Research On The Optimization Of Page Management Based On DRAM/NVM

With the rapid development of embedded systems and the widespread use of data-intensive software,traditional memory systems based on dynamic random access memory(DRAM)have further deficiencies in high static power consumption,small storage density,and limited expansion.A hybrid memory system based on a new type of non-volatile memory(NVM)provides the possibility to solve these problems.However,NVM write operations generally have two shortcomings: one is poor durability,and the same location can only be written a limited number of times;the second is that write imbalance accelerates the wear of NVM.To this end,this thesis focuses on the optimization of NVM write operations in hybrid memory to reduce write wear and energy consumption of NVM.Specifically,this thesis mainly completed the following work:1.Aiming at the characteristics of poor NVM durability and unbalanced writing,a page classification strategy based on hybrid memory is proposed.When the system DRAM space is insufficient,the pages with read characteristics are preferentially replaced from DRAM to NVM,thereby reducing the write frequency of NVM.Improve its system performance and energy consumption.Secondly,from the write imbalance of NVM itself,the two-way queue algorithm is used to manage NVM free pages,so as to make NVM page frame wear more balanced and improve the life of NVM.2.Since the NVM free page updates the page,its bit flip cost is different,and a page allocation mechanism based on reservoir sampling is proposed.This mechanism finds the most similar free pages in the NVM free page set for allocation before allocating free NVM pages for the requested pages,thereby reducing the bit update of NVM storage units.In order to reduce the page comparison overhead,a page sampling sample was extracted using a reservoir sampling algorithm,and a page similarity calculation method based on Jaccard was designed to accelerate the calculation of sample similarity and allocate pages by the optimal page allocation algorithm.3.Using gem5 and NVmain simulators to simulate the NVM/DRAM hybrid memory architecture,based on the SPEC CPU benchmark test program,the proposed strategy was experimentally verified in terms of NVM write times,write distribution,energy consumption and delay.The results show that the page classification strategy in hybrid memory can significantly reduce the write operations in NVM,making more writeoperations occur in DRAM,and the two-way queue algorithm makes the distribution of write operations on NVM page frames more balanced,compared to traditional The random allocation method,the page allocation mechanism of the reservoir sampling proposed in this thesis,further reduces the write operation.In summary,the hybrid memory based page classification management strategy and page allocation mechanism based on reservoir sampling proposed in this thesis can effectively reduce the NVM write operation in the hybrid memory and reduce the system energy consumption.It has good theory and practical significance.