15–18 Dec 2026
Duke University
America/New_York timezone
Artificial Intelligence for the Electron Ion Collider (2026)

AI-Enabled Operations and Predictive Resource Management in PanDA

Not scheduled
20m
LSRC B101 Love Auditorium (Duke University)

LSRC B101 Love Auditorium

Duke University

Durham, NC 27710
AI/ML in Production and Distributed ML AI/ML in Production, Distributed ML

Speaker

Tasnuva Chowdhury (Brookhaven National Laboratory)

Description

PanDA (Production and Distributed Analysis) is a large-scale workload management system supporting distributed scientific computing for ATLAS and expanding to ePIC. As computing infrastructures and scientific workflows grow in complexity, integrating AI into workload management offers opportunities to improve operational efficiency, resource utilization, and reliability. This talk presents ongoing efforts to incorporate AI into PanDA through operational assistance, system understanding, predictive resource management, and intelligent monitoring.

For operational assistance, AskPanDA combines large language models (LLMs), knowledge retrieval, and log diagnostics to support operators. PandaMCP complements these capabilities by exposing PanDA APIs as tools through the Model Context Protocol (MCP), allowing LLM-based assistants to access operational information and interact with PanDA functionality. Its dynamic API-to-tool mapping reduces the need to implement and maintain individual tool integrations. Bamboo V2 extends this direction through code-derived graphs and graph-guided, multi-hop reasoning, connecting system structure and dependencies with operational evidence to support diagnostics and action planning.

For predictive resource management, deep neural network (DNN) and graph neural network (GNN) models use historical workload data to estimate resource requirements and workflow turnaround times. These efforts are being extended toward a PanDA Digital Twin: a data-driven framework for exploring resource behavior, evaluating scheduling strategies, and supporting future real-time optimization.

For intelligent monitoring, an AI-enhanced interface facilitates access to operational information across distributed PanDA resources. Complementary work uses LLM orchestration to generate site-status reports that help operators interpret site conditions and identify potential operational issues.

Together, these initiatives illustrate a multifaceted approach to AI integration in distributed workload management, linking operational knowledge and system context with predictive models and monitoring information. The talk will discuss their development and integration, with a focus on supporting scalable scientific computing for ATLAS, ePIC, and other data-intensive scientific applications.

Author

Tasnuva Chowdhury (Brookhaven National Laboratory)

Presentation materials

There are no materials yet.