The Design And Implementation Of Wear-leveling Aware Fine-grained Memory Management Based On Non-volatile Memory

In recent years,with the popularization of information technology,a large number of data are continuously generated,processed and stored,so the demand for storage technology is increasingly high.Non-volatile Memory: as an emerging storage technology,is regarded as a potential storage technology with high access speed,high storage density,byte addressable and non-volatile,etc.From the computers with embedded systems to common computers to the computers with large data centers,the Non-volatile Memory is being deployed into single-type or hybrid storage systems.In these systems,the different application scenarios will produce a lot of Fine-grained data.But these data updated frequently and scattered distribution are a huge test for the using time of computer literacy and access performance.Non-volatile memory(NVM)as a new storage technology,there are two questions needed to face when it is applied:(1)the storage unit of NVM has a disadvantage of low tolerance level,making it difficult to afford the operation of data frequently updating.(2)The traditional storage memory management technology doesn't take the question that the using time of computer literacy brought by data update in to account.(3)The traditional storage management technology does not take the characteristics of the hybrid storage system into account,and this way can not combine the advantages of traditional memory technology and give full play of the NVM.Therefore,this paper aims to project a storage memory based on NVM,especially with hybrid storage medium such as mixed DRAM and NVM,and design a memory management mechanism with balance small granularity.Besides highly memory access efficiency,at the same time,the memory allocation of Fine-grained data can be optimized to improve the efficiency of system using.The main research contents of this paper are as following:(1)The efficient memory management of Fine-grained data.Fine-grained data distributed and scattered in the memory,causing the access frequency is high.So the smallest basic memory unit is a cache line as 64 byte,at the same time,the storage space within the pages is divided according to the basic unit.Adopting the way of pages to manage the data.(2)The wear balance strategy of small size memory space.With the frequent updating of Fine-grained data,it is bound to affect the efficiency of the system by adding complex statistical mechanism to optimize wear balance.In this paper,the sequential writing mechanism is adopted to solve the problem of wear balance,so that the system efficiency is very high.(3)The management optimization of mixed storage medium.Due to the problem of writing tolerance of NVM,the introduction of DRAM hybrid can use DRAM to share the writing wear problem of partial data.At the same time,the introduction of DRAM hybrid memory system can not only reduce the use cost,but also use DRAM to share some data's write wear problem.This paper updates the temporary data of metadata and page level management data in DRAM and updates the mechanism of conditional persistence from DRAM to NVM,which further solve the problem of wear and tear in the metadata area.(4)Persistence mechanism.Due to the configuration of NVM,therefore,the system power recoverability is an important indicator.In the process of the DRAM data brush back to NVM,if power outages,the data produced will be inconsistency.Therefore,the method of writing logging is used to maintain consistency.(5)Multi-threading optimization of the memory allocation mechanism.The single-threaded memory allocation mechanism generates resource preemption in the face of multi-threading,resulting in waste of system performance.Therefore,the use of thread-local resources allows each thread to have a local resource buffer,which can reduce the problem of resource competition and improve the performance of multi-threading.Based on the proposed problem,we designed Wear-leveling Aware Fine-grained allocator(WAFA).The experimental results show that the total write counts of pages in WAFA can achieve 81.1% and 40.1% reduction over that of NVMalloc and nvm_malloc,the state-of-the-art wear-conscious allocator for NVM.For the write counts of per memory unit,the co-efficient of variation of WAFA is 69.4% and 97.3% smaller than that of NVMalloc and malloc,respectively.Meanwhile,WAFA can achieve 9.6% higher performance than glibc malloc.