(RP13) An Application Parameter Search Method Based on the Binary Search Algorithm for Performance Tuning
Performance Analysis and Optimization
Scientific Software Development
TimeWednesday, June 19th10:10am - 11am CEST
LocationAnalog 1, 2
DescriptionBecause of the complexity of modern computer systems, such as many-core processors and heterogeneous memory systems, performance tuning is necessary to achieve high performance in scientific and engineering applications. The performance tuning needs to adjust many tuning parameters provided by systems and applications. As the number of parameters increases, the number of their combinations to be searched increases. Moreover, the execution times of simulation applications become long because of their problem sizes and resolutions to solve. This situation causes an increase in a time for selecting an appropriate combination of the tuning parameters. This poster proposes the method to reduce the time to search the appropriate combination of the tuning parameters by the method based on the binary search algorithm. The proposed method select the appropriate application parameters such as inputs and iteration counts whiling keep the feature of the application. By executing the application with the smaller resolution and iteration counts than with original ones, each execution time for parameter tuning can be reduced. The evaluation results by using the tsunami simulation on Intel Xeon Phi Knights Landing show that the proposed method can select the most appropriate combination of the tuning parameters, and the parameter tuning time is reduced by 78.7% compared with the full search algorithm.