The Design of Discrete Memory-Accessing Library of Unstructured-grid for Domestic Heterogeneous Many-Core Architecture

Based on the “SunWay TaihuLight” system, it designs high performance batch-processing algorithm of unstructured-grid discreted accessing memory, which is suitable for different application scenarios. This method uses coordinate preprocessing technology to improve the bandwidth utilization of the slave core on-chip array by establishing a grouped message queue model, and uses barrier-free data distribution technology to make full used of the network performance of the domestic heterogeneous many-core architecture. Through testing different types of applications and combining performance models and experimental results to verify, it is found that the average effective memory bandwidth utilization of the algorithm is more than 70% under different grid models and scales, which is an average of 10 times and the highest performance acceleration is of up to 45 times compared with the master-core serial algorithm. The test and analysis of the application in many different fields have proved the universality of the algorithm, and though a large number of experimental results show that the algorithm has good acceleration performance for different grid types and scales in different application fields.