Power is the most important resource on the next-generation supercomputers, and they will be operated under power constraint. Therefore, there is a need to maximize performance of HPC application under power constraint. To do such optimization, we've developed and reported a method to improve performance by power allocation for each processor, which is called the variation-aware power budgeting. In this study, we carried out large-scale performance evaluation of a proposed method for two mini-applications related to molecular science, Modylas-mini and NTChem-mini. As a result, our method can improve their performance under power constraint up to 1.99 times speedup compared to conventional power constraint.