Parsl at OLCF
30 Sep 2022 - Ketan Maheshwari, Sean Wilkinson, Tyler Skluzacek, and Rafael Ferriera da Silva (NCCS and OLCF, ORNL)
In this blog post, we discuss three developments surrounding Parsl at OLCF.
First, the setup, management, and support of the Parsl software module on the Summit supercomputer. Second, an ongoing project where we are using Parsl to bridge the gap between Quantum Computing and HPC. Third, another ongoing project where we are developing a distributed task orchestration system and are in the process of designing a Parsl plugin.
In January 2022, we analyzed the usage data on Summit login nodes that we have been collecting hourly since January 1, 2020. While looking at the workflow systems usage in this data, we found that Parsl is one of the top five workflow systems used on Summit.
More information about our analysis and the data itself may be found in our ICCS’22 paper and the associated published data.
After looking into the results of this analysis, we approached the core OLCF admin team and requested them to let us make Parsl and other workflow systems available to all Summit users as modules. The request was approved.
OLCF admins created /sw/summit/workflows and let us install Parsl there. Parsl is installed as a conda
package that is based on the pre-installed Anaconda module on Summit. A module definition was written and
the module was set up for Parsl.
One question we would like to ask the community is: Is this the right approach to make Parsl available to the users of a cluster? Are there any best practices surrounding this topic that we should be aware of to make using Parsl more effective? Please reach out to us at km0@ornl.gov.
Alongside the installation and setup, we put together documentation and gave a tutorial / training talk about using workflows at OLCF resources. The talk video and slides are available.
In another development, earlier in 2022, we applied for and were awarded a small allocation on the ORNL Quantum Computing User Program (QCUP). The objective of this project is to combine the classical HPC and Quantum Computing using the workflows paradigm as a bridge.
One idea we explored is to make HPC generate input data and send this data along with quantum algorithms, implemented as python-qiskit programs, to the remote quantum systems, run them there and obtain the results back – with all the steps coordinated with Parsl. We have some early success on this work that was performed jointly with summer intern student Samuel Bieberich.
The linked Github repo has more info and code.
Last but not least is ongoing work where we are developing a distributed task orchestration system that can seamlessly orchestrate science campaigns, initially across the OLCF systems and ultimately across the DOE complex. We call it Zambeze.
For Zambeze, we plan to reuse existing solutions where possible as a plugin-based architecture and are currently developing a Parsl plugin for handling user-submitted workflows.
Acknowledgements: This work is supported by UT-Battelle, LLC, under contract DE-AC05-00OR22725 with the US Department of Energy (DOE) and the Exascale Computing Project (17-SC-20-SC), a collaborative effort of the U.S. DOE Office of Science and the NNSA.