System-level Data Partitioning Policy Based On Hybrid Memory

Some Non-Volatile Memories,such as Phase-Change Memory(PCM),have been proposed to solve the high energy consumption and low density problems.The disadvantages of byte addressable,high density and low energy consumption make PCM become an alternative for the Dynamic Random Access Memory(DRAM).However,some drawbacks,especially higher read/write latency,restrict PCM replace DRAM completely.Therefore,hybrid main memory,which consists of both PCM and DRAM,is a promising main memory architecture for current.The biggest challenge for the PCM/DRAM hybrid main memory is to use advantages of both mediums to prevent drawbacks.In other words,it must to allocate data into higher write cost PCM or higher energy consumption DRAM.So,the key is to partition data into different characteristics.According to the data characteristic,determine allocation.The mainstream research almost based on migration.Allocate page directly by the traditional method.When the page shown the different behaviors of the expected results,it would be migrated to another storage medium.It is to control the writing and reading distribution,thus reduce the access memory latency.The algorithm based on migration bring some extra options and energy consume,which affected the average access time of the whole system.This paper analyzes the shortcomings of the existing partition algorithm and what caused that.In order to reduce the migration,this paper present Virtual Page Behavior based Page Management Policy for Hybrid Main Memory(VBPM).The contribution of this paper as follow:(1)Introduced the VPBM page partitioning strategy research background,significance(1)Introduced the VPBM page partitioning strategy research background,significance and research status at home and abroad,and describes the hybrid architecture and related simulator related technologies.(2)Analyzed the merits and demerits of the existing hybrid memory partitioning strategy in detail,and proposes the page partitioning strategy based on the virtual page memory behavior according to the existing partitioning strategy.Through the simulator in the cloud obtain the access behavior data of each virtual page,according to the page to visit the behavior of data to predict the virtual page visit behavior,the page directly divided.Because the same process virtual page memoryaccess behavior is similar in many operations,through the visit to the behavior data allocation page,you can avoid running page migration,improve system efficiency and reduce memory consumption,extended PCM life.(3)Proposed an adaptive partitioning method for predicting the access behavior in multi-process environment to solve the problem that the difficult of the simulation is geometric growth in the environment of multi-core and multi-process concurrency.Through the frequency of the page to predict the frequency of multiprocess environment to determine the page partition and preemptive strategy.In the case where the simulator acquires sufficient access behavior data,it is possible to calculate the slope of the access based on the difference between the number of times the page has been accessed and the expected number of times of access.The final position of the page is determined based on the result of the comparison of the calculated access frequency with the expected frequency of the competing page.This adaptive partitioning strategy can effectively reduce the simulation cost and improve the overall efficiency of the system.(4)Summarized the work and important contribution,and makes the prospect and the plan for the future work.