Data Allocation Optimization Based On Non-volatile Memory In Embedded Systems

Tranditional storage architure of embedded systems mainly suffers from two problems: firstly,DRAM as the main memory has the defects of high energy consumption and low expansitive;secondly,SPM based on SRAM has been ultilized to replace tranditional cache of hardware-control,however,SRAM include of the strawbacks of large area and high leakage power.Due to its advantages of lower static power consumption and higher density,novel NVMs can not only significantly reduce the static energy consumption generated by DRAM,but also can fill the gap of performance between cache,main memory and external memory.Nevertheless,NVMs include the problems of that the performance on read operation and write operation is asymmetric,and that the performances of MLC and SLC are complementary.Futhermore,the extremely unbalanced access pattern of problems will further aggravate the above problems.Therefore,based on the novel storage architectures of DRAM/NVM hybrid memory and SLC/MLC hybrid SPM,this thesis will deeply explore the access characteristics of read and write operations.And this thesis will make full use of the advantages of NVMs,and avoid the effects on the disadvantages of NVMs,to present the data allocation technologes in heterogeneous storage,which can satisfy the requirements on performance and energy of embedded systems.It mainly includes:1.Based on DRAM/NVM hybrid memory in embedded systems,this thesis proposes a real-time guaranted data allocation strategy.This thesis analyzes the programs in embedded systems with compiler and conditional branching probability.Then,considering about the writing characteristics of each data,the technology proposed by this paper decide the overall arrangement of each data and heterogeneous storage.And the data whose write operation is frequnent will be stored in DRAM with lower writing energy.On the contrary,the data whose read operation is frequnent will be stored in NVM with higher density.In this way,this paper can achieve the optimal data access energy,under the precondition of guaranteeing the real-time requirement of embedded systems.2.Based on NVM SLC/MLC hybrid SPM in embedded systems,this thesis proposes a dynamic data allocation strategy.In order to optimize the data access overhead of SLC/MLC hybrid SPM,this thesis considers about the complementary advatantage of NVM SLC/MLC on access performance and density at first.Then,this thesis ultilize the characteristic of SLC/MLC dynamically switchability.Next,this thesis unit the contriants of storage location,heterogeneous storage overhead,and memory size in novel architecture.Finally,an integer linear programming formulation and a near optimal algorithm are constructed to solve these problems.In the end,experimental simulation platforms of DRAM/NVM hybrid memory and NVM SLC/MLC hybrid SPM are constructed.Especially,the experimental results show that the technologys this thesis proposed can efficiently improve the performance,decrease the energy consumption,and reduce the data access latency of embedded systems.Through ultilizing the characteristics of embedded system programs,this thesis studies data allocation based on novel NVMs to break through the bottleneck of the storage in embedded systems and promote the applications and developments of embedded devices in the Internet of things,big data and other fields.