Data Allocation Strategy For Hybrid Memory Using A Double Fitness Genetic Algorithm

Nowadays,the development of multi-core systems meets many challenges.Since the development gap between processor and memory has increased rapidly,cache controlled by hardware in traditional computing systems can't meet the development requirements.On the other hand,the difference in development speed between CPU and memory may constrain the development of on-chip multiprocessor systems and even cause more energy cost.However,Scratch-Pad Memory(SPM)can make use of resources with lower energy cost and less time latency because of its merits in energy cost,chip area and predictability.Due to its advantages,SPM are widely used in multi-core systems.The new byte addressable Non-volatile Memory(NVM)has many advantages in terms of energy cost,storage density,read write performance and data update,which has attracted researchers' attention and became a hot spot in the memory.This paper aims to find a data allocation strategy with low energy cost and write operations in the hybrid memory,which can reduce the total energy cost,improve the performance of system and prolong the lifetime of NVM.In our paper,the on-chip hybrid memory consists of a SRAM and a NVM,while the off-chip memory is made of DRAM.In the hybrid memory consists of a SRAM and a PRAM,the asymmetry of read and write operations of NVM limits the efficiency of data allocation,so this paper take both energy cost and write operations of NVM into consideration.Traditional data allocation strategy can only consider one objective function at a time,some algorithms can consider two objective functions,but they are very complex and their applicability is limited.While there are two factors should be considered in this paper and traditional data allocation strategy based on greedy algorithm can't meet our requirements.Therefore,in order to rationally utilize the advantages and avoid the disadvantages of various parts of memory,this paper proposes an adapted genetic algorithm(AGA)and an adapted double fitness genetic algorithm(DFGA)to realize our two goals at a time.Finally,this paper compares these three kinds of algorithms in data allocation strategy.According to the experimental results analyze,DFGA reduces 11.62%of energy cost,43.88%of write operations to NVM compared with greedy algorithm.Moreover,DFGA reduces 33.76%of write operations to NVM compared with AGA on average.In a word,AGA mainly reduces energy cost compared with greedy algorithm while DFGA reduces number of write operations to NVM based on the reduction of energy cost of AGA.While the NVM's lifetime is associated with the number of write operations to NVM,DFGA can not only reduce energy cost but also prolong the lifetime of NVM.