Production-scale Performance Analysis and Optimization of Vlasiator

As HPC systems scale toward the exascale era, the ability to efficiently utilize the growing number of parallelism available becomes critical. Traditional performance validation methods - such as scalability studies, and aggregated performance metrics obtained by profilers and mini-apps - often fall short when applied to complex, full-scale production runs. In this paper, we present a novel methodology for in-depth performance analysis of large-scale simulations, demonstrated on Vlasiator, a space plasma simulation code executed on up to 500 nodes of a European pre-exascale supercomputer. Our approach combines targeted instrumentation, programming model-based performance metrics, and iterative investigation to identify and address performance bottlenecks that are otherwise hidden in conventional analyses. The methodology led to optimizations, resulting in improved parallel efficiency and measurable speedups. This work highlights the importance of production-level analysis for HPC applications and provides a practical framework for similar efforts across other scientific domains.