Memory Management Mechanism For New Hybrid Memory Architectures

With the rapid development of large-scale memory computing,the existing Dynamic Random Access Memory(DRAM)have been unable to meet the requirements of performance,energy consumption and equipment area of some real time systems and high performance systems.Industry and academia are proposing to use hybrid memory system which consists of the new Non-Volatile Memory(NVM)and DRAM to relieve the situation.NVM storage has the advantages of low power consumption,high density,low latency and byte-addressable,but also has some disadvantages,such as writing power consumption is higher than reading,write latency is higher than read,and the write counts of NVM is limited.The asymmetry problem about read and write in NVM should be paid more attention when using the hybrid NVM and DRAM memory.This thesis propose a new kind of efficient memory page management mechanism called HMPM(Hybrid Memory Page Management)based on the hybrid NVM and DRAM memory Architecture.The main idea of this mechanism is to save the data pages with different access characteristics in the appropriate memory space according to the characteristics of different memory media,so as to reduce the number of system migration operations and improve the system performance.The HMPM mechanism mainly optimizes the migration strategy of data pages.Unlike existing hybrid memory management mechanisms,HMPM allows a small number of writes directly taken place in the NVM area to reduce the migration of data pages.For data pages in DRAM space,HMPM reconsiders its access characteristics and selects read-intensive data pages for migration.At the same time,the HMPM mechanism uses a simple and efficient wearleveling algorithm to improve NVM's lifetime.This paper also optimizes the memory mapping access and concurrently apply for memory.In this thesis,the HMPM mechanism is simulated in the Linux kernel.The experiment includes two parts.One is the comparison between the HMPM mechanism and the existing memory management mechanism.The experimental results show that HMPM can reduce the migration operation by 79.4% and improve the system performance by 128.5% compared with the existing memory management mechanism.And the NVM lifetime is improved by 6.9 times;the second is the comparison between before and after optimization of memory mapping and concurrently apply for memory,the experimental results show that,after optimization,the migration operation of memory mapping can be reduced by an average of 41.5%,and the average cost of time is reduced by 46.3% when apply for memory concurrently.