Analysis And Visualization Of Memory Access Characteristics In Heterogeneous Memory

Heterogeneous memory is a hot research area in memory computing.Heterogeneous memory consists of a small amount of traditional memory and a large number of nonvolatile new types of memory.In heterogeneous memory,it is important to study the memory access behavior of an application and optimize memory access for program and system performance.In the current research on the analysis of memory access behavior,the accuracy of the data obtained by these technologies is high,but the semantic ability of the data is not strong.It is impossible to help researchers understand the access mode of the application deeper.Memory computing requires a fine-grained memory access analysis solution.An analysis and visualization tool of memory access features is proposed: HMProfiler.HMProfiler is designed for heterogeneous memory features,supports tracking of multithreaded programs,and finally analyzes and visualizes the collected memory access data.First,it uses the Intel Pin tool to collect fine-grained application access data for specific memory objects and function bodies.By simulating the CPU cache process,the memory access instructions are filtered to improve the accuracy of the data.Then,the collected memory data will be analyzed.The memory access data is summarized into the corresponding memory objects.The memory objects of the application runtime are associated with the variable in the source code.It calculates performance data for memory accesses at the granularity of memory objects.Finally,the memory access data is persisted and visualized to facilitate subsequent re-emergence and secondary processing of the experimental data.The experimental results show that the memory access data of applications tracked by HMProfiler is true and reliable.Compared to the same type of Metric memory tracking tool,the performance of HMProfiler is much higher.In addition,the data interface of HMProfiler has an lower average latency and an higher average throughput.Through the visualization of HMProfiler,researchers can discover and diagnose performance issues in the application and be provided useful insights into the application's access behavior.