In-situ feature extraction has become increasingly important for understanding complex fluid behaviors in hydrodynamics simulations. However, achieving high accuracy without introducing significant overhead remains challenging. We propose a real-time, in-situ analysis approach performing auto-regression on temporal and spatial data as simulations progress to predict future states of simulation. Our method uses a linear autoregressive model and mini-batch training scheme to ensure real-time adaptability, enabling efficient extraction of dynamic features with good speed-ups. We demonstrate the method's accuracy and scalability through applications to material deformation analysis (LULESH) and white dwarf (WD) merger detonation detection (Castro), achieving 94.44% to 99.60% accuracy and$2.40 \times$to$3.03 \times$speed-up.

Efforts to optimize parallel scientific applications have focused on enhancing metrics related to either the computing aspects, such as execution performance, or the computational aspects, such as computation precision. However, attempting to simultaneously optimize both aspects often proves unproductive due to the differing expertise, benchmarks, and applications utilized by computer scientists and computational scientists. This paper introduces a Runtime Tuning (RTune) method designed to facilitate automated, runtime-time analysis and adjustment of application and system parameters during the simulation process to achieve user-defined objectives. The RTune API allows users to program computing and computational objectives within parallel iterative programs, while the RTune system manages measurement, modeling, and analysis based on these objectives throughout the computation process. We have evaluated the RTune method through three use cases. In simulations of wave propagation and for the objective of finding region of interest of the simulation, an RTune-optimized LULESH can decrease the number of iterations needed by 48 % to 98 % compared to a full simulation. For solver acceleration, an RTune-tuned Jacobi kernel can automatically adjust iteration counts based on the desired solver result precision. Additionally, for parallel applications, RTune can automatically determine the optimal thread count to minimize execution time, demonstrating its capability to efficiently configure system settings. The RTune library is available as open-source software under the BSD license and can be accessed from https://github.com/passlab/rtune.