Aaron Culich's picture

Jetstream cloud support for multi-institutional data science workshops and research

Jupyter notebook showing neuroimaging visualization, over Jetstream banner

Data scientists at the Berkeley Institute for Data Science (BIDS) and the University of Washington’s eScience Institute teamed up with UCSF researchers to deliver a workshop on data-driven analysis and machine learning for neuroscience imaging data. The workshop was held in January 2017, and had 30-40 participants comprised of faculty, postdocs, graduate students, and data science fellows from UCSF, UC Berkeley, Lawrence Berkeley Lab, and the University of Washington.

As a part of this workshop, Berkeley’s Research IT worked with the organizers to host all course materials on an XSEDE cluster, allowing students access to their own python environments that were accessed via a web browser. This was made possible by a collaboration between BIDS, UW eScience, and Research IT’s Berkeley Research Computing (BRC) program, and was originally conceived at an earlier Moore-Sloan Data Science Environment Summit focused on “building community around data science for research.”

This article describes the Advanced Cyberinfrastructure (ACI) support that made this workshop possible, and highlights reusable and shareable patterns to build on for future work. You can read more about the collaboration and content of the workshop itself in a related blog post from our collaborators at BIDS, excerpted below to provide context for this article:

“All participants received their own session on the XSEDE cluster that they accessed via a unique IP address, allowing them to perform computationally demanding analysis using their own laptops. This made the learning experience very effective and reduced the amount of difficulty related to customizing the programming environment for each student. In addition, these online resources were made available for several days after the event so that students could re-run the code and analyze the data on their own time.”

The Jetstream cloud platform provided the computational resources for the workshop. Jetstream’s core capabilities include the ability to create interactive Virtual Machines (VMs), access to remote desktops through a web browser, and publishing VMs with a DOI. Jetstream is attractive to communities who have not been users of traditional HPC systems, but who would benefit from advanced computational capabilities.

Access to Jetstream is available to researchers at no cost through the NSF-funded XSEDE (Extreme Science and Engineering Discovery Environment) project which offers a portfolio of supercomputers and high-end visualization and data-analysis resources across the country to address increasingly diverse scientific and engineering challenges.

To obtain access, a qualified PI writes a resource justification and submits an allocation request. To help speed up the process of choosing and obtaining access to the resource, many campuses have local XSEDE Campus Champions who can facilitate quick access and help prepare an allocation request.

For the neuroimaging workshop, the local Campus Champion worked with BIDS and eScience data scientists to prepare an Education Allocation request. Below are some key excerpts from the 1-page allocation request, which you can read in full from the list of example allocation requests:

  • 50 Virtual Machines running simultaneously (40 students + 5 instructors + test/spare/debug VMs)
  • Each VM will need to be a: Jetstream m1.medium VM (6 vCPUs, 16GB RAM, 60GB Storage)
  • Each VM will need an external IP address so students can connect remotely with a web browser to a Jupyter Notebook running on the machine
  • We are requesting 10,000SUs in total.

The technology we used to deploy the workshop in addition to the Jetstream cloud platform includes Docker, Dockerhub, and the docker-stacks maintained by the Jupyter project.

Each of the instructors initially used their own laptops to develop Jupyter Notebook-based tutorials on computer vision and machine learning for neuroscience, using state-of-the-art deep learning methods and software such as Tensorflow and scikit-learn.

Research IT staff worked with BIDS and eScience data scientists to build a customized container from the Jupyter project’s datascience-notebook image. This provides a pre-configured Jupyter Notebook 4.3.x; Conda Python 3.x and Python 2.7.x environments; and several common libraries including: pandas, matplotlib, scipy, seaborn, scikit-learn, and scikit-image. Additional neuroscience-specific packages were included such as Dipy for diffusion magnetic resonance imaging (dMRI) analysis.

This customized container ensured that each student had an identical environment on the day of the workshop, including all required software dependencies. The container made it possible for each participant to easily run the software without installing each of the components, often a lengthy and error-prone process at the start of many workshops. The container can also be used as a snapshot-in-time or a “time capsule” so the software is preserved for future use. Months or years for now it is possible to re-run the notebooks again, even if external software packages and dependencies have changed.

The container image was pushed to https://hub.docker.com/ which provides a centralized resource for container image discovery, distribution and change management, user and team collaboration, and workflow automation.

While it is possible to download and run a customized container directly on a participant’s laptop, the instructors wanted to simplify the workshop experience. A Docker container for each participant was therefore provisioned to a virtual machine (VM) running remotely on the XSEDE Jetstream cloud platform. This allowed instructors to dive directly into the material, without a download-and-install step. The participant only needed to connect to their assigned IP address and type in a password provided on the whiteboard.

After the workshop the participants were allowed to continue accessing their notebook on the Jetstream platform for a limited time using the Education Allocation for the workshop. After the allocation expired, each individual could either:

  • install Docker for Mac or Docker for Windows to download and run the container on their own laptop
  • and/or apply for their own Startup and Research Allocations on XSEDE Jetstream

This project represents a first step towards a flexible and easy way to deploy computational environments on existing cloud infrastructure for the purposes of teaching data analytic methods to scientists. Research IT and BIDS will refine this process in the coming months in order to accommodate new research domains and training events, and to make it more straightforward for instructors to set up course infrastructure without need for exceptional technical knowledge.

We invite you to reach out to us if you would like access to the resources described above for your own research, or to support running a workshop of your own. Please send email to: research-it@berkeley.edu.


BIDS Fellow Chris Holdgraf contributed to this article.