GSoC Final Submission- What Zarrtraj is and why you should use it

What is Zarrtraj?

Zarrtraj is an MDAKit for storing and analyzing trajectories in MDAnalysis from the cloud, representing a major milestone towards the MDAnalysis 3.0 goal of cloud streaming as well as a proof-of-concept for a new paradigm in the field of molecular dynamics as a whole. It can stream H5MD trajectories from cloud storage providers including AWS S3, Google Cloud buckets, and Azure Blob storage and data lakes. Using Zarrtraj, anyone can reproduce your analyses or train their ML model on trajectory data without ever having to download massive files to their disk.

This is possible thanks to Zarr, fsspec, and kerchunk, packages that have created a foundation for storing and interacting with large datasets in a uniform way across different storage backends. Interestingly, these projects were intially developed by geoscientists in the Pangeo project to make use of cloud computing credits from an NSF partnership with cloud providers, but have since undergone wider adoption by the broader Python community. This project also represents one of the first forays of the molecular dynamics field into the excellent Pangeo ecosystem of tools.

Zarr is especially well-suited to the task of reading cloud-stored files due to its integration with dask for parallelized reading. This parallelization offsets the increased IO time in cloud-streaming, speeding up common analysis algorithms up to ~4x compared to sequential analysis. See the zarrtraj benchmarks for more.

In this project, we also decided to experiment with storing trajectories directly in Zarr-backed files using the same specification that H5MD uses, so Zarrtraj can read both .h5md and H5MD-formatted .zarrmd files. See this explanation of the modified format to learn more.

While this GSoC project started with the goal of building a new, Zarr-backed trajectory format, we pivoted to making the existing H5MD format streamable after getting feedback from the community that supporting widely adopted formats in the MD ecosystem makes code more sustainable and simplifies tool adoption.

The next section is a walkthrough of Zarrtraj’s features and usage also available here.

How can I use it?

Zarrtraj is currently available via PyPI and Conda Forge.

Pip installation

pip install zarrtraj

Conda installation

conda install -c conda-forge zarrtraj

For more information on installation, see the installation guide

This walkthrough will guide you through the process of reading and writing H5MD-formatted trajectories from cloud storage using AWS S3 as an example. To learn more about reading and writing trajectories from different cloud storage providers, including Google Cloud and Azure, see the API documentation.

Reading H5MD trajectories from cloud storage

Uploading your H5MD file

First, upload your H5MD trajectories to an AWS S3 bucket. This requires that an S3 Bucket is setup and configured for write access using the credentials stored in “sample_profile”. If you’ve never configured an S3 Bucket before, see this guide. You can setup a profile to easily manage AWS credentials using this VSCode extension. Here is a sample profile (stored in ~/.aws/credentials) where the key is an access key associated with a user that has read and write permissions for the bucket

[sample_profile]
aws_access_key_id = <key>

MDAnalysis can write a trajectory from any of its supported formats into H5MD. We recommend using the chunks kwarg with the MDAnalysis H5MDWriter with a value that yields ~8-16MB chunks of data for best S3 performance. Once written locally, you can upload the trajectory to S3 programmatically:

import os
from botocore.exceptions import ClientError
import boto3
import logging

os.environ["AWS_PROFILE"] = "sample_profile"
# This is the AWS region where the bucket is located
os.environ["AWS_REGION"] = "us-west-1"

def upload_h5md_file(bucket_name, file_name):
    s3_client = boto3.client("s3")
    obj_name = os.path.basename(file_name)

    response = s3_client.upload_file(
        file_name, bucket_name, obj_name
    )


if __name__ == "__main__":
    # Using test H5MD file from the Zarrtraj repo
    upload_h5md_file("sample-bucket-name", "zarrtraj/data/COORDINATES_SYNTHETIC_H5MD.h5md")

You can also upload the H5MD file directly using the AWS web interface by navigating to S3, the bucket name, and pressing “upload”.

Reading your H5MD file

After the file is uploaded, you can use the same credentials to stream the file into MDAnalysis:

import zarrtraj
import MDAnalysis as mda
# This sample topology requires installing MDAnalysisTests
from MDAnalysisTests.datafiles import COORDINATES_TOPOLOGY
import os

os.environ["AWS_PROFILE"] = "sample_profile"
os.environ["AWS_REGION"] = "us-west-1"

u = mda.Universe(COORDINATES_TOPOLOGY, "s3://sample-bucket-name/COORDINATES_SYNTHETIC_H5MD.h5md")
for ts in u.trajectory:
    pass

You can follow this same process for reading .zarrmd files with the added advantage that Zarrtarj can write .zarrmd files directly into an S3 bucket.

Writing trajectories from MDAnalysis into a zarrmd file in an S3 Bucket

Using the same credentials with read/write access, you can write a trajectory into your bucket.

You can change the stored precision of floating point values in the file with the optional precision kwarg and pass in any numcodecs.Codec compressor with the optional compressor kwarg. See numcodecs for more on the available compressors.

Chunking is automatically determined for all datasets to be optimized for cloud storage and is not configurable by the user. Initial benchmarks show this chunking strategy is effective for disk storage as well.

import zarrtraj
import MDAnalysis as mda
from MDAnalysisTests.datafiles import PSF, DCD
import numcodecs
import os

os.environ["AWS_PROFILE"] = "sample_profile"
os.environ["AWS_REGION"] = "us-west-1"

u = mda.Universe(PSF, DCD)
with mda.Writer("s3://sample-bucket-name/test.zarrmd", 
                n_atoms=u.trajectory.n_atoms, 
                precision=3,
                compressor=numcodecs.Blosc(cname="zstd", clevel=9)) as W:
                for ts in u.trajectory:
                    W.write(u.atoms)

If you have additional questions, please don’t hesitate to open a discussion on the zarrtarj github. The MDAnalysis discord is also a great resource for asking questions and getting involved in MDAnalysis.

What’s next for Zarrtraj?

Zarrtraj is currently in a fully operational state and is ready for use! However, I’m excited about creating some new features in the future that will make Zarrtraj more flexible and faster.

Lazytimeseries

In MDAnalysis, many trajectory readers expose a timeseries method for getting access to coordinate data for a subselection of atoms across a trajectory. This provides a viable way to sidestep the Timestep (eagerly-loaded frame-based) paradigm that MDAnalysis uses for handling trajectory data. Zarrtraj could implement a “lazytimeseries” that returns a lazy dask array of a selection of atoms’ positions across the trajectory. Early benchmarks show that analysis based on such a lazy array can outperform Timestep-based analysis.

Asynchronous Reading

The performance impact of network IO could be reduced by creating a multi-threaded ZARRH5MDReader that isn’t blocked by analysis code executing. The reader could eagerly load the cache with the next frames the analysis code will need to reduce the impact of network IO on exeution time.

Acknowledgements

Thank you to Dr. Hugo MacDermott-Opeskin and Dr. Yuxuan Zhuang for their mentorship and feedback throughout this project and to Dr. Jenna Swarthout Goddard for supporting the GSoC program at MDAnalysis.

I also want to thank Dr. Oliver Beckstein and Edis Jakupovic for lending their expertise in H5MD and all things MDAnalysis.

Finally, another thanks to Martin Durant, author of Kerchunk, who was incredibly helpful in refining and merging a new feature in his codebase necessary for this project to work.

Citations

Alistair Miles, jakirkham, M Bussonnier, Josh Moore, Dimitri Papadopoulos Orfanos, Davis Bennett, David Stansby, Joe Hamman, James Bourbeau, Andrew Fulton, Gregory Lee, Ryan Abernathey, Norman Rzepka, Zain Patel, Mads R. B. Kristensen, Sanket Verma, Saransh Chopra, Matthew Rocklin, AWA BRANDON AWA, … shikharsg. (2024). zarr-developers/zarr-python: v3.0.0-alpha (v3.0.0-alpha). Zenodo. https://doi.org/10.5281/zenodo.11592827

de Buyl, P., Colberg, P. H., & Höfling, F. (2014). H5MD: A structured, efficient, and portable file format for molecular data. In Computer Physics Communications (Vol. 185, Issue 6, pp. 1546–1553). Elsevier BV. https://doi.org/10.1016/j.cpc.2014.01.018

Gowers, R., Linke, M., Barnoud, J., Reddy, T., Melo, M., Seyler, S., Domański, J., Dotson, D., Buchoux, S., Kenney, I., & Beckstein, O. (2016). MDAnalysis: A Python Package for the Rapid Analysis of Molecular Dynamics Simulations. In Proceedings of the Python in Science Conference. Python in Science Conference. SciPy. https://doi.org/10.25080/majora-629e541a-00e

Jakupovic, E., & Beckstein, O. (2021). MPI-parallel Molecular Dynamics Trajectory Analysis with the H5MD Format in the MDAnalysis Python Package. In Proceedings of the Python in Science Conference. Python in Science Conference. SciPy. https://doi.org/10.25080/majora-1b6fd038-005

Michaud‐Agrawal, N., Denning, E. J., Woolf, T. B., & Beckstein, O. (2011). MDAnalysis: A toolkit for the analysis of molecular dynamics simulations. In Journal of Computational Chemistry (Vol. 32, Issue 10, pp. 2319–2327). Wiley. https://doi.org/10.1002/jcc.21787

Extras: Upstream code merged during this GSoC project

MDAnalysis

One feature I needed for testing Zarrtraj was full use of writer kwargs when aligning a trajectory and writing it immediately rather than storing it in memory. However, this feature hadn’t yet been implemented, and luckily, it was a small change, so I worked with the core developers to merge this PR with the new feature and tests:

• https://github.com/MDAnalysis/mdanalysis/pull/4565

MDAKits

Since Zarrtraj is based on the MDAnalysis MDAKit Cookiecutter (which is a fantastic tool for getting started in making and distributing Python packages), I was able to find and fix a few small bugs along the way in my GSoC journey including:

• https://github.com/MDAnalysis/cookiecutter-mdakit/pull/115
• https://github.com/MDAnalysis/cookiecutter-mdakit/pull/133
• https://github.com/MDAnalysis/MDAKits/pull/140

Kerchunk

Kerchunk is central to Zarrtraj’s ability to read hdf5 files using Zarr. However, H5MD files (stored in hdf5) have linked datasets as per the H5MD standard, but Kerchunk did not translate these from hdf5 to Zarr previously. I was able to work alongside a Kerchunk core developer to add the ability to translate hdf5 datasets into Zarr along with comprehensive tests of this new feature.

• https://github.com/fsspec/kerchunk/pull/463

Extras: Lessons Learned

Here are a bunch of random things I learned while doing this project!

• Any time you need to run code that will take several hours to execute due to file size or some other factor, create a minimal, quickly-executing example of the code to work out bugs before running the full thing. You will save yourself so, so much frustration
• It is worth investing time into getting a debugging environment properly configured. If you’re hunting down a specific bug, it is worth the 20 minutes it will take to create a barebones example of the bug instead of trying to hunt it down “in-situ”. A lot of the time, just creating the example will make you realize what was wrong. GH Actions runners sometimes behave differently than your development machine. This action for SSHing into a runner is FANTASTIC!
• Maintain a “tmp” directory in your locally cloned repos, gitignore it, and use it for testing random ideas you have or working through bugs. Take the time to give each file in it a descriptive name! Having these random scripts and ideas all in one place will pay off massively later on.
• Take risks with ideas if you suspect they might result in cleaner and faster code, even if you’re not 100% sure! Experimenting is worth it.
• Don’t be afraid to read source code! Sometimes the fastest way to solving a problem is seeing how someone else solved it, and sometimes the fastest way to learning why someone else’s code isn’t doing what you expected is to read the code rather than the docs.