Write Scheduling And Data Movement Optimization For Solid State Drives

NAND flash memory-based solid-state drives(SSDs)are widely leveraged in digital devices,because of the high performance and reliability.Due to the feature of erase-before-write in SSDs,it requires to carry out garbage collection for reclaiming the space that occupied by invalid data in SSDs.But,frequent garbage collection operations may lead to an unevenness of block erasures across all SSD blocks.Therefore,the technique of wear leveling is adopted to balance the erasure of blocks and then extend the lifetime of SSDs.On the other hand,a considerable amount of read requests may result in read interference in SSDs,which impacts the SSD reliability.Therefore,the approach of read refresh is proposed to account for this problem.In other words,it is extremely significant to overcome how to more efficiently conduct the wear leveling operations and dispatch the read refresh operations,to expand the lifespan of SSDs and speedup the processing on I/O requests.To address this issue,this thesis completes the following three pieces of work.Firstly,by analyzing the frequency distribution of addresses of I/O requests,this thesis presents a write request mapping scheme when dispatching I/O requests,as well as a scheme of data regrouping when conducting wear leveling.On the one hand,the logical addresses of frequent write requests sometimes occur as patterns,which means these requests more possibly being updated together.Therefore,dispatching these write requests into the same block can cut down the overhead of valid page moves in garbage collection operation.On the other hand,read requests also have cold and hot addresses.Therefore,distributing hot read data to different parallel units during wear leveling operations can balance the distribution of read requests,for consequently increasing the access performance of read data.Secondly,the time intervals between I/O requests do exist idle periods,but it is difficult to finish a read refresh process within one of the idle periods.This thesis argues that the read refresh operation should be split into some partial operations and such partial operations can be separately completed.Though making use of the reinforcement learning,it dispatches the partial read refresh in the right time.That is to say,carrying out partial read refresh operations by largely utilizing the idle interval between two I/O requests can reduce their negative impacts on the response time of normal I/O requests.Finally,in order to evaluate the effectiveness of the proposed mechanisms,this thesis designs and implements the proposed optimization methods and state-of-the-art approaches.The experimental results illustrate that the new proposed schemes can effectively decrease I/O response time and improve the endurance of SSDs,in contrast to the selected state-of-the-art approaches.