3_TORCH_CONNECTOR_LOGS_DIR_PATH - The path to a local directory where you have write permissions. When configured, the logs from the Rust components will be appended to a file at this location. This will result in a log file located at ${S3_TORCH_CONNECTOR_LOGS_DIR_PATH}/s3torchconnectorclient.log.yyyy-MM-dd-HH, rolled on an hourly basis. The log messages of the latest run are appended to the end of the most recent log file.

Examples

  • Configure INFO level logs to be written to STDOUT:
  export S3_TORCH_CONNECTOR_DEBUG_LOGS=info
  • Enable TRACE level logs (most verbose) to be written at /tmp/s3torchconnector-logs:
  export S3_TORCH_CONNECTOR_DEBUG_LOGS=trace
  export S3_TORCH_CONNECTOR_LOGS_DIR_PATH="/tmp/s3torchconnector-logs"

After running your script, you will find the logs under /tmp/s3torchconnector-logs. The file will include AWS CRT logs.

  • Enable TRACE level logs with AWS CRT logs filtered out, written at /tmp/s3torchconnector-logs:
  export S3_TORCH_CONNECTOR_DEBUG_LOGS=trace,awscrt=off
  export S3_TORCH_CONNECTOR_LOGS_DIR_PATH="/tmp/s3torchconnector-logs"
  • Set up different levels for inner components:
  export S3_TORCH_CONNECTOR_DEBUG_LOGS=trace,mountpoint_s3_client=debug,awscrt=error

This will set the log level to TRACE by default, DEBUG for mountpoint-s3-client and ERROR for AWS CRT.

For more examples please check the env_logger documentation.

Fine Tuning

Using S3ClientConfig you can set up the following parameters for the underlying S3 client:

  • throughput_target_gbps(float): Throughput target in Gigabits per second (Gbps) that we are trying to reach. 10.0 Gbps by default (may change in future).

  • part_size(int): Size (bytes) of file parts that will be uploaded/downloaded. Note: for saving checkpoints, the inner client will adjust the part size to meet the service limits. (max number of parts per upload is 10,000, minimum upload part size is 5 MiB). Part size must have values between 5MiB and 5GiB. Is set by default to 8MiB (may change in future).

  • unsigned(bool): Set to true to disable signing S3 requests.

For example this can be passed in like:

from s3torchconnector import S3MapDataset, S3ClientConfig

# Setup for DATASET_URI and REGION.
...
# Setting part_size to 5 MiB and throughput_target_gbps to 15 Gbps.
config = S3ClientConfig(part_size=5 * 1024 * 1024, throughput_target_gbps=15)
# Passing this on to an S3MapDataset.
s3_map_dataset = S3MapDataset.from_prefix(DATASET_URI, region=REGION, s3client_config=config)
# Updating the configuration for checkpoints.
# Please note that you can also pass in a different configuration to checkpoints.
s3_checkpoint = S3Checkpoint(region=REGION, s3client_config=config)
# Works similarly for Lightning checkpoints.
s3_lightning_checkpoint = S3LightningCheckpoint(region=REGION, s3client_config=config)

# Disable signing to make requests without AWS credentials
config = S3ClientConfig(unsigned=True)
s3_map_dataset = S3MapDataset.from_prefix(DATASET_URI, region=REGION, s3client_config=config)

When modifying the default values for these flags, we strongly recommend to run benchmarking to ensure you are not introducing a performance regression.