Optimizing I/O for an Exascale Implicit Kinetic Plasma Simulation using the Rabbit Storage System

I/O optimization for an exascale implicit kinetic plasma simulation on LLNL’s Rabbit near-node NVMe storage.

Exascale computing revolutionizes scientific research by facilitating large-scale simulations and data analysis. As applications generate massive, heterogeneous data streams, the I/O subsystem increasingly dominates runtime, limiting scalability and time-to-solution. Emerging storage accelerators such as the Rabbit system offer a promising path forward. Rabbit provides dynamically configurable hybrid storage, combining high-bandwidth, node-local NVMe (e.g., XFS) with shared, distributed file systems (e.g., Lustre), to match diverse I/O patterns more efficiently. In this work, we evaluate Rabbit's capability to mitigate I/O bottlenecks using iPIC3D, a representative implicit Particle-in-Cell (PIC) code for exascale scientific workloads. We perform a detailed characterization of the key phases of I/O (restart, field, and moment) and identify the dominant access patterns in these phases. We then benchmark these patterns using IOR in Rabbit hybrid storage configurations, identifying the ideal performance achievable for each access pattern under node-local and distributed storage modes. We use phase-aware mapping of these patterns to Rabbit storage systems, achieving a 4.85 times improvement in I/O throughput, reducing the I/O share of runtime from 38% to 11% and delivering an end-to-end speedup of 1.45 times. These results outline the importance of hardware-software co-design and highlight Rabbit as a scalable, data-intensive storage solution.