Research On The Design Of NAND Flash-based Hybrid Solid State Drive

Although the technology of SSD is improving,and the advantages of SSD in terms of read / write speed,power consumption,security,shock resistance and drop size are more obvious,single flash memory based SSDs in terms of performace cost ratio are still unable to meet consumer's demand.Therefore,experts have proposed the use both of SLC(Single-Level Cell)and MLC(Multi-Level Cell)flash memory chips together to build low-cost,high-performance and high-capacity hybrid solid-state hard drives.In the design of hybrid solid-state drives,hot data identification and wear leveling are critical issues.On the one hand,one key point of data allocation is the hot data identification,and accurate identification and reasonable allocation of hot data can improve the read/write performance and lifetime of hybrid SSDs;On the other hand,the durability of SLC and MLC flash chips is quite different,and achieving wear balance between SLC and MLC is critical to the lifetime of hybrid SSDs.Therefore,this paper mainly focuses on the optimization of wear leveling and the accuracy of hot data identification algorithm in the hybrid SSDs design.In the aspect of wear leveling,this paper proposes a wear leveling aware flash translation layer(WLAFTL)for hybrid SSDs.Firstly,WLAFTL proposes a dynamic data allocation scheme,which is based on the fusion of wear leveling and write-request size.Specifically,it adaptively adjusts the threshold value of cold/hot data identification according to the wear speed of SLC and MLC region,and then,the small write requests whose size is less than the threshold value are send to SLC region,while the large write requests are send to MLC region.Secondly,WLAFTL proposes a cold data garbage collection/migration scheme of SLC,which is based on the fusion of wear leveling and first in first out(FIFO).This scheme can reduce the number of data migration from SLC to MLC region.Our experiment results demonstrate that under the condition of using the same address mapping scheme,compared to CombFTL and CFTL,WLAFTL can achieve up to 13.6% and 12.7% average response time improvement,as well as 9.2% and 20.4% total block erasure reduction,respectively.Moreover,it can be able to achieve better wear leveling between SLC and MLC than CombFTL and CFTL.In the aspect of hot data identification,this paper proposes a hot data identification algorithm which is based on the fusion of the K-means and KNN(K-Nearest Neighbor).Firstly,we cluster requests by using K-means algorithms according to the size of request to determine that the request is cold data or hot data;then,we classify the request by KNN algorithms according to the logical page number of request;finally,we modify the judgement result by nearest neighboring rule according to the logical page number if the classification results of these two methods aren't the same.Our experiment results demonstrate that,compared whith 2-means clustering,2-level LRU and MBF,KKH can reduce false identification rates by an average of 9.7%?8.3% and 2.9%,as well as more than 75% of memory space reduction on average in memory overhead,respectively.