Study On Performance Optimization Of Hybrid Memory System Based On Non-volatile Memory

Dynamic Random Access Memory(DRAM)has been widely used in computer systems over the past few decades due to its excellent read-write performance.With the development of computer technology and the advent of the era of big data,storage systems present problems of "storage walls" and "energy walls",because DRAM is limited by low storage density and high refresh energy consumption.The emerging Non-Volatile Memory(NVM)has the advantages of non-volatile,byte-addressable,zero leakage power consumption,high scalability,etc.It is considered to be a potential replacement for traditional memory devices.However,Non-Volatile Memory also has defects such as high write latency,high write power consumption,and limited write endurance,so that it cannot directly replace traditional memory devices.The hybrid memory architecture based on DRAM and NVM can make full use of the advantages of DRAM and NVM.Most existing researches on hybrid memory data page management identify cold and hot data by writing counts or historical heat statistics,which greatly increases the computation and storage overhead.In addition,the unconditional migration proposed some research will produce a large number of invalid migrations and reduce system performance.In response to the above problems,this thesis makes the following contributions:First,in response to the problem of frequent migration of hot/cold data in a hybrid memory system and a large amount of computing overhead,which results in performance degradation,a hybrid memory data management strategy(Data Management with Appropriate Writing,DMAW)with a small amount of calculation overhead and allowing appropriate NVM writing is proposed.DMAW guides data distribution and data updates by combining time locality and access interval to accurately identifies hot and cold,to maintain hot data in DRAM and cold data in NVM as much as possible.At the same time,allowing appropriate NVM writes through state transitions in NVM reduces a lot of unnecessary data migration.Second,due to the shortcomings of limited write endurance of NVM,this thesis proposes a hot-address wear leveling(HAWL)strategy based on hot address maintenance.This strategy balances the number of writes to the NVM memory space by maintaining the hot addresses generated by data migration during the data management process,thereby extending the life of the NVM.Final,this thesis chooses ten classic benchmark programs in SPEC CPU 2006 to experiment with the proposed strategy and gives detailed results analysis.Experiments show that DMAW can reduce NVM write operations and data migration with a small amount of computing overhead and can improve system performance by up to 75%.In addition,HAWL can effectively balance the wear rate of NVM storage units,which can extend the life of NVM by 7.2 times on average.