Compression And Wear-Leveling For 3D-stacked Non-Volatile Memory

In the era of big data,the amount of data to be processed is huge and the generation speed is fast.With the increasing memory footprint of big data applications,today's computer systems require large amounts of memory for high-performance data processing.Traditional Dynamic Random-Access Memory(DRAM)has been increasingly unable to meet the demand of large memory for big data applications due to the challenges of manufacturing technology.Compared with DRAM,emerging Non-Volatile Memory(NVM)usually has the advantages of high density,low static power consumption and low cost.However,the disadvantages of NVM such as high write latency,high write energy consumption and limited write lifetime make it a great challenge to replace DRAM.Aiming at the shortcomings of NVM,such as limited write lifetime,high write latency and high write power consumption,the 3D-stacked NVM memory compression system is designed and implemented,which reduces the amount of data written to NVM and extends the lifetime of NVM.Firstly,combined with the characteristics of high proportion of zero data and frequent locality of memory data,a ZD-FVC data compression strategy based on zero data and frequent values is designed.ZD-FVC makes full use of the characteristics of high zero-data ratio in most program memory,optimizes the design mode of FVC,and greatly reduces the storage overhead of compression coding and improves the compression ratio.Then,by implementing the compression/decompression unit for the above data compression strategy on the logic layer of the 3D-stacked NVM memory,the data amount of the write operation is reduced by data compression,thereby reducing the write delay and write power consumption of the NVM.Furthermore,the data placement strategy in NVM is studied to efficiently manage the compressed data,and the storage space saved by data compression is used to realize a rotation-based wear leveling mechanism,so that the write operation can be distributed as evenly as possible to extend the lifetime of NVM.Using a hybrid simulator consisting of gem5 and NVMain,12 representative applications were selected from the SPEC CPU 2006,and seven systems were compared.Experimental results show that compared with the compression system using frequent value compression algorithm,the 3D-stacked NVM memory compression system can improve data compression ratio by 55%,reduce the write latency of NVM by 41%,reduce the write energy consumption of NVM by 20% and extend the lifetime of NVM by 74%.