Birds of a Feather Archive

HPC Developments for Fusion Energy Applications


Authors: Quynh Nguyen (NVIDIA Corporation), Shantenu Jha (Princeton Plasma Physics Laboratory), Justin Angus (Lawrence Livermore National Laboratory (LLNL)), Nitin Shukla (Cineca HPC Center)

Abstract: The growing interest in fusion energy to power massive data centers has created an urgent need to accelerate plasma science codes. Both experimental capabilities and computational workflows are complex and often render current technologies inefficient, which results in a severe lack of data. Implementation of plasma codes becomes complicated due to variations in software and hardware architectures. A potential solution is to develop an abstraction layer to connect infrastructure to application codes. This BoF focuses on identifying the needs for development and standardization of an abstraction layer, and the relevant functionalities/data structures that would be most useful to abstract.

Long Description: The growing interest to commercialize fusion energy as a power source for massive compute data centers have created an urgent need to accelerate plasma science codes. Both experimental capabilities and computational workflows are complex and often render current technologies inefficient, which result in a severe lack of data. In this BoF session, we want to gather inputs and build a community to inform upcoming computational developments that can enable implementation and optimization of plasma codes.

A primary focus of this session includes Particle-in-Cell codes. They directly benefit fusion energy by solving for the underlying physics that take place in a reactor, such as plasma properties and instabilities. However, they are computationally too expensive and typically requires filling largest HPC systems. These new scientific findings from PIC codes can help address current design challenges and generate datasets that can be used to train AI surrogate models. The models can become surrogates for engineering applications in fusion energy. Hence, we need to migrate from science apps to engineering apps.

Specifically, the challenges and complexity of implementation of PIC code include: • Implementation of PIC on different architectures. • Different phases of the code (grid, particles, with/without collisions) requires careful tuning for optimal implementation on a given architecture. • Implementation in a hardware agnostic way is challenging, as developers are forced to think of these low-level details. • On the other hand, these implementation and optimization pose challenges for hardware design and vendors to innovate as the application codes are tuned for a given architecture to reach optimal performance, including runtime and accuracy.

A potential solution is to develop a universal abstraction layer to help establish a clear separation of concerns. For example, there have been various efforts for establishing such a standard have been made in the past, such as community PIC codes including WarpX or PIConGPU to establish a layer that connects to other application codes, particle-based efforts such as CoPA to establish relevant APIs.

The goal of this BOF is to identify the desire for such a standard in developing an abstraction layer, and which functionalities/data structures would be most useful to abstract. In addition, we will discuss community approaches that include both AI and HPC for fusion/plasma codes.

By bringing together application developers, framework developers, hardware vendors, and academic and industry users, we will strive to identify the most important functions. Based on the outcome of this workshop, we will target future workshops and implementations.


Back to Birds of a Feather Archive Listing