Research And Implementation Of Cache Optimization For Hybrid Storage Systems

In recent years,NVMs have been widely used in cache,main memory,and external memory to alleviate the performance differences between CPUs and memories.Among them,there are more and more researches on integrating NVM into SPM.NVM has the advantages of fast reading and writing speed,high density and persistence,but it also has the shortcomings of limited number of writes.This thesis aims to study the data variable allocation method of NVM-based SPM storage architecture,which can improve the system performance while giving full play to the advantages of NVM.Firstly,this thesis proposes a utilization-aware data variable allocation method on NVM-based SPM,named DVA(Data Variable Allocation).Based on the genetic algorithm,DVA obtains the optimal solution through multiple iterations of genetic operations.It can not only determine whether to allocate data variables to SPM,but also distribute data variables evenly on SPM.Experiments show that DVA can obtain an allocation scheme which is very close to the optimal solution.Compared with other methods,DVA can extend the lifetime by 9.17%.Then,this thesis proposes a writing-aware data variable allocation method on hybrid SRAM+NVM SPM,named DVAWF(Data Variable Allocation based on Writing Frequency).By comparing writing frequency and writing threshold,DVAWF judges the type of data variables and determine the storage location of data variables on hybrid SPM.Experiments show that hybrid SRAM+NVM SPM architecture can reduce the number of writing operations to NVM by 17.9% compared to the pure NVM-based SPM architecture.Compared with other methods,DVAWF can reduce the number of writing operations to NVM by 50.13%.Finally,this thesis proposes two energy optimization of branch-aware data variable allocation methods based on hybrid SPM,named BSA+EDA(Branch-Based Static Analysis+Energy-Based Data Allocation)and NNBP+EDA(Neural Network Branch Prediction+Energy-Based Data Allocation).BSA+EDA includes a branch-based static analysis strategy and an energy-based data allocation strategy to reduce energy consumption by reducing the number of reading and writing operations.NNBP+EDA introduces neural network as a dynamic branch prediction strategy to improve the energy consumption by improving the accuracy of branch prediction.Experiments show that compared with other methods,the maximum and average energy consumption improvement of BSA+EDA are 39.4% and 25.1%,respectively.BSA+EDA is suitable for data variable allocation of simple structure program.Compared with BSA+EDA,NNBP+EDA is more suitable for data variable allocation of complex structure programs.When the program structure is more complicated,the energy consumption optimization and writing optimization effect of NNBP+EDA is more obvious.