Porting Applications with OpenMP Using Similarity Analysis

Computer architecture has undergone dramatic changes due to technology innovation. Some emerging architectures, such as GPUs and MICs also have been successfully used for parallel computation in the today’s HPC field. Nowadays, people frequently have to port application to a new architecture or system and to expand its functionality for a better performance while in the meantime to meet the new hardware environment need. However, many scientific application legacy codes have a relative large size and long development cycle, so it’s a very challenging job to port legacy codes to a new environment. And current codes porting process is a manual, time-consuming, expensive and error-prone process, which requires a team of people work together. Barely any useful tools can be used to ease the porting process in High Performance Computing (HPC). In this paper, we present a tool called Klonos, which is designed for assisting scientific application porting. Based on similarity analysis of code syntax and cost-model provided metrics, we are able to find codes which can be optimized similarly without the need of profiling the codes. The proposed porting plan can systematically guide users for selecting subroutines in a way which maximizes the reuse of similar porting strategy. We evaluate Klonos by applying it to a real scientific application porting to a shared memory environment using OpenMP. According to our experiment result, which shows that Klonos is very accurate to detect similar codes which can be ported similarly.