1 of 3

JUSTIN M WOZNIAK

  • Computer Scientist, Data Science & Learning, ANL
  • Scientist-At-Large, CASE, University of Chicago
  • Impacts of shared filesystem performance on real-time data acquisition and analysis. �Proc. NRDPISI 2025.

Swift/T Workflow System

Diaspora FS Prediction Model

Braid AI Provenance

  • Tracking dubious data: Protecting scientific workflows from invalidated experiments.�Proc. ReWoRDS 2022.
  • ExaWorks Software Development Kit: Interoperable Workflows Technologies.�Frontiers in HPC 2024.

2 of 3

ABSTRACT: STPR: STRUCTURED PARALLELISM

  • Lead: Mihael Hategan-Marandiuc; with Nicholson Collier, Amal Gueroudji, and Jonathan Ozik
  • Want to express higher-level concurrency patterns to the workflow runtime
  • Want to support AI-relevant algorithmic patterns such as optimization, classification, feature detection, etc.
  • Implemented with automated re-writing of Python functions via decorators and async / await
  • Enable easier construction of AI-controlled workflows with complex control patterns & concurrency

Concurrency Patterns and Primitives in Modern �AI/ML Scientific Applications

3 of 3

FUTURE OF WORKFLOWS

Many challenges that workflow techniques can address:

  • New applications need to integrate across data sources, computing approaches, numerical and text data
  • Increasing user detachment from software implementations and data sources
  • Possible technology borrowing from AI data tools, adapters, etc., will make workflow development and component integration easier
  • Pervasiveness of AI/ML control will enable targeted application searches
  • Increased user abstraction risks more easily erroneous results,�additional difficulties in reproducibility, regression analysis, human insight

More Computing, More Insight?