The International Conference for High Performance Computing, Networking, Storage, and Analysis
The International Conference for High Performance Computing, Networking, Storage, and Analysis
Authors: Frank Indiviglio (National Oceanic and Atmospheric Administration (NOAA)), Ron Bewtra (Self), Manish Parashar (University of Utah), James Rogers (Oak Ridge National Laboratory (ORNL))
Abstract: The convergence of traditional HPC simulations and large-scale AI is reshaping data center infrastructure. This session addresses the unique challenges of supporting both modeling and machine learning workloads on a unified platform. Key topics include navigating the demands of specialized hardware like GPUs and custom accelerators, managing complex software stacks with containers and workflow managers, and optimizing high-performance storage for diverse I/O patterns. We'll also discuss scheduling strategies for fair and efficient resource allocation. Join this interactive discussion to share your experiences, challenges, and solutions for building and managing a truly converged HPC and AI infrastructure
Long Description: Topic and Relevance: This Birds of a Feather (BoF) session will explore the critical and rapidly evolving intersection of High-Performance Computing (HPC) and Artificial Intelligence (AI). The traditional lines between large-scale simulation and data-intensive machine learning are blurring, creating a new paradigm where AI is used to steer simulations, and HPC systems are essential for training massive models. This convergence presents a profound challenge for data center architects, administrators, and researchers who must design, deploy, and manage infrastructure capable of efficiently supporting both workload types. This topic is essential for any HPC professional navigating the technical and operational complexities of building a unified, future-ready scientific computing environment.
Goals: The primary goal of this session is to establish a collaborative space for the community to address the practical challenges of converged HPC and AI infrastructure. We aim to:
Identify and discuss the primary architectural and operational friction points when supporting both traditional modeling and AI/ML workloads.
Share real-world strategies and best practices for hardware selection, software stack management, and resource scheduling in a mixed-workload environment.
Foster a network of peers to share solutions and lessons learned, helping institutions avoid common pitfalls and accelerate their infrastructure's evolution.
Discussion Areas: This session will be a moderated, interactive discussion focused on the core technical hurdles of this convergence. Guiding topics will include:
Heterogeneous Hardware: What are the best strategies for integrating and managing diverse accelerators (GPUs, TPUs, etc.) alongside traditional CPUs? How do we balance system design for double-precision simulation versus mixed-precision AI training?
Software Stack Complexity: How can we effectively manage the competing software ecosystems of HPC (MPI, OpenMP) and AI (PyTorch, TensorFlow, JAX) on the same system? What is the role of containerization (Singularity/Apptainer, Docker) and workflow managers?
Storage and I/O Patterns: How do we design storage systems that can handle the large, sequential reads/writes of classic HPC and the small, random-access I/O patterns typical of AI training data pipelines?
Scheduling and Resource Allocation: What scheduling policies and tools can ensure equitable and efficient allocation of specialized resources (like GPUs) between long-running simulations and interactive, bursty AI training jobs?
Networking Demands: How do the networking requirements for large-scale AI model training (e.g., all-to-all communication) differ from traditional HPC interconnect needs, and how can we build a fabric that serves both well?
Facility Limitations: How can data centers sustainably meet the massive power demands of HPC and AI workloads? Are current facilities capable of accommodating emerging hardware? Can onsite power generation (e.g., SMR or renewables) offset grid limitations?
Expected Outcome: Attendees will leave this BoF with a comprehensive understanding of the key challenges and emerging solutions in converged HPC-AI infrastructure. Participants will gain practical insights from peer institutions, a clearer perspective on the current technology landscape, and new connections with colleagues who are designing and operating these complex, next-generation systems.