Shared Memory Parallelization Strategies and Speedup Scaling of the Hygrothermal Building Component Simulation Program HAM4D_VIE

Hygrothermal building simulation is helpful for predicting and enhancing building performance, durability, and energy efficiency. Simulating different scenarios allows for better design, addresses concerns like mold, material decay and energy waste, which is increasingly important in light of climate challenges. Given the computational intensity of such simulations, the efficient use of computational resources like available CPU cores is paramount for manageable execution times.The simulation program, which has been developed and used in research and teaching at TU Wien for many years, does not rely on linear algebra operations or linear solvers like many other simulation algorithms, but rather implements an alternative approach iterating over cell and face objects of the construction component’s mesh. Using this program as an example, this paper outlines challenges and solutions in the parallelization process of such an algorithm using multithreading with the OpenMP API. Different implementation strategies are evaluated with respect to their resulting execution time, and speedup scaling is measured on an office machine as well as on compute nodes of the VSC-4 supercomputer.Execution time measurements on both machines reveal a speedup of several folds which shows that our implementation improves the overall performance of the simulation. However, when comparing the degree of parallelization suggested by Amdahl’s Law with the one suggested by performance profiling, a significant difference was observed, which indicates substantial overhead due to thread management. This underscores the need for further optimization in both parallelization techniques and the used data structures.