Research On Optimization Techniques Of Memory Deduplication

Memory deduplication is a technique to remove the redundant data,which can save memory space and improve the whole system performance.After comparison,identified duplicate pages are merged through Copy-On-Write(COW)operation,where only a single backup needs to be kept in the memory.Now many literatures improve the performance of memory deduplication,but there are still some problems.First,existing methods identify memory pages that need to be checked repeatedly by scanning processes' address space.It wastes some CPU cycles when duplication ratio is high or some specific pages need to be ignored.Secondly,searching duplicated pages incurs additional computation overhead.Although some papers use spatial locality to improve the efficiency,they still suffer from low continuous distribution and high failure of prediction under unstable workloads.Finally,with the emergence and application of non-volatile memory,the new features should be taken into account on merging duplicated pages,and corresponding adjustments can improve the efficiency of memory usage.To solve the problem of inefficient memory scanning,this paper proposes a physical address label-based scanning method,which can find unmerged pages quickly by generating hints and using bitmap.Our experiments demonstrate that this method can reduce invalid scanning overhead.On the other hand,considering the large capacity of non-volatile memory,this paper designs a multithreaded scanning scheme to improve memory deduplication performance.To reduce the overhead of identifying duplicated pages,this paper proposes a memory deduplication algorithm of using enhanced spatial locality.On one hand,the logical addresses of different pages can keep better continuity,which can make use of spatial locality.On the other hand,comparing several potential locations to find duplication can improve hit rate.During the prediction procedure,duplicated pages are merged directly and regular routine is avoided.Experimental results under several environments show that this method can increase predictable opportunity,improve prediction hit ratio and reduce identification time of duplication.To make use of characteristics of non-volatile memory,such as low power consumption,this paper makes some adjustments when merging duplicated pages and handling COW break.The single backup is saved in non-volatile memory as much as possible to reduce of system power consumption,and new private pages are allocated from DRAM for frequent write operations.Finally,this paper verifies the practicability of this scheme through some experiments.