A Study On Flash Memory Three-Dimensional Stacked Structure For Big Data Applications

Nowadays, the development of the big data applications is rapid. The different data characteristics with the traditional desktop applications have led to a big challenge of the processor's performance, and it also increase the distance between the computing capacity and the storage capacity. Three-dimensional stacked storage structure has the advantagements of low latency, large capacity, high bandwidth, and these can effectively relieves the problems of storage wall. For the aspect of storage media, beyond the nonvolatile characteristic, nonvolatile storage medium has the superiority in the aspects of capacity and energy. The non-volatile storage stacked structure has combined the superiority of the three-dimensional structure and the characteristics of non-volatile storage, which makes it necessary to be in development. Flash memory is as the most mature storage medium, it has the significance of reality. This paper mainly studies on the flash memory three-dimensional stacked structure facing with the big data applications.Firstly, according to the current situation of the three-dimensional stacked storage structure, we found the four aspects of the key research directions including the storage level stacked, the use of storage medium, the communication with the stacked structure and faced with the big data applications. We mainly focus on the DRAM(dynamic random access memory) stacked structure and NVM(non-volatile memory) stacked structure, and to summarize the characteristics of three-dimensional stacked storage structure. Based on these, the flash memory stacked structure is focused on and briefly introduced the new type of non-voliate storage stacked structure, which can point out the research thought in the view of big data applications.Secondly, by using the method of simulation and test, we make an analysis on the characteristics of the flash memory stacked structure. We select the typical variables can reflect the characteristics of the flash memory and the common characteristics when it being a common non-volatile storage. Uses these variables to make a test throgh observing the memory access delay. By analysing the multiple results, it has summared four typical characteristics: one is the bandwidth of the three-dimensional stacked structure has a significant effect on the big data applications, the second one is that with the increasement of bandwidth advantages, the incfluence of the latency of the flash read, write or erase operations is gradually reduced, the third one is the page size influenced little for the delay results, the last one and important one is the capacity of flash memory layer brings a great influence, especially when the channel number is big. Beyond of these, it also gives the next optimized points.After stacked flash memory layer on top of the DRAM layer, through the optimized points because of the difference between flash memory and DRAM, for optimizing the flash memory stacked structure and taking advantagement of the difference, two optimization strategies are proposed in this paper. They are granularity difference strategy and read and write partition strategy. With the two strategies, the objective is to improve the system's memory access performance and facilitate data management, and according to it, the design thought and realization are given.Finally, for finding the effect of the opmization strategies, the paper gives the implementation process, and test the memory access delay with the original flash memory stacked structure and the optimization stacked structure. Based on the existed code to modify and use tools to collect the necessary information on the access trace, an easy simulator is realized. Using the simulator and choosing appropriate experiment parameters, some results are got. The experimental results show that the two strategies can improve the performance of the memory access to the original flash memory stacked structure. Under the granularity difference strategy, the performance increased by 11.3%, with the partition strategy, it is by 5.4%, and it has the benefit of write endurance. Using the two strategies, it can be increased by 17.2%.Above all, faced with the big data applications, this paper mainly makes an analysis on the characteristics of the flash memory stacked structure, and gives optimization strategies. This can offer an important significance for further study.