Browsing by Author "Pan, Qifeng"

Filter results by typing the first few letters
Now showing 1 - 1 of 1
  • Thumbnail Image
    ItemOpen Access
    Code intrinsic uncertainty quantification
    (2025) Pan, Qifeng; Resch, Michael (Prof. Dr.-Ing.)
    Uncertainty quantification (UQ) has been widely applied in numerical simulations in recent years. UQ deals with estimating the statistical output from the quantities of interest given a computational model and probability distribution of input. The Monte Carlo (MC) method has played an important role in the field of UQ due to its simplicity and robustness. However, the MC method suffers from a low order of convergence rate and low execution efficiency in high-performance computing (HPC) systems. To improve the efficiency of the MC method, this thesis proposes a new framework method: the code intrinsic UQ (CIUQ). The core idea of CIUQ is to transform the deterministic code (DC) into an uncertainty intrinsic code (UIC). The UIC extends the variables by one extra dimension holding all statistical information of uncertain variables. Instead of evaluating the instances generated by MC one by one, the UIC is able to conduct all results via one invocation. Compared to the DC, the UIC has fewer redundant operations with higher calculation intensity, ultimately resulting in higher efficiency on an HPC system. To transform a DC into a UIC, the interactions of uncertain variables through the code will be explored. Specific algorithms are developed based on traversing the Abstract Syntax Tree (AST) of the original code. This exploration of uncertainty propagation details leads to a comprehensive understanding of variable dynamics within the UIC. Based on the algorithms, the ATUQ toolkit is able to perform the two algorithms to accomplish a general uncertainty propagation exploration for Fortran code by employing the component of ROSE compiler front-end, In addition, a general procedure of automatically transforming a DC to UIC is also developed based on ROSE. The ATUQ toolkit offers a potential solution for domain-specific compilers for MC simulations. This thesis performed the transformation methodology on different numerical solvers. The transformation methodology showed its generality and capability to reveal how uncertainty propagates through the code. The further numerical experiments demonstrate that the UIC can improve the vectorization of recursive loops, and take advantage of single instruction multiple data schemes on NEC and Intel machines. Moreover, the UIC framework can further increase the data intensity of computational kernels and thus significantly improve strong scalability. These improvements, when compared to standard MC approaches, showcase the effectiveness of the new framework across various types of parallelism, significantly enhancing both efficiency and scalability.