Optimized Design Of Page Management For Hybrid Memory Systems

Memory system plays an important role in the whole storage system.It is mainly responsible for caching external storage data and ensuring fast calculation and access of processor.With the development of big data and cloud computing,the amount of data loaded by the system is increasing rapidly,which requires more and more mem-ory capacity and performance.However,traditional Dynamic Random Access Mem-ory(DRAM)has encountered bottlenecks in integration density and power consump-tion,which limits the large-scale expansion of the memory system.In order to solve the bottleneck of memory system development,a large number of researchers are com-mitted to building hybrid memory systems with Non-Volatile Random Access Mem-ory(NVRAM).However,because NVRAM is still in the development stage,the re-search on hybrid memory systems is still not mature enough,mainly manifested in the following aspects:the existing hybrid memory management strategy is not per-fect enough,and the optimization of data migration between hybrid media is lacking;In large-scale server deployment,multi-processor memory access is not uniform,multi-processor hybrid memory deployment is diverse,and multi-task deployment is complex,resulting in multi-dimensional isomerism in the server;The existing NVRAM simula-tion method is not accurate enough,and simulating hybrid memory management in the kernel of a real system has great challenges.This paper mainly studies the page management strategy of hybrid memory sys-tems,which include:The optimization of data migration between media in hybrid mem-ory systems;Under large NUMA architecture server,hybrid memory system manage-ment optimization for multi-dimensional isomerism scenarios;The real system simu-lates NVRAM access delay to complete the simulation of hybrid memory systems.The main research contents and contributions are as follows:(1)Optimization of Data Migration in Hybrid Memory SystemsThe existing hybrid memory architecture design can reasonably allocate and man-age pages according to page characteristics.However.in order to realize reasonable page allocation management,additional data migration between different media is in-troduced.Excessive data migration will seriously affect the performance of the sys-tem.This paper proposes an optimization strategy to mitigate data migration.Different from the traditional hybrid memory allocation strategy,our strategy to mitigate data migration is based on multi-level queue management and optimizes the conversion re-lationship between multi-level queues.On the other hand,we depict the dependency of data pages and predict the allocation strategy of related data pages according to the characteristics of the allocated data pages.The experimental results show that our opti-mization strategy can improve the performance of at least 10%?20%for scenarios with large data migration.(2)Optimization of Multidimensional Hybrid Scenes under NUMA Architec-tureWith the development of processor technology,NUMA architecture has been pop-ularized in servers.With the increase of NUMA nodes,the access inconsistency of each node in the server becomes more and more serious.At the same time,with the develop-ment of new memory devices and the complexity of application scenarios,the present situation of multi-dimensional isomerism is presented under NUMA architecture.This brings huge challenges to hybrid memory management.This paper proposes a hybrid memory allocation strategy based on NUMA architecture,which achieves better perfor-mance for multi-dimensional isomerism scenarios.Different from the traditional man-agement of single-node hybrid memory systems,we add a management strategy for task allocation to NUMA multi-node management,and analyze the characteristics of various tasks at the same time,thus realizing a more dimensional management strategy of hybrid memory systems.Experimental results show that our optimization strategy can signif-icantly improve the overall performance of the system for the multi-task deployment scenario in NUMA architecture.Compared with hybrid memory system management that does not identify multi-nodes,we can achieve 65%performance improvement.(3)Simulation of Hybrid Memory Systems in Real SystemsSince NVRAM has not been commercialized,the experimental analysis of hybrid memory systems is based on DRAM to simulate NVRAM.However,the existing hybrid memory simulators do not go deep into the kernel memory management level,but are implemented by external application simulation or by task application self-management.The realization of external simulation is difficult to accurately depict,and the applica-tion self-management is too intrusive for large-scale deployment.Both applications have great limitations.This paper proposes a simulation method to realize hybrid mem-ory system in real system.This simulation method can be completely transparent to the application deployed on the upper layer of the system and will not be invasive to the application.On the other hand,this design completely follows the existing hard-ware control flow and the control flow of system kernel software.The reliability of the simulation is ensured.