Research On Hybrid Memory Management Scheduling Strategy Based On Multi-granularity Pages

As the memory consumption of data-intensive applications gradually increases,the existing Dynamic Random Access Memory(DRAM)as a computer memory has been difficult to meet the needs of large capacity,high density and low energy consumption.The emerging non-volatile memory technology(Non-Volatile Memory,NVM)has the characteristics of large capacity,high density and low power consumption.It can be combined with DRAM to form a large-capacity hybrid memory to meet application requirements.However,due to the current gap between NVM and DRAM in read and write performance,in a large-capacity hybrid memory system,hot and cold data needs to be reasonably migrated between different types of memory to achieve optimal performance.At the same time,the overhead of virtual and real address translation is becoming increasingly prominent in large-capacity hybrid memory systems.Therefore,how to efficiently manage large-capacity hybrid memory is a major research focus.This paper proposes a multi-granularity paging management for NVM / DRAM hybrid memory(MGPM)based on the combination of superpages and regular pages.MGPM uses pages of different granularity to manage NVM and DRAM memory,which manages DRAM with conventional 4KB page granularity and manages NVM with superpage granularity.Based on the spatial locality of hot spot data distribution in superpages,MGPM innovatively splits the superpage into multiple groups(including several consecutive regular pages),and uses the group as the basic granularity for migration scheduling within the superpage to avoid frequent migration of regular pages as a unit,which saves many CPU cycles and system resources.At the same time,MGPM also adopts a periodic dynamic migration threshold adjustment strategy based on the current DRAM cache utilization rate,and places as much data as possible in DRAM to take full advantage of DRAM performance and capacity advantages.In addition,in order to achieve accurate monitoring of lightweight hot data,MGPM has proposed a two-level hot data recognition mechanism that combines hotness prediction and accurate counting.It achieves accurate hot data recognition without adding too much on-chip space overhead.In order to reduce the address translation overhead,MGPM has designed a group migration mapping table,which accelerates the speed of address translation while maintaining the continuity of superpages.The MGPM hybrid memory management scheduling strategy was implemented and tested and verified on a system-wide simulation platform composed of ZSim and NVMain.Experimental results show that,compared with the current state-of-the-art hybrid main memory system that supports superpages,Rainbow,MGPM's system IPC has increased by an average of 25.14% and energy consumption has been reduced by an average of 41.8%.In addition,MGPM also shows certain advantages in terms of data traffic generated during page migration,the impact of dynamic threshold adjustment strategies,and space overhead.