Catalyst: Expanding In Situ Analysis and Visualization Workflows for Exascale Computing

Modern GPU-accelerated supercomputers expose a growing imbalance between computational throughput and I/O bandwidth, making traditional post-hoc workflows increasingly impractical for large-scale simulations. This work introduces two advances to the Catalyst 2 in situ framework that reduce data movement and improve hardware utilization on heterogeneous nodes. First, we present a production-ready, zero-copy workflow that allows simulations to pass GPU-resident mesh and field data directly to Catalyst without staging through host memory. By extending Catalyst’s Conduit interface with memory-space descriptors and constructing non-owning Viskores ArrayHandles, data remains on the device throughout the analysis pipeline. This removes redundant device–host transfers and reduces memory footprint, with support available in ParaView 6.0 and later. Second, we outline an asynchronous on-node analysis model that uses idle CPU resources to process data concurrently with GPU-based simulation timesteps. By overlapping lightweight extraction, statistics, or ML inference with ongoing GPU computation, this approach can significantly reduce or fully hide, the apparent cost of in situ operations. Together, these capabilities improve scalability and efficiency for GPU-centric simulation workflows on exascale systems such as Frontier, Aurora, and El Capitan.