| With the rapid development of big data technology,traditional mechanical hard disks can no longer meet the performance requirements of current massive data.Flash memory stands out due to its high performance,low energy consumption,high reliability,and shock resistance.However,flash memory has characters including asymmetrical I/O latency,erase-before-write and limited lifetime.Therefore flash-specific buffer algorithms are required in order to optimize the performance of flash storage systems.Current research on buffer optimization of flash storage systems can be further improved.On the one hand,current buffer management algorithms do not take the locality of the workload into account,which results in limited performance.On the other hand,the address mapping algorithms have low hit rate and high garbage collection cost.Aiming at the above two problems,this paper proposes a buffer management algorithm based on dynamic page weight and a locality-aware page-level address mapping algorithm.The innovation and main work of this paper are reflected in the following aspects:1.This paper proposes a buffer management algorithm based on dynamic page weight(DPW-LRU),aiming at the problem that the existing flash buffer management algorithms do not take the locality of the workload into account and cause limited performance.The characteristic of page's access and frequency are used to classify the pages.The working region and exchange region of the buffer and the migration strategy are designed.In addition,by evaluating the time locality,eviction cost and recency of the buffered pages,the dynamic page weight is calculated to support page migration.By considering the locality of the pages comprehensively,DPW-LRU algorithm improves the hit rate of buffer and reduces the overhead of page access.2.This paper proposes a locality-aware page-level mapping algorithm(LPM),aiming at the problem of low hit rate and high garbage collection cost of existing page-level address mapping algorithms.Novel buffer partitioning strategy is designed based on locality of mapping items.The buffer is divided into sequential cached mapping table,working cached mapping table and exchange cached mapping table.The compression and prefetching of sequential address mapping items are implemented.The proximity of the mapping items is used to reduce the buffer occupation and improve the hit rate.In addition,cluster-based write-back aggregation strategy is designed to improve the locality of the page in the data block in order to reduce the number of block erasures and response time.3.In this paper,the DPW-LRU algorithm and the LPM algorithm are implemented by using two flash storage system simulation platforms,Flash-DBSim and FlashSim.After several sets of experimental comparisons,the effectiveness of the proposed algorithm is verified in terms of hit rate,write operations,total latency etc.Experimental results demonstrate that DPW-LRU improves the hit rate by 8.3%,reduces write operations by 22.6%and overall latency by 18.8%compared with LRU algorithm.Compared with DFTL algorithm,LPM algorithm improves the hit rate by 8.5%,reduces the number of block erasures by 10.5%and the response time by 12.3%. |
