Cache Performance Improvement Method Based On Data Layout Optimization

Memory access latency is one of the performance bottlenecks of most modern applications.Data layout optimization can improve the memory access performance of programs by improving the local behavior of program access cache,which is one of the effective methods to improve cache utilization.Most modern memory intensive applications have more structure data,and the memory access performance of structures has become a key factor to improve the performance of such programs.Designing an effective structure layout optimization scheme can improve the local performance of the cache,which is of great significance to the performance optimization of the program.This paper studies the cache performance improvement based on the data layout optimization of the structure.The main work is as follows:(1)Aiming at the problem of analyzing the behavior of structural field access,a spatiotemporal model of structural field access is established.First,according to the statistical information of field access,the fields of the structure are separated from each other.Then,by combining the time stamp information of hot field access and its spatial information in the cache,the relative access times between hot fields are counted.Finally,based on Apriori algorithm and periodic correction dynamic time warping method,the similarity of memory access behavior between hot fields is calculated.(2)A cache performance modeling tool based on memory access trace is designed and implemented to solve the problem of selecting optimal scheme for structure layout.Based on the program memory access trace information generated by dynamic detection,the cache behavior of the program is modeled to compare the cache performance of the optimization schemes under different parameters,and then the most appropriate structure layout optimization scheme is selected according to the optimization objectives.(3)Aiming at the optimization of structure field reorder,a structure field reorder algorithm based on Louvain clustering is proposed.Firstly,the Louvain community is constructed according to the relative memory access count between the hot fields of the structure and the similarity of memory access behavior.Then,the final reorder classification scheme is formed through multiple community tests on the hot fields.Finally,considering the impact of structure field alignment on cache performance,alignment optimization is carried out for each category of the reorder scheme.(4)Aiming at the problem of structure splitting optimization,a structure splitting optimization algorithm based on field memory access spatiotemporal model is proposed.According to the spatio-temporal model of field memory access,combined with the idea of approximate optimal segmentation,a heuristic optimization algorithm for structure splitting is implemented to generate the optimization scheme for structure splitting.For different split scenarios,pointer index optimization and memory pooling optimization are implemented respectively.Compared with the cutting-edge structure splitting optimization method,the optimization performance is further improved,and the average time speedup ratio of the benchmark program reaches 1.37x.To sum up,this paper proposes a new structure layout optimization model and algorithm,and has achieved remarkable optimization results on the open benchmark program.