Sept 2026
Paper accepted at NeurIPS2026-AI4Science
- Agentic Job Scheduling for Scientific Distributed Computing via Active Knowledge Extraction
Performance Modelling · AI systems · Distributed Computing · Photonic Computing
Sairam Sri Vatsavai
Assistant Computational Scientist, Computing and Data Sciences Department,
Brookhaven National Laboratory, Upton, New York
My research focuses on understanding and improving computing systems across two complementary fronts. First, I evaluate emerging computing technologies like superconducting digital logic , photonics and in-memory computing , by developing analytical and AI-enabled modeling frameworks that characterize their performance, identify their potential, and enable exploration of the cross-layer design space spanning device, architecture, and applications. Second, I study existing large-scale distributed computing systems for scientific workflows, with an emphasis on improving their resilience and efficiency through better workload modeling, scheduling, and resource-management policies. Together, these efforts aim to build principled tools for understanding both future computing platforms and the complex systems that support large-scale science today.
Before joining Brookhaven National Laboratory, I completed my Ph.D. in Electrical Engineering at the UCAT Lab at the University of Kentucky in spring 2024, under the guidance of Dr. Ishan Thakkar. My doctoral research focused on the cross-layer design and optimization of photonic integrated-circuit-based AI accelerators.
News
Sept 2026
Paper accepted at Transactions on Computers
- A System-level Exploration of Superconducting Supercomputers for Emerging Workloads
Sept 2026
Three workshop papers accepted at SC26 in Ai4S, AionHPC and PMBS workshops
- LightshowAI-Agent: An MCP-Based Agentic Framework for Real-Time and Reproducible X-Ray Absorption Spectral Analysis at Scientific User Facilities
- HOLOS: A Gradient-Descent-Based Design-Space Exploration Framework for Hierarchical AI Accelerator Systems
- DMSim: Configurable Data Management Simulation Framework for Scientific Computing Grid
Aug 2026
Delivered a lightning talk and presented a poster at ModSim 2026, Seattle, WA
Presented “AI-Enabled ModSim for Distributed Computing: A Simulator-to-Surrogate Methodology for Scalable Performance Modeling.”
Aug 2026
Two Papers Accepted at IEEE Generative AI for Secure Systems (GAISS)
- HorizonGuard: Measuring Hindsight Contamination in Generative AI Cyber Threat Intelligence
- ChronoEntity: Auditable Agentic AI for Point-in-Time Entity Integrity in Financial Data Pipelines
July 2026
Paper accepted at IEEE eScience 2026, Naples, Italy
“CGSim: A Simulation Framework for Large Scale Distributed Computing Infrastructures with AI-Driven Applications” was accepted to the 22nd IEEE International Conference on eScience, to be held September 28 to October 2, 2026.
Jul 2026
Presented a talk and poster on the REDWOOD project at the ASCR Computer Science PI Meeting, on behalf of Alexei Klimentov
Presented progress updates and future directions for the REDWOOD project.
Jul 2026
Delivered a plenary talk at ATLAS Software and Computing Week, CERN, Geneva, Switzerland
Presented “Resilient Federated Workflows in a Heterogeneous Computing Environment: The REDWOOD Project.”
Jun 2026
Delivered a talk and served as a panelist at a mini-symposium during PASC 2026, University of Bern, Switzerland
Presented “AI-Enabled Modeling, Simulation, and Optimization of Distributed Computing Systems.”
May 2026
May 2026
Invited talk at Optics-Laser 2026, 4th Innovations in Optics, Photonics, and Laser Engineering Conference, Boston, MA
Delivered an invited talk titled “Cross-Layer Design of Highly Scalable and Energy-Efficient AI Accelerator Systems Using Photonic Integrated Circuits” on 18 to 19 May 2026.
May 2026
Promoted to Research Staff 3/Assistant Computational Scientist, at Brookhaven National Laboratory
Transitioned from Postdoctoral Research Associate to a research staff appointment in the Computing and Data Sciences Department, within the Systems, Architectures, and Advanced Technologies group.
Research
Two connected threads: modeling complex computing systems, and designing alternative hardware for AI.
Modeling and surrogates for distributed computing
Discrete-event simulators reproduce the behavior of production grids, and learned surrogates stand in for them once a search needs thousands of rollouts. The goal is scheduling-policy search at facilities where a single real experiment costs weeks of grid time.
Photonic and beyond-CMOS accelerators
Microring and interferometric meshes can carry many computations at once across wavelength, time, spatial mode, and polarization. My work spans the device models, the mapping from workload to degrees of freedom, and the benchmarking that decides whether any of it beats a GPU.
Publications
Selected peer-reviewed work, followed by manuscripts currently under review or in preparation.
2026
Predicting job turnaround time in large-scale distributed computing environments with graph neural networks
2026
MCP agents in PanDA
2026
Machine learning-driven predictive resource management in complex science workflows
2025
Simulating the ATLAS distributed computing infrastructure to develop and optimize workload allocation strategies
2025
Error analysis of globally distributed workflow management system
Projects
Current research efforts spanning scientific infrastructure, photonic computing, X-ray science, and system co-design.
REDWOOD
A DOE ASCR project on modeling and optimizing scientific cyberinfrastructure. I build the simulation and surrogate stack, CGSim and TRASSE, and the training pipelines that cover 74 WLCG sites.
COSMOS
A Brookhaven LDRD on simulation-driven co-design of multidimensional photonic computing systems. I am a co-investigator leading benchmarking and evaluation against GPU, TPU, and single-degree-of-freedom photonic baselines.
LightshowAI
An X-ray absorption spectroscopy platform for NSLS-II. I work on federated authentication, streaming beamline data, and the agentic layer that exposes surrogate models and materials databases as tools.
HOLOS
A gradient-descent design space exploration framework that searches parallelism strategy and accelerator geometry together, across CMOS and superconducting digital technologies.
Service
Reviewing activity separated by journals and conferences for faster scanning.
40+
papers reviewed to date
| Journals |
|
|---|---|
| Conferences |
|