Design And Verification Of GPU Cache Miss Analysis Model Based On Reuse Distance

Over the past decade, GPU has gradually evolved from a dedicated graphics processor into a general purpose computing platform. Due to its powerful parallel computing capabilities and power control capabilities, GPU general computing has been widely applied in scientific computing area. Since most of GPU chip area is assigned to computing units, the performance of a large number of GPU applications is limited to memory access speed, rather than computing power For GPU memory-limited applications, cache efficiency has a significant effect on the overall performance. To help application developers understand GPU cache behavior characteristics and choose the appropriate cache optimization method, accurate, fast, and full-featured GPU cache miss analysis tool is particularly essential.In this thesis, the author optimizes reuse distance algorithm accuracy and speed on GPU platform based on GPU cache micro-architecture feature, and designs a GPU cache miss analysis model based on reuse distance theory. First the accuracy of reuse distance algorithm under GPU platform is optimized by modifying reuse distance algorithm core data updating time and amending GPU thread blocking mechanism; Second, the speed of reuse distance algorithm under GPU platform is optimized by dividing reuse distance calculating task to sub tasks and merging memory trace records to save storage space. Third, a GPU cache miss analysis model is designed base on reuse distance theory. In addition to traditional cache miss type analysis, cache size sensitivity analysis and instruction-level cache miss analysis are implemented to help GPU application developers better understand GPU cache miss characteristics.Based on verification experiment on PolyBench/GPU and Parboil benchmarks, after optimization, the average error of GPU platform reuse distance algorithm decreased from 13.6% down to 5.74%, the average time cost reduced from 7749ms to 1297ms. Besides, the GPU cache miss analysis model is used to analyze cache miss behavior characteristics of typical GPU applications and choose the right cache performance optimization method. Result shows that, cache miss rate and execution time of the GPU applications dropped significantly, validating the practicality of our GPU cache miss analysis model.