Study On Page Migration Algorithm Of Hybrid Memory System Based On Emerging Non-volatile Memory

Dynamic Random Access Memory(DRAM)is widely used in various computing devices for its high performance.However,DRAM is limited by storage density and power consumption,and it is difficult to further expand the capacity.Compared with DRAM,the emerging non-volatile Memory(NVM)promises higher density and lower energy consumptions,thus is becoming an attractive solution for replacing or complementing DRAM in future computer systems.But NVM also has shortcomings such as poor write performance and unbalanced read/write performance.Therefore,building hybrid memory using DRAM and NVM is a viable solution that combines the advantages of both.In order to take full advantage of the hybrid memory heterogeneous hardware features and overcome NVM performance defects,frequently accessed hot data can be migrated to DRAM through data page migration.However,the existing research uses sperate mechanisms to identify hot and cold pages in DRAM and NVM,respectively.This leads to invalid data page exchange and affects overall performance of hybrid memory.In response to the above questions,this paper has done the following work:Initially,we introduced and analyzed the existing hybrid memory page migration algorithm in detail,and expounded two main problems in the existing researches: invalid page migration introduced by inaccurate hot/cold pages identification and additional NVM writes introduced by frequent page migrations.Secondly,for the problem of invalid page migrations,we proposes a page migration method based on a unified hotness identification mechanism,named UIMigrate.UIMigrate focuses on the comparisons of pages' hotness and revenue of page migrations.It mainly includes two aspects: firstly,the global unified hotness identification mechanism is adopted to reduce the invalid page migrations to improve the hybrid memory performance;secondly,the related threshold is adjusted in real time through the dynamic adjustment mechanism to adapt to changing of memory access characteristics.Thirdly,for the problem of additional NVM writes,we proposes a fine-grained page writeback strategy based on NVM non-volatility,named LazyWriteback.When pages are swapped back into the NVM from DRAM,LazyWriteback can reuse some of the old data that was not rewritten in the NVM frame,eliminating some of the extra NVM write operations,thereby improving the performance of the hybrid memory system and reducing NVM memory wear.Finally,we experiment and evaluate the above methods and strategies with a dozen typical loads of SPEC CPU 2006.Experiment results show that UIMigrate can effectively reduce invalid page swapping in hybrid memory systems,and the performace of hybrid memory system can be increased by up to 90%.In addition,LazyWriteback can reduce the NVM write operation introduced by page migration by up to 75% when the DRAM memory is small and the memory access of workload is sparse.