Rudder: Steering Prefetching in Distributed GNN Training using LLM Agents
#Graph Neural Network #Distributed Training #Prefetching #LLM Agent #AWS DistDGL #In‑Context Learning #Zero‑Shot Tasks #NERSC Perlmutter #Performance Improvement #Communication Overhead
📌 Key Takeaways
- Rudder embeds LLM agents into the AWS DistDGL framework to perform adaptive prefetching.
- The system leverages emerging properties of contemporary LLMs (in‑context learning, logical multi‑step reasoning) to make zero‑shot decisions.
- Evaluation on standard datasets and unseen configurations on Perlmutter achieved up to a 91% speed‑up over baseline DistDGL (no prefetching) and an 82% improvement over static prefetching.
- Communication volume was cut by over 50% compared to static schemes.
- The paper demonstrates a practical application of generative AI for real‑time control in high‑performance computing.
- Source code is publicly available on the project’s repository.
📖 Full Retelling
Aishwarya Sarkar and colleagues (Sayan Ghosh, Nathan Tallent, Aman Chadha, Tanya Roosta, Ali Jannesari) introduced **Rudder**, a software module that autonomously steers prefetching in distributed Graph Neural Network (GNN) training using Large Language Model (LLM) agents. The work was completed on the NERSC Perlmutter supercomputer, submitted to arXiv on 26 Feb 2026, and accepted for presentation at the 40th ACM International Conference on Supercomputing (ICS 2026). The goal was to reduce irregular communication stalls and improve end‑to‑end training performance by adapting prefetching policies to dynamic graph, distribution, and sampling conditions.
🏷️ Themes
Distributed graph neural networks, High‑performance computing, Communication optimization, Adaptive prefetching, Large language models, Zero‑shot learning
Entity Intersection Graph
No entity connections available yet for this article.
Original Source
--> Computer Science > Machine Learning arXiv:2602.23556 [Submitted on 26 Feb 2026] Title: Rudder: Steering Prefetching in Distributed GNN Training using LLM Agents Authors: Aishwarya Sarkar , Sayan Ghosh , Nathan Tallent , Aman Chadha , Tanya Roosta , Ali Jannesari View a PDF of the paper titled Rudder: Steering Prefetching in Distributed GNN Training using LLM Agents, by Aishwarya Sarkar and 5 other authors View PDF HTML Abstract: Large-scale Graph Neural Networks are typically trained by sampling a vertex's neighbors to a fixed distance. Because large input graphs are distributed, training requires frequent irregular communication that stalls forward progress. Moreover, fetched data changes with graph, graph distribution, sample and batch parameters, and caching polices. Consequently, any static prefetching method will miss crucial opportunities to adapt to different dynamic conditions. In this paper, we introduce Rudder, a software module embedded in the state-of-the-art AWS DistDGL framework, to autonomously prefetch remote nodes and minimize communication. Rudder's adaptation contrasts with both standard heuristics and traditional ML classifiers. We observe that the generative AI found in contemporary Large Language Models exhibits emergent properties like In-Context Learning for zero-shot tasks, with logical multi-step reasoning. We find this behavior well-suited for adaptive control even with substantial undertraining. Evaluations using standard datasets and unseen configurations on the NERSC Perlmutter supercomputer show up to 91% improvement in end-to-end training performance over baseline DistDGL (no prefetching), and an 82% improvement over static prefetching, reducing communication by over 50%. Our code is available at this https URL . Comments: Accepted to the 40th ACM International Conference on Supercomputing (ICS 2026) Subjects: Machine Learning (cs.LG) ; Artificial Intelligence (cs.AI); Distributed, Parallel, and Cluster Computing (cs.DC); Multiage...
Read full article at source