Research On Hybrid Cache Optimization Strategy Combined With New Non-Volatile Memory

As an important part of the computer system,Cache is used to reduce the huge gap in access speed between the processor and the main memory.However,with the development of processor technology and the advancement of semiconductor manufacturing technology,it exhibits many restrictions in traditional RAM-based cache technology,such as static power consumption and the increase in the area of memory cells.As a new type of non-volatile memory,STT-RAM not only has the advantages of high integration and low power consumption,but also has faster access speeds,greater storage density and longer write lifetime.STT-RAM is considered to be the most likely to be used as Cache instead of SRAM.However,the slow write speed and high write energy consumption limit the practical application of STT-RAM.In order to combine the advantages of high storage density of STT-RAM,low static power consumption and fast access speed of SRAM,and minimize the write operation of STT-RAM cache,this thesis studies the data allocation and migration strategies of hybrid caches in order to optimize the performance of hybrid caches as LLC(Last Level Cache)under non-inclusive and exclusive policies.The main research contents and contributions of this thesis are as follows:(1)Aiming at the hybrid Cache under the non-inclusive strategy(referred to as non-inclusive hybrid Cache),a write-intensive prediction technology based on perceptual learning,WIPPL(Write-Intensive Predictor based on Perceptual Learning),is proposed.WIPPL implements a write-intensive prediction algorithm,which can predict whether a cache block is a write-intensive one,and guide the data distribution of non-inclusive hybrid cache according to the prediction result,thereby reducing the number of write operations of the STT-RAM cache area.Compared with the existing non-inclusive hybrid Cache management strategy,the proposed non-inclusive hybrid Cache management strategy based on perceptual learning has higher prediction accuracy,which can effectively reduce the write operation of the STT-RAM cache area and save the energy consumption of the Cache system.(2)Aiming at the hybrid cache under the exclusive strategy(referred to as exclusion hybrid cache),a reuse distance prediction technology RDP(Reuse Distance Prediction)is proposed.RDP divides the cache blocks into three reuse categories,i.e.,far,medium,and near according to reuse distance.The near-reuse distance cache blocks that may be reused in a short period of time are placed in the SRAM cache area of the exclusive hybrid cache.The far-reuse distance cache block,which will not be accessed before being evicted from the cache,bypasses the exclusive hybrid cache,thereby reducing the number of write operations of the STT-RAM cache.Compared with SRAM-based Cache,the exclusive hybrid Cache management strategy based on reused distance prediction proposed in this paper can save more Cache energy consumption.The hybrid cache management strategy proposed in this paper can effectively alleviate the contradiction between the high static energy consumption of SRAM in hybrid cache and the high write energy consumption and large write delay of STT-RAM,so as to effectively reduce the overall energy consumption of Cache while meeting the requirements of Cache system delay.