An appropriate computing system and its system parameters selection based on bottleneck prediction of applications

Recent computing systems have different characteristics as they consist of various processors such as a scalar processor, an accelerator, and a vector processor. As calculations patterns best suited for individual systems are different and diverged, it is difficult to determine in advance which computing system is appropriate when an application is given. Furthermore, due to the increase in the complexity of computing systems such as the many-core and new memory technologies, it is necessary to adjust many system parameters and find appropriate system parameters for each application to deliver the high performance of a computing system. One of the ways to find a computing system and its system parameter combinations suitable for an HPC application is a bruteforce approach of trial and error executions of an application on various computing systems and various system parameter combinations, which is no longer a realistic way due to its high cost of time and efforts. This paper proposes a method to carefully select a candidate of a computing system and system parameter combinations appropriate for executions of an HPC application by considering both characteristics of the application and computing systems. Thus, the proposed method can narrow down a large search space of computing systems and their system parameter combinations suitable for executing the HPC application.